Fix BRCA training collapse, mobile layout and results history (#3)

- Fix mobile dashboard layout (9a10d98c1ae2d54e01852d5dd5f5a7cda0f8e577)
- Fix BRCA training collapse with lab-grid defaults (62a8628cbfa0da789c989c19608ee2c2f6bb536b)
- Add BRCA training diagnostics and lab grid results (d1ec2eed8c6cd02479da5b4234a056cb83761ab5)
- Keep the action bar visible without scrolling (bb42e347c5bc5557015683583eba1f7f6def7354)
- Scroll long result tables inside their own box (e413269a95818f6de610533927deab1f2d19ad52)
- Align Results tables and fix overlapping table headers (c5021fff6b8ddc17cd0286c07d68bc7db95e569c)
- Align the confusion matrix with the training history chart (adaf2dbccdb6e262631b7f94b561b5d2618fd8f4)


Co-authored-by: Faruk Yildirim <F4RUKYLDRM@users.noreply.huggingface.co>

app.py

@@ -34,6 +34,7 @@ from biolmnet.data import (
     validate_and_align_omics,
 )
 from biolmnet.training import (
+    DEFAULT_VALIDATION_FRACTION,
     Hyperparameters,
     ModelBundle,
     pathway_importance,
@@ -44,8 +45,11 @@ from biolmnet.ui import components as ui
 from biolmnet.ui.styles import CSS
 
 
+# Long result tables scroll inside a box of this height (px) instead of stretching the page.
+TABLE_MAX_HEIGHT = 320
+
 GENEPT_OPTIONS = {
-    "Auto — bulk or single-cell based on source": "auto",
+    "Auto — lab-recommended single-cell context": "auto",
     "Bulk · large-3 context": "embedding_original_large_3.parquet",
     "Bulk · original ada-text context": "embedding_original_ada_text.parquet",
     "Single-cell · cell type, tissue, drug & pathway": (
@@ -101,9 +105,8 @@ def _resolve_embedding_file(option: str, source_mode: str, example_dataset: str)
     selected = GENEPT_OPTIONS[option]
     if selected != "auto":
         return selected
-    if source_mode == "BioLM-NET examples" and example_dataset == "scTrioseq2":
-        return "embedding_associations_cell_type_tissue_drug_pathway_openai_large.parquet"
-    return "embedding_original_large_3.parquet"
+    # Bozdag Lab recommendation (J. Rifat, 2026-10-17) for every source.
+    return "embedding_associations_cell_type_tissue_drug_pathway_openai_large.parquet"
 
 
 def _graph_preview_html(mask_density: str, gene_nodes: str, pathway_units: str, genept_dim: str) -> str:
@@ -514,6 +517,14 @@ def _loss_plot(history: list[dict[str, float]]) -> go.Figure:
     return figure
 
 
+def _validation_split_note(validation_fraction: float) -> str:
+    validation_percent = round(float(validation_fraction) * 100)
+    return (
+        f'<div class="field-hint">{validation_percent}% of samples are held out for validation '
+        f"(stratified by class). The remaining {100 - validation_percent}% are used for training.</div>"
+    )
+
+
 def estimate_training_duration(
     workspace: PreparedWorkspace | None,
     epochs: int,
@@ -651,8 +662,7 @@ def train_workspace(
         importance = pathway_importance(bundle).head(100)
         architecture_audit = _architecture_audit(bundle)
         confusion_html = ui.blueprint_div(
-            f'<div class="mono" style="margin-bottom:12px">Confusion matrix · validation split</div>'
-            + ui.confusion_matrix_html(result.confusion.tolist(), bundle.label_names)
+            ui.confusion_matrix_html(result.confusion.tolist(), bundle.label_names)
         )
 
         run_meta_update = {
@@ -949,6 +959,7 @@ def refresh_results(bundle: ModelBundle | None, run_meta: dict):
                 "Build the biological architecture, train a model, or run prediction to populate this page."
             ),
             gr.update(visible=False),
+            None,
         )
 
     macro_auc = run_meta.get("macro_auc")
@@ -965,7 +976,8 @@ def refresh_results(bundle: ModelBundle | None, run_meta: dict):
     )
         + "</div>"
     )
-    return stats, gr.update(visible=True), "", gr.update(visible=True)
+    loss_figure = _loss_plot(bundle.history) if bundle.history else None
+    return stats, gr.update(visible=True), "", gr.update(visible=True), loss_figure
 
 
 THEME = gr.themes.Base(primary_hue="gray", neutral_hue="gray")
@@ -986,18 +998,19 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
     session_id_state = gr.State("")
     intro_step_state = gr.State(0)
     intro_dismissed_state = gr.BrowserState(False, storage_key="biolmnet_intro_dismissed")
+    nav_ids = ["nav-data", "nav-train", "nav-export", "nav-predict", "nav-results"]
 
     with gr.Row(elem_classes=["app-shell"]):
         with gr.Column(elem_classes=["rail-col"]):
             with gr.Column(elem_classes=["rail"]):
-                gr.HTML(ui.brand_block_html())
+                gr.HTML(ui.brand_block_html(), elem_id="app-header")
                 rail_row_htmls = []
                 rail_row_buttons = []
                 with gr.Column(elem_classes=["rail-nav"]):
                     for index, label in enumerate(STAGE_LABELS):
                         with gr.Column(elem_classes=["rail-row-wrap"]):
                             row_html = gr.HTML(ui.rail_row_html(index + 1, label, "on" if index == 0 else "off", "Active" if index == 0 else "Locked"))
-                            row_button = gr.Button("", elem_classes=["rail-row-click"])
+                            row_button = gr.Button("", elem_id=nav_ids[index], elem_classes=["rail-row-click"])
                         rail_row_htmls.append(row_html)
                         rail_row_buttons.append(row_button)
                 with gr.Column(elem_classes=["rail-spacer", "rail-plate-wrap"]):
@@ -1011,7 +1024,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                 intro_open_button = gr.Button("Introduction", size="sm", variant="secondary", elem_classes=["btn-ghost", "intro-open-btn"])
 
             # ── Data & Priors ──────────────────────────────────────���─
-            with gr.Column(visible=True) as data_page:
+            with gr.Column(visible=True, elem_id="page-data", elem_classes=["workspace-page", "data-workspace-page"]) as data_page:
                 with gr.Row(elem_classes=["pghd"]):
                     gr.HTML(
                         ui.title_block_html(
@@ -1028,7 +1041,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                         source_mode = gr.Radio(
                             ["BioLM-NET examples", "GitHub folder", "Upload files"],
                             value="BioLM-NET examples", show_label=False, container=False,
-                            elem_classes=["seg-radio"],
+                            elem_id="dataset-source", elem_classes=["seg-radio"],
                         )
                         with gr.Row(elem_classes=["srow-row", "first"]):
                             with gr.Column(scale=1, min_width=0):
@@ -1038,7 +1051,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                                             gr.HTML('<div class="field-label">Repository dataset</div>')
                                             example_dataset = gr.Dropdown(
                                                 ["BRCA", "COAD", "GBM", "scTrioseq2"], value="BRCA",
-                                                show_label=False, container=False,
+                                                show_label=False, container=False, elem_id="dataset-select",
                                             )
                                 with gr.Column(visible=False) as github_group:
                                     github_folder = gr.Textbox(
@@ -1056,7 +1069,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                                         uploaded_dna_pathway = gr.File(label="DNA → pathway mapping", file_types=[".csv"], type="filepath", elem_classes=["file-slot"])
 
                         gr.HTML('<div class="mono sect sect-spaced">Resolved files</div>')
-                        resolved_files_table = gr.Dataframe(
+                        resolved_files_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, 
                             headers=["File", "Shape", "Rows matched", "Status"], interactive=False, wrap=False,
                             value=pd.DataFrame(),
                         )
@@ -1134,8 +1147,8 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                         graph_preview = gr.HTML(_graph_preview_html("—", "—", "—", "—"))
 
                 with gr.Row(visible=False):
-                    architecture_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
-                    enrichment_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                    architecture_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                    enrichment_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
 
                 with gr.Row(elem_classes=["actbar-row"]):
                     data_actionbar_note = gr.HTML(f'<div class="actbar-note">{ui.esc("Choose a source, then build the biological architecture.")}</div>')
@@ -1144,12 +1157,13 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                             "Re-validate inputs", size="sm", variant="secondary", elem_classes=["actbar-secondary-btn"]
                         )
                         prepare_button = gr.Button(
-                            "Build biological architecture", size="sm", variant="primary", elem_classes=["actbar-primary-btn"]
+                            "Build biological architecture", size="sm", variant="primary",
+                            elem_id="build-architecture-btn", elem_classes=["actbar-primary-btn"]
                         )
                 preparation_status = gr.HTML(visible=False)
 
             # ── Train Model ──────────────────────────────────────────
-            with gr.Column(visible=False) as train_page:
+            with gr.Column(visible=False, elem_id="page-train", elem_classes=["workspace-page"]) as train_page:
                 with gr.Row(elem_classes=["pghd"]):
                     gr.HTML(
                         ui.title_block_html(
@@ -1165,22 +1179,22 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Epochs"))
                             with gr.Column():
-                                epochs = gr.Slider(5, 200, value=50, step=5, show_label=False, container=False)
+                                epochs = gr.Slider(5, 200, value=Hyperparameters.epochs, step=5, show_label=False, container=False)
                         with gr.Row(elem_classes=["srow-row"]):
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Batch size"))
                             with gr.Column():
-                                batch_size = gr.Dropdown([8, 16, 32, 64, 128], value=16, show_label=False, container=False)
+                                batch_size = gr.Dropdown([8, 16, 32, 64, 128], value=Hyperparameters.batch_size, show_label=False, container=False)
                         with gr.Row(elem_classes=["srow-row"]):
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Learning rate"))
                             with gr.Column():
-                                learning_rate = gr.Number(value=0.001, minimum=0.000001, show_label=False, container=False)
+                                learning_rate = gr.Number(value=Hyperparameters.learning_rate, minimum=0.000001, show_label=False, container=False)
                         with gr.Row(elem_classes=["srow-row"]):
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("L2 weight decay"))
                             with gr.Column():
-                                weight_decay = gr.Number(value=0.01, minimum=0, show_label=False, container=False)
+                                weight_decay = gr.Number(value=Hyperparameters.weight_decay, minimum=0, show_label=False, container=False)
                         with gr.Row(elem_classes=["srow-row", "last"]):
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Dropout"))
@@ -1202,7 +1216,13 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Validation fraction"))
                             with gr.Column():
-                                validation_fraction = gr.Slider(0.1, 0.4, value=0.2, step=0.05, show_label=False, container=False)
+                                validation_fraction = gr.Slider(
+                                    0.1, 0.4, value=DEFAULT_VALIDATION_FRACTION, step=0.05, show_label=False, container=False
+                                )
+                                validation_split_note = gr.HTML(_validation_split_note(DEFAULT_VALIDATION_FRACTION))
+                                validation_fraction.change(
+                                    _validation_split_note, inputs=[validation_fraction], outputs=[validation_split_note]
+                                )
                         with gr.Row(elem_classes=["srow-row"]):
                             with gr.Column(min_width=140):
                                 gr.HTML(ui.esc("Optimizer"))
@@ -1221,11 +1241,16 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                         gr.HTML('<div class="mono" style="margin-bottom:9px">Training result</div>')
                         training_status = gr.HTML(_status("Prepare the architecture before training."))
                         loss_plot = gr.Plot(label="Loss by epoch", show_label=False)
-                        gr.HTML('<div class="mono" style="margin:20px 0 9px">Paper-faithful defaults</div>')
+                        gr.HTML('<div class="mono" style="margin:20px 0 9px">Architecture defaults</div>')
                         gr.HTML(
                             "".join(
                                 [
                                     ui.kv_plain_row("First layer", "trainable W ⊙ M"),
+                                    ui.kv_plain_row(
+                                        "Activations",
+                                        f"{Hyperparameters.biological_activation} / {Hyperparameters.projection_activation} / "
+                                        f"{Hyperparameters.fusion_activation}",
+                                    ),
                                     ui.kv_plain_row("PDI weights", "binary"),
                                     ui.kv_plain_row("PPI weights", "STRING norm."),
                                     ui.kv_plain_row("Attention", "GenePT query"),
@@ -1238,11 +1263,12 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                     train_actionbar_note = gr.HTML('<div class="actbar-note">Set hyperparameters, then train.</div>')
                     with gr.Row(elem_classes=["actbar-buttons"]):
                         train_button = gr.Button(
-                            "Train BioLM-NET on ZeroGPU", size="sm", variant="primary", elem_classes=["actbar-primary-btn"]
+                            "Train BioLM-NET on ZeroGPU", size="sm", variant="primary",
+                            elem_id="train-btn", elem_classes=["actbar-primary-btn"]
                         )
 
             # ── Export Artifacts ───────────────────────���─────────────
-            with gr.Column(visible=False) as export_page:
+            with gr.Column(visible=False, elem_id="page-export", elem_classes=["workspace-page"]) as export_page:
                 with gr.Row(elem_classes=["pghd"]):
                     gr.HTML(
                         ui.title_block_html(
@@ -1270,7 +1296,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                     gr.HTML('<div class="actbar-note">Bundle written to a temp file on export · not persisted between Space restarts</div>')
 
             # ── Predict ──────────────────────────────────────────────
-            with gr.Column(visible=False) as predict_page:
+            with gr.Column(visible=False, elem_id="page-predict", elem_classes=["workspace-page"]) as predict_page:
                 with gr.Row(elem_classes=["pghd"]):
                     gr.HTML(
                         ui.title_block_html(
@@ -1326,7 +1352,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                     with gr.Column(scale=6, min_width=0):
                         gr.HTML('<div class="mono" style="margin-bottom:9px">Predictions · last successful run</div>')
                         distribution_panel = gr.HTML(ui.empty_note_html("No predictions yet."))
-                        prediction_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                        prediction_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
                         prediction_download = gr.File(label="Prediction CSV", interactive=False, elem_classes=["file-slot", "btn-block"])
 
                 with gr.Row(elem_classes=["actbar-row"]):
@@ -1334,12 +1360,12 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                     with gr.Row(elem_classes=["actbar-buttons"]):
                         predict_button = gr.Button(
                             "Run inference", size="sm", variant="primary", interactive=False,
-                            elem_classes=["actbar-primary-btn"],
+                            elem_id="predict-btn", elem_classes=["actbar-primary-btn"],
                         )
                 prediction_status = gr.HTML(visible=False)
 
             # ── Results ──────────────────────────────────────────────
-            with gr.Column(visible=False) as results_page:
+            with gr.Column(visible=False, elem_id="page-results", elem_classes=["workspace-page"]) as results_page:
                 with gr.Row(elem_classes=["pghd"]):
                     gr.HTML(
                         ui.title_block_html(
@@ -1351,21 +1377,27 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
                 results_empty_note = gr.HTML(ui.empty_note_html("Build the biological architecture, train a model, or run prediction to populate this page."))
                 with gr.Row(elem_classes=["results-stat-row"], visible=False) as results_stats_row:
                     results_stats = gr.HTML("")
-                with gr.Row(elem_classes=["results-grid"], equal_height=False, visible=False) as results_content:
-                    with gr.Column(scale=6, min_width=0):
-                        gr.HTML('<div class="mono sect">Training history</div>')
-                        results_loss_plot = gr.Plot(show_label=False)
-                        gr.HTML('<div class="mono sect sect-spaced">Sparse architecture audit</div>')
-                        results_architecture_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
-                    with gr.Column(scale=6, min_width=0):
-                        results_confusion = gr.HTML("")
-                        gr.HTML('<div class="mono sect sect-spaced">Pathway attention · top retained</div>')
-                        results_pathway_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                # Two rows so the side-by-side panels in each row start at the same height.
+                with gr.Column(visible=False) as results_content:
+                    with gr.Row(elem_classes=["results-grid"], equal_height=False):
+                        with gr.Column(scale=6, min_width=0):
+                            gr.HTML('<div class="mono sect">Training history</div>')
+                            results_loss_plot = gr.Plot(show_label=False)
+                        with gr.Column(scale=6, min_width=0):
+                            gr.HTML('<div class="mono sect">Confusion matrix · validation split</div>')
+                            results_confusion = gr.HTML("")
+                    with gr.Row(elem_classes=["results-grid"], equal_height=False):
+                        with gr.Column(scale=6, min_width=0):
+                            gr.HTML('<div class="mono sect">Sparse architecture audit</div>')
+                            results_architecture_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                        with gr.Column(scale=6, min_width=0):
+                            gr.HTML('<div class="mono sect">Pathway attention · top retained</div>')
+                            results_pathway_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
                 with gr.Column(visible=False) as results_tables_bottom:
                     gr.HTML('<div class="mono sect sect-spaced">Validation predictions</div>')
-                    results_validation_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                    results_validation_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
                     gr.HTML('<div class="mono sect sect-spaced">Predictions and class probabilities</div>')
-                    results_prediction_table = gr.Dataframe(interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
+                    results_prediction_table = gr.Dataframe(max_height=TABLE_MAX_HEIGHT, interactive=False, wrap=False, value=pd.DataFrame(), headers=[""], column_count=(1, "dynamic"))
                 with gr.Row(elem_classes=["actbar-row"]):
                     gr.HTML(
                         '<div class="actbar-note">Validate cohorts, preprocessing and performance before '
@@ -1642,7 +1674,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
     ).then(
         refresh_results,
         inputs=[model_state, run_meta_state],
-        outputs=[results_stats, results_stats_row, results_empty_note, results_content],
+        outputs=[results_stats, results_stats_row, results_empty_note, results_content, results_loss_plot],
         show_progress="hidden",
     ).then(
         lambda df: gr.update(value=df), inputs=[validation_predictions_state], outputs=[results_validation_table],
@@ -1718,7 +1750,7 @@ with gr.Blocks(title="BioLM-NET Workbench", fill_width=True) as demo:
         show_progress="hidden",
     ).then(
         refresh_results, inputs=[model_state, run_meta_state],
-        outputs=[results_stats, results_stats_row, results_empty_note, results_content],
+        outputs=[results_stats, results_stats_row, results_empty_note, results_content, results_loss_plot],
         show_progress="hidden",
     )
 

biolmnet/training.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import copy
+import math
 import random
 from dataclasses import asdict, dataclass
 from typing import Callable
@@ -25,21 +26,36 @@ from .model import BioLMNet
 
 ProgressCallback = Callable[[float, str], None]
 
+# 75% training / 25% validation. The validation partition drives early
+# stopping and best-checkpoint selection, so it is not an untouched test set.
+DEFAULT_VALIDATION_FRACTION = 0.25
 
+# Activations BioLMNet accepts.
+ACTIVATIONS = ("relu", "sigmoid", "tanh")
+
+
+# Defaults follow the Bozdag Lab grid (J. Rifat, 2026-10-17): the configuration that stayed
+# stable across 8 seeds on BRCA (macro F1 0.749 +/- 0.018). The published activations
+# (relu/sigmoid/tanh) collapse at this learning rate and L2; see
+# diagnostics/2026-10-30-lab-hyperparameter-grid.
 @dataclass(frozen=True)
 class Hyperparameters:
-    epochs: int = 50
-    batch_size: int = 16
-    learning_rate: float = 0.001
+    epochs: int = 200
+    batch_size: int = 32
+    learning_rate: float = 0.01
     weight_decay: float = 0.01
     dropout: float = 0.3
     projection_dim: int = 64
     fusion_dim: int = 12
-    validation_fraction: float = 0.2
+    validation_fraction: float = DEFAULT_VALIDATION_FRACTION
     optimizer: str = "Adam"
     class_weighting: bool = True
     early_stopping_patience: int = 12
     seed: int = 42
+    biological_activation: str = "tanh"
+    projection_activation: str = "relu"
+    fusion_activation: str = "tanh"
+    sgd_momentum: float = 0.03
 
     def validate(self) -> None:
         if not 1 <= self.epochs <= 1000:
@@ -54,6 +70,11 @@ class Hyperparameters:
             raise ValueError("Validation fraction must be between 0.05 and 0.5.")
         if self.optimizer.lower() not in {"adam", "sgd"}:
             raise ValueError("Optimizer must be Adam or SGD.")
+        for name in ("biological_activation", "projection_activation", "fusion_activation"):
+            if getattr(self, name) not in ACTIVATIONS:
+                raise ValueError(f"{name.replace('_', ' ').capitalize()} must be one of {', '.join(ACTIVATIONS)}.")
+        if not 0 <= self.sgd_momentum < 1:
+            raise ValueError("SGD momentum must be in [0, 1).")
 
 
 @dataclass
@@ -86,6 +107,36 @@ def set_reproducible_seed(seed: int) -> None:
         torch.cuda.manual_seed_all(seed)
 
 
+def split_train_validation(
+    labels: np.ndarray, validation_fraction: float, seed: int
+) -> tuple[np.ndarray, np.ndarray]:
+    """Stratified sample-index split shared by both omics and the labels."""
+    classes, class_counts = np.unique(labels, return_counts=True)
+    if class_counts.min() < 2:
+        raise ValueError(
+            "Each class needs at least two samples for a stratified train/validation split."
+        )
+    sample_count = len(labels)
+    validation_count = math.ceil(validation_fraction * sample_count)
+    train_count = sample_count - validation_count
+    if min(validation_count, train_count) < len(classes):
+        raise ValueError(
+            f"A validation fraction of {validation_fraction:.2f} on {sample_count} samples "
+            f"gives {train_count} training and {validation_count} validation samples; "
+            f"each partition needs at least one sample per class ({len(classes)} classes) "
+            "for a stratified split. Adjust the validation fraction or add samples."
+        )
+
+    indices = np.arange(sample_count)
+    train_index, validation_index = train_test_split(
+        indices,
+        test_size=validation_fraction,
+        random_state=seed,
+        stratify=labels,
+    )
+    return train_index, validation_index
+
+
 def _fit_scaler(values: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
     mean = values.mean(axis=0, dtype=np.float64).astype(np.float32)
     scale = values.std(axis=0, dtype=np.float64).astype(np.float32)
@@ -121,6 +172,9 @@ def _make_model(
         projection_dim=hyperparameters.projection_dim,
         fusion_dim=hyperparameters.fusion_dim,
         dropout=hyperparameters.dropout,
+        biological_activation=hyperparameters.biological_activation,
+        projection_activation=hyperparameters.projection_activation,
+        fusion_activation=hyperparameters.fusion_activation,
     )
 
 
@@ -153,18 +207,8 @@ def train(
     hyperparameters.validate()
     set_reproducible_seed(hyperparameters.seed)
     labels = workspace.labels
-    classes, class_counts = np.unique(labels, return_counts=True)
-    if class_counts.min() < 2:
-        raise ValueError(
-            "Each class needs at least two samples for a stratified train/validation split."
-        )
-
-    indices = np.arange(len(labels))
-    train_index, validation_index = train_test_split(
-        indices,
-        test_size=hyperparameters.validation_fraction,
-        random_state=hyperparameters.seed,
-        stratify=labels,
+    train_index, validation_index = split_train_validation(
+        labels, hyperparameters.validation_fraction, hyperparameters.seed
     )
     gene_mean, gene_scale = _fit_scaler(
         workspace.gene_expression[train_index]
@@ -204,6 +248,7 @@ def train(
         optimizer = torch.optim.SGD(
             model.parameters(),
             lr=hyperparameters.learning_rate,
+            momentum=hyperparameters.sgd_momentum,
             weight_decay=hyperparameters.weight_decay,
         )
     else:
@@ -357,9 +402,9 @@ def train(
             "projection_dim": hyperparameters.projection_dim,
             "fusion_dim": hyperparameters.fusion_dim,
             "dropout": hyperparameters.dropout,
-            "biological_activation": "relu",
-            "projection_activation": "sigmoid",
-            "fusion_activation": "tanh",
+            "biological_activation": hyperparameters.biological_activation,
+            "projection_activation": hyperparameters.projection_activation,
+            "fusion_activation": hyperparameters.fusion_activation,
         },
     }
     bundle = ModelBundle(

biolmnet/ui/styles.py

@@ -236,6 +236,12 @@ body, .gradio-container {
   flex: 1 1 auto !important;
   width: calc(100% - 236px) !important;
 }
+.workspace-page,
+.workspace-page > .form {
+  min-height: calc(100vh - 45px) !important;
+  display: flex !important;
+  flex-direction: column !important;
+}
 
 /* rail row (overlay: visible .rr under an invisible full-size button) */
 .rail-row-wrap { position: relative; }
@@ -312,7 +318,11 @@ body, .gradio-container {
 .sect { margin: 0 0 9px; padding-bottom: 8px; border-bottom: 1px solid var(--color-text); }
 .sect-spaced { margin-top: 26px !important; }
 
-.actbar-row { margin-top: auto !important; align-items: center !important; justify-content: space-between !important; flex-wrap: nowrap !important; gap: 20px !important; padding: 16px 30px !important; border-top: 1px solid var(--color-divider) !important; background: color-mix(in srgb, var(--color-text) 3%, transparent) !important; }
+/* Gradio's container sets overflow: hidden, which disables position: sticky inside it;
+   clip still clips stray content without breaking the sticky action bar below. */
+.gradio-container { overflow: clip !important; }
+/* Pinned to the bottom of the visible frame so the page's main action never needs a scroll to find. */
+.actbar-row { position: sticky !important; bottom: 0; z-index: 20; margin-top: auto !important; align-items: center !important; justify-content: space-between !important; flex-wrap: nowrap !important; gap: 20px !important; padding: 16px 30px !important; border-top: 1px solid var(--color-divider) !important; background: color-mix(in srgb, var(--color-text) 3%, var(--color-bg)) !important; box-shadow: 0 -8px 18px -14px color-mix(in srgb, var(--color-text) 45%, transparent); }
 .actbar-note { font: 500 10px/1.4 var(--font-mono); color: var(--color-muted-2); }
 .actbar-note.error { color: var(--color-error); }
 .actbar-buttons {
@@ -424,6 +434,46 @@ body, .gradio-container {
 .strip.error .strip-kicker { color: var(--color-error); }
 .strip .strip-body { margin-top: 4px; font-size: 12.5px; line-height: 1.5; }
 .empty-note { padding: 14px 0; color: var(--color-muted); font-size: 13px; }
+.field-hint { margin-top: 6px; color: var(--color-muted); font-size: 12px; line-height: 1.4; }
+
+@media (min-width: 901px) and (max-height: 820px) {
+  .topbar {
+    padding-top: 10px;
+    padding-bottom: 10px;
+  }
+  .pghd {
+    padding-top: 18px;
+  }
+  .pghd h2 {
+    font-size: 31px;
+  }
+  .pghd .desc {
+    margin-top: 4px;
+    line-height: 1.35;
+  }
+  .stage-grid {
+    padding-top: 16px !important;
+    gap: 22px !important;
+  }
+  .data-workspace-page .section-heading-row {
+    margin-bottom: 8px !important;
+  }
+  .data-workspace-page .sect-spaced {
+    margin-top: 18px !important;
+  }
+  .data-workspace-page .srow-row {
+    min-height: 42px;
+  }
+  .data-workspace-page .dataframe,
+  .data-workspace-page .table-wrap {
+    max-height: 118px !important;
+    overflow: auto !important;
+  }
+  .actbar-row {
+    padding-top: 11px !important;
+    padding-bottom: 11px !important;
+  }
+}
 
 /* ── slots (one-line gr.File) ─────────────────────────────────────────── */
 .file-slot .file-preview { border: 1px solid var(--color-divider) !important; border-radius: var(--radius-md); background: #fff; font-size: 12.5px !important; }
@@ -571,7 +621,9 @@ body, .gradio-container {
 .gradio-container table { color: var(--color-text) !important; font-size: 12.5px !important; }
 .gradio-container thead th {
   font: 500 10px/1 var(--font-mono) !important; letter-spacing: .1em; text-transform: uppercase;
-  color: var(--color-muted-2) !important; background: transparent !important;
+  color: var(--color-muted-2) !important;
+  /* opaque so rows scrolling under the pinned header stay hidden */
+  background: var(--color-bg) !important;
   border-bottom: 1px solid var(--color-divider) !important;
 }
 .gradio-container tbody td { border-bottom: 1px solid var(--color-row-rule) !important; }
@@ -723,13 +775,223 @@ body, .gradio-container {
 
 /* ── responsive ───────────────────────────────────────────────────────── */
 @media (max-width: 1000px) {
-  .app-shell { display: block !important; }
-  .rail-col { min-width: 100% !important; max-width: 100% !important; }
-  .rail { position: relative; min-height: auto; border-right: 0; border-bottom: 1px solid var(--color-divider); }
-  .stage-grid, .results-grid { grid-template-columns: 1fr !important; }
+  .app-shell {
+    display: block !important;
+    min-height: 100vh !important;
+  }
+  .rail-col {
+    min-width: 100% !important;
+    max-width: 100% !important;
+    width: 100% !important;
+    flex: none !important;
+  }
+  .rail {
+    position: sticky;
+    top: 0;
+    z-index: 40;
+    height: auto !important;
+    min-height: 0 !important;
+    padding: 12px 12px 10px;
+    border-right: 0;
+    border-bottom: 1px solid var(--color-divider);
+    background: color-mix(in srgb, var(--color-bg) 96%, #fff);
+  }
+  .rail-brand {
+    display: flex;
+    align-items: baseline;
+    gap: 8px;
+    padding: 0 4px 9px;
+  }
+  .rail-brand .name {
+    font-size: 22px;
+  }
+  .rail-brand .mono {
+    margin-top: 0 !important;
+    font-size: 8.5px;
+  }
+  .rail-kicker,
+  .rail-spacer {
+    display: none !important;
+  }
+  .rail-nav {
+    display: grid !important;
+    grid-auto-flow: column;
+    grid-auto-columns: minmax(154px, max-content);
+    gap: 6px;
+    overflow-x: auto;
+    overscroll-behavior-x: contain;
+    padding: 0 2px 2px;
+    scrollbar-width: none;
+  }
+  .rail-nav::-webkit-scrollbar {
+    display: none;
+  }
+  .rail-row-wrap {
+    min-width: 154px !important;
+  }
+  .rr {
+    min-height: 36px;
+    padding: 8px 9px 8px 8px;
+    gap: 8px;
+    border-left-width: 2px;
+    font-size: 12.5px;
+    white-space: nowrap;
+  }
+  .rr .rn {
+    width: 19px;
+    height: 19px;
+    font-size: 9.5px;
+  }
+  .rr .st {
+    font-size: 8px;
+    letter-spacing: .08em;
+  }
+  .workspace-col {
+    width: 100% !important;
+    min-width: 0 !important;
+  }
+  .topbar {
+    padding: 10px 116px 10px 14px;
+    gap: 10px;
+  }
+  .topbar > .num {
+    display: none;
+  }
+  .intro-open-btn {
+    right: 12px;
+    top: 5px;
+  }
+  .pghd {
+    padding: 20px 14px 0;
+    gap: 10px;
+  }
+  .pghd h2 {
+    font-size: 30px;
+  }
+  .pghd .desc {
+    max-width: none;
+    font-size: 13px;
+  }
+  .stage-grid,
+  .results-grid {
+    display: grid !important;
+    grid-template-columns: minmax(0, 1fr) !important;
+    padding: 20px 14px 0 !important;
+    gap: 22px !important;
+  }
+  .stage-grid > *,
+  .results-grid > * {
+    min-width: 0 !important;
+    max-width: 100% !important;
+    width: 100% !important;
+  }
+  .results-stat-row,
+  .stat-grid {
+    padding: 20px 14px 0 !important;
+  }
+  .metric-strip {
+    grid-template-columns: repeat(2, minmax(0, 1fr));
+  }
+  .field-pair {
+    display: grid !important;
+    grid-template-columns: minmax(0, 1fr) !important;
+    gap: 12px !important;
+  }
+  .srow-row {
+    display: grid !important;
+    grid-template-columns: minmax(118px, 34%) minmax(0, 1fr) !important;
+    gap: 12px !important;
+    padding: 8px 0 !important;
+  }
+  .srow-row > * {
+    min-width: 0 !important;
+  }
+  .actbar-row {
+    align-items: flex-start !important;
+    flex-wrap: wrap !important;
+    padding: 12px 14px !important;
+    gap: 12px !important;
+  }
+  .actbar-buttons {
+    width: 100% !important;
+    flex: 1 1 100% !important;
+    justify-content: flex-start !important;
+    flex-wrap: wrap !important;
+  }
+  .actbar-buttons button {
+    min-height: 34px !important;
+    padding: 0 12px !important;
+  }
+  .gradio-container .table-wrap,
+  .gradio-container .dataframe,
+  .gradio-container .wrap.svelte-1ipelgc,
+  .tbl {
+    max-width: 100%;
+    overflow-x: auto;
+  }
+  .intro-card {
+    width: min(556px, calc(100vw - 28px)) !important;
+    padding: 22px 20px 18px !important;
+  }
 }
 @media (max-width: 620px) {
-  .pghd h2 { font-size: 26px; }
-  .stat-grid, .metric-strip { grid-template-columns: repeat(2, minmax(0,1fr)) !important; }
+  .topbar {
+    padding-right: 112px;
+  }
+  .topbar .mono {
+    font-size: 8.5px;
+    letter-spacing: .06em;
+  }
+  .pghd h2 {
+    font-size: 27px;
+  }
+  .section-heading-row {
+    display: block !important;
+  }
+  .section-heading-row .mono {
+    margin-top: 6px;
+    text-align: left !important;
+  }
+  .seg-radio label {
+    padding: 7px 8px !important;
+    font-size: 12px !important;
+  }
+  .stat-grid,
+  .metric-strip {
+    grid-template-columns: 1fr !important;
+  }
+  .srow-row {
+    grid-template-columns: 1fr !important;
+    gap: 6px !important;
+  }
+  .srow-value-row {
+    justify-content: flex-start !important;
+  }
+  .rr {
+    font-size: 12px;
+  }
+  .rr .st {
+    display: none;
+  }
+  .rail-nav {
+    grid-auto-columns: minmax(126px, max-content);
+  }
+  .rail-row-wrap {
+    min-width: 126px !important;
+  }
+  .intro-card h4 {
+    font-size: 24px;
+  }
+  .intro-card p {
+    font-size: 13px;
+  }
+  .intro-card-foot {
+    align-items: flex-start !important;
+  }
+  .intro-card-actions {
+    justify-content: flex-start !important;
+    flex: 1 1 100% !important;
+    gap: 9px !important;
+  }
 }
 """

diagnostics/2026-10-16-brca-training-collapse/README.md

@@ -0,0 +1,37 @@
+# BRCA training collapse diagnostics (2026-10-16)
+
+These diagnostic runs explain why BRCA training in the Workbench predicted one class for every sample. None of this is production code: the scripts call the project's code unchanged and only override `Hyperparameters` fields.
+
+## Layout
+
+- `scripts/`
+  - `diagnose.py`: builds the BRCA workspace through `app.prepare_workspace`, then reports class counts, class weights, mask sparsity, per-epoch predictions and activations.
+  - `sweep.py`: weight-decay sweep at seed 42.
+  - `seeds.py`: 8 seeds at weight_decay 1e-4 and 0.01.
+  - `epochs.py <limit>`: 8 seeds at weight_decay 1e-4 with the given epoch limit.
+  - `epochs_summary.py`: summarises the three `epochs_*.json` files.
+- `results/`: raw outputs from the runs on 2026-10-16.
+  - `brca_weighted.txt`: full `diagnose.py` output with the Workbench defaults.
+  - `weight_decay_sweep.txt`: weight-decay sweep.
+  - `paper_config_sgd.txt`: the notebook's SGD configuration.
+  - `effective_fanin_init.txt`: initialisation experiment, run inline with a monkeypatch.
+  - `seed_runs.csv` and `seed_runs.txt`: 8 seeds × 2 weight-decay values.
+  - `epochs_{50,100,200}.json`: per-run histories.
+  - `epochs_runs.csv`: per-run summary table.
+- `upstream/`: copies of the bozdaglab/BioLM-NET notebooks used for the comparison.
+
+## Reproduce
+
+Run from the repository root with the project venv:
+
+```bash
+D=diagnostics/2026-10-16-brca-training-collapse/scripts
+PYTHONPATH=$PWD .venv/bin/python $D/diagnose.py BRCA 1 50   # downloads data; caches workspace_BRCA.pkl (~37 MB) next to the scripts
+PYTHONPATH=$PWD .venv/bin/python $D/seeds.py
+for L in 50 100 200; do PYTHONPATH=$PWD .venv/bin/python $D/epochs.py $L; done
+.venv/bin/python $D/epochs_summary.py
+```
+
+`workspace_BRCA.pkl` is a regenerable cache. Don't commit it.
+
+Note: `epochs.py` and `epochs_summary.py` write their output files next to the scripts.

diagnostics/2026-10-16-brca-training-collapse/results/brca_weighted.txt

@@ -0,0 +1,147 @@
+
+======== RAW FILES (BRCA) ========
+gene           shape=(875, 999)  first cols=['A2ML1', 'ABAT', 'ABCA13']
+dna            shape=(875, 1000)  first cols=['A2LD1', 'A2ML1', 'ABAT']
+labels         shape=(875, 1)  first cols=['label']
+  raw label counts: {3.0: 436, 4.0: 147, 1.0: 131, 0.0: 115, 2.0: 46}
+gene_pathways  shape=(5614, 2)  first cols=['SYMBOL', 'PathwayID']
+dna_pathways   shape=(5410, 2)  first cols=['SYMBOL', 'PathwayID']
+
+======== WORKSPACE ========
+warnings: []
+samples: 875  classes: ['0.0', '1.0', '2.0', '3.0', '4.0']
+full class counts: {'0.0': 115, '1.0': 131, '2.0': 46, '3.0': 436, '4.0': 147}
+gene: shape=(875, 999) min=0 max=1 mean=0.408 zero-var cols=0
+dna: shape=(875, 1000) min=0 max=1 mean=0.52 zero-var cols=0
+
+======== MASKS ========
+gene: bio mask (999, 1008) density=0.2328% nnz=2,344 PDI=369 PPI=3559
+   inputs with 0 outgoing edges: 695/999   hidden genes in-degree median=1 max=32
+   nonzero mask values: min=0.988 median=0.998 max=1.000
+   pathway mask (1008, 133): hidden genes in no pathway=0/1008  genes/pathway median=35
+dna: bio mask (1000, 840) density=0.1423% nnz=1,195 PDI=158 PPI=1566
+   inputs with 0 outgoing edges: 792/1000   hidden genes in-degree median=1 max=6
+   nonzero mask values: min=0.988 median=0.998 max=1.000
+   pathway mask (840, 142): hidden genes in no pathway=0/840  genes/pathway median=32
+
+======== TRAINING  Hyperparameters(epochs=50, batch_size=16, learning_rate=0.001, weight_decay=0.01, dropout=0.3, projection_dim=64, fusion_dim=12, validation_fraction=0.25, optimizer='Adam', class_weighting=True, early_stopping_patience=12, seed=42) ========
+training class counts:   {'0.0': 86, '1.0': 98, '2.0': 35, '3.0': 327, '4.0': 110}
+validation class counts: {'0.0': 29, '1.0': 33, '2.0': 11, '3.0': 109, '4.0': 37}
+
+======== CLASS WEIGHTS ========
+weights passed to CrossEntropyLoss: [array([1.526, 1.339, 3.749, 0.401, 1.193], dtype=float32)]
+weight attribute seen inside loss.forward: [1.526 1.339 3.749 0.401 1.193]
+loss.forward calls: 1271
+
+======== HISTORY ========
+ epoch  training_loss  validation_loss  validation_accuracy  validation_f1_macro
+   1.0       1.642350         1.601534             0.150685             0.052381
+   2.0       1.629108         1.576027             0.150685             0.052381
+   3.0       1.624403         1.574691             0.497717             0.132927
+   4.0       1.613764         1.588255             0.168950             0.057813
+   5.0       1.628252         1.574925             0.497717             0.132927
+   6.0       1.620439         1.678558             0.168950             0.057813
+   7.0       1.626308         1.564224             0.497717             0.132927
+   8.0       1.618647         1.550071             0.497717             0.132927
+   9.0       1.617843         1.545797             0.497717             0.132927
+  10.0       1.614722         1.622764             0.168950             0.057813
+  11.0       1.616705         1.577437             0.497717             0.132927
+  12.0       1.619570         1.585625             0.168950             0.057813
+  13.0       1.617500         1.541410             0.497717             0.132927
+  14.0       1.619081         1.613365             0.132420             0.046774
+  15.0       1.616881         1.572324             0.497717             0.132927
+  16.0       1.616340         1.557646             0.497717             0.132927
+  17.0       1.615816         1.569439             0.497717             0.132927
+  18.0       1.613130         1.558936             0.497717             0.132927
+  19.0       1.611621         1.538758             0.497717             0.132927
+  20.0       1.610360         1.565398             0.497717             0.132927
+  21.0       1.607601         1.592779             0.132420             0.046774
+  22.0       1.612177         1.580287             0.497717             0.132927
+  23.0       1.614574         1.590151             0.168950             0.057813
+  24.0       1.612279         1.569710             0.497717             0.132927
+  25.0       1.607981         1.565809             0.497717             0.132927
+  26.0       1.611930         1.564028             0.497717             0.132927
+  27.0       1.612068         1.584551             0.132420             0.046774
+  28.0       1.613480         1.597781             0.132420             0.046774
+  29.0       1.613598         1.575038             0.497717             0.132927
+  30.0       1.617835         1.584576             0.497717             0.132927
+  31.0       1.610216         1.583199             0.497717             0.132927
+
+======== PER-EPOCH VALIDATION PREDICTION COUNTS ========
+epoch   1: counts=[0, 219, 0, 0, 0]  mean max-prob=0.205
+epoch   2: counts=[0, 219, 0, 0, 0]  mean max-prob=0.240
+epoch   3: counts=[0, 0, 0, 219, 0]  mean max-prob=0.216
+epoch   4: counts=[0, 0, 0, 0, 219]  mean max-prob=0.215
+epoch   5: counts=[0, 0, 0, 219, 0]  mean max-prob=0.216
+epoch   6: counts=[0, 0, 0, 0, 219]  mean max-prob=0.234
+epoch   7: counts=[0, 0, 0, 219, 0]  mean max-prob=0.223
+epoch   8: counts=[0, 0, 0, 219, 0]  mean max-prob=0.236
+epoch   9: counts=[0, 0, 0, 219, 0]  mean max-prob=0.239
+epoch  10: counts=[0, 0, 0, 0, 219]  mean max-prob=0.219
+epoch  11: counts=[0, 0, 0, 219, 0]  mean max-prob=0.212
+epoch  12: counts=[0, 0, 0, 0, 219]  mean max-prob=0.208
+epoch  13: counts=[0, 0, 0, 219, 0]  mean max-prob=0.232
+epoch  14: counts=[219, 0, 0, 0, 0]  mean max-prob=0.245
+epoch  15: counts=[0, 0, 0, 219, 0]  mean max-prob=0.216
+epoch  16: counts=[0, 0, 0, 219, 0]  mean max-prob=0.226
+epoch  17: counts=[0, 0, 0, 219, 0]  mean max-prob=0.218
+epoch  18: counts=[0, 0, 0, 219, 0]  mean max-prob=0.224
+epoch  19: counts=[0, 0, 0, 219, 0]  mean max-prob=0.234
+epoch  20: counts=[0, 0, 0, 219, 0]  mean max-prob=0.221
+epoch  21: counts=[219, 0, 0, 0, 0]  mean max-prob=0.214
+epoch  22: counts=[0, 0, 0, 219, 0]  mean max-prob=0.210
+epoch  23: counts=[0, 0, 0, 0, 219]  mean max-prob=0.209
+epoch  24: counts=[0, 0, 0, 219, 0]  mean max-prob=0.218
+epoch  25: counts=[0, 0, 0, 219, 0]  mean max-prob=0.223
+epoch  26: counts=[0, 0, 0, 219, 0]  mean max-prob=0.223
+epoch  27: counts=[219, 0, 0, 0, 0]  mean max-prob=0.212
+epoch  28: counts=[219, 0, 0, 0, 0]  mean max-prob=0.215
+epoch  29: counts=[0, 0, 0, 219, 0]  mean max-prob=0.217
+epoch  30: counts=[0, 0, 0, 219, 0]  mean max-prob=0.212
+epoch  31: counts=[0, 0, 0, 219, 0]  mean max-prob=0.211
+
+======== FINAL (best checkpoint) VALIDATION ========
+prediction counts: {'0.0': 0, '1.0': 0, '2.0': 0, '3.0': 219, '4.0': 0}
+average predicted probability per class: {'0.0': 0.19499999284744263, '1.0': 0.210999995470047, '2.0': 0.16200000047683716, '3.0': 0.23399999737739563, '4.0': 0.1979999989271164}
+training class prior: {'0.0': 0.131, '1.0': 0.149, '2.0': 0.053, '3.0': 0.498, '4.0': 0.168}
+std of each class prob across samples: [0.0, 0.0, 0.0, 0.0, 0.0]
+max-prob: min/median/max = [0.234, 0.234, 0.234]
+margin top1-top2: median = 0.0232
+first 8 validation rows (prob vectors):
+ [[0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]
+ [0.195 0.211 0.162 0.234 0.198]]
+metrics: {'validation_loss': 1.5387582778930664, 'accuracy': 0.4977168949771689, 'f1_macro': 0.1329268292682927, 'f1_weighted': 0.3307996436128745, 'precision_macro': 0.09954337899543378, 'recall_macro': 0.2, 'epochs_completed': 31, 'training_samples': 656, 'validation_samples': 219}
+confusion:
+ [[  0   0   0  29   0]
+ [  0   0   0  33   0]
+ [  0   0   0  11   0]
+ [  0   0   0 109   0]
+ [  0   0   0  37   0]]
+
+======== LOGITS / ACTIVATIONS (best checkpoint, validation rows) ========
+gene: bio pre-act mean=-4.48e-19 std=1.18e-16  dead ReLU units=6.6%  pathway std across samples=4.02e-17  sigmoid saturated(<.02|>.98)=0.0%  proj std across samples=0
+dna: bio pre-act mean=-2.92e-21 std=1.15e-18  dead ReLU units=9.4%  pathway std across samples=2.91e-19  sigmoid saturated(<.02|>.98)=0.0%  proj std across samples=0
+fusion tanh saturated(|x|>.98)=0.0%  fused std across samples=0
+logits mean per class: [-0.057  0.019 -0.246  0.124 -0.042]  logits std across samples per class: [0. 0. 0. 0. 0.]
+
+======== PARAMETER CHANGE FROM INIT ========
+gene_branch.biological.weight                 rel change=1.0000
+gene_branch.biological.bias                   rel change=1.0000
+gene_branch.pathway_attention.query           rel change=1.0000
+gene_branch.projection.weight                 rel change=1.0000
+gene_branch.projection.bias                   rel change=0.9941
+dna_branch.biological.weight                  rel change=1.0000
+dna_branch.biological.bias                    rel change=1.0000
+dna_branch.pathway_attention.query            rel change=1.0000
+dna_branch.projection.weight                  rel change=1.0000
+dna_branch.projection.bias                    rel change=0.9789
+fusion.weight                                 rel change=0.3960
+fusion.bias                                   rel change=0.2371
+output.weight                                 rel change=0.4115
+output.bias                                   rel change=0.2730

diagnostics/2026-10-16-brca-training-collapse/results/effective_fanin_init.txt

@@ -0,0 +1,16 @@
+Diagnostic only, NOT committed to biolmnet/model.py.
+BioMaskedLinear.reset_parameters was monkeypatched inside the run to use LeCun-uniform
+initialisation on each hidden gene's real (masked) in-degree, with zero bias:
+
+    def reset(self):
+        fan_in = self.mask.gt(0).sum(1).clamp(min=1).float()   # real in-degree per hidden gene
+        bound = (3.0 / fan_in).sqrt().unsqueeze(1)
+        with torch.no_grad():
+            self.weight.copy_((torch.rand_like(self.weight) * 2 - 1) * bound)
+            if self.bias is not None: self.bias.zero_()
+    biolmnet.model.BioMaskedLinear.reset_parameters = reset
+
+BRCA, seed 42, 75/25 split, class weighting on:
+
+eff-fan-in init + current default (Adam wd=.01)    acc=0.680 f1_macro=0.493 epochs=50 pred=[0, 34, 10, 159, 16] prob std=0.1422
+eff-fan-in init + paper (SGD lr=.01 wd=.01 200ep)  acc=0.511 f1_macro=0.497 epochs=148 pred=[80, 32, 9, 70, 28] prob std=0.1255

diagnostics/2026-10-16-brca-training-collapse/results/epochs_100.log

@@ -0,0 +1,9 @@
+limit=100 seed=0 done: best=49/61 acc=0.763 f1m=0.724
+limit=100 seed=1 done: best=50/62 acc=0.772 f1m=0.744
+limit=100 seed=2 done: best=50/62 acc=0.790 f1m=0.761
+limit=100 seed=3 done: best=71/83 acc=0.740 f1m=0.719
+limit=100 seed=7 done: best=32/44 acc=0.772 f1m=0.752
+limit=100 seed=42 done: best=38/50 acc=0.767 f1m=0.747
+limit=100 seed=123 done: best=37/49 acc=0.763 f1m=0.740
+limit=100 seed=2024 done: best=41/53 acc=0.758 f1m=0.737
+limit=100 ALL DONE

diagnostics/2026-10-16-brca-training-collapse/results/epochs_200.log

@@ -0,0 +1,9 @@
+limit=200 seed=0 done: best=49/61 acc=0.763 f1m=0.724
+limit=200 seed=1 done: best=50/62 acc=0.772 f1m=0.744
+limit=200 seed=2 done: best=50/62 acc=0.790 f1m=0.761
+limit=200 seed=3 done: best=71/83 acc=0.740 f1m=0.719
+limit=200 seed=7 done: best=32/44 acc=0.772 f1m=0.752
+limit=200 seed=42 done: best=38/50 acc=0.767 f1m=0.747
+limit=200 seed=123 done: best=37/49 acc=0.763 f1m=0.740
+limit=200 seed=2024 done: best=41/53 acc=0.758 f1m=0.737
+limit=200 ALL DONE

diagnostics/2026-10-16-brca-training-collapse/results/epochs_50.json

@@ -0,0 +1,3105 @@
+[
+ {
+  "limit": 50,
+  "seed": 0,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 0
+  },
+  "epochs_completed": 50,
+  "best_epoch": 49,
+  "early_stopped": false,
+  "hit_max": true,
+  "train_loss_at_best": 0.6369955467014778,
+  "val_loss_at_best": 0.6139159798622131,
+  "accuracy": 0.7625570776255708,
+  "f1_macro": 0.7241977003359705,
+  "f1_weighted": 0.7688921750314536,
+  "recall": {
+   "0.0": 0.7931034482758621,
+   "1.0": 0.8181818181818182,
+   "2.0": 0.5454545454545454,
+   "3.0": 0.8073394495412844,
+   "4.0": 0.6216216216216216
+  },
+  "pred_counts": {
+   "0.0": 47,
+   "1.0": 28,
+   "2.0": 8,
+   "3.0": 105,
+   "4.0": 31
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6172636282153245,
+    "validation_loss": 1.5241289138793945,
+    "validation_accuracy": 0.4977168949771689,
+    "validation_f1_macro": 0.1329268292682927
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.5862088872165214,
+    "validation_loss": 1.4785966873168945,
+    "validation_accuracy": 0.6484018264840182,
+    "validation_f1_macro": 0.4379253414135677
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.486709772086725,
+    "validation_loss": 1.3735746145248413,
+    "validation_accuracy": 0.5844748858447488,
+    "validation_f1_macro": 0.4254201336812392
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.3238095277693214,
+    "validation_loss": 1.1953068971633911,
+    "validation_accuracy": 0.593607305936073,
+    "validation_f1_macro": 0.4059264209867135
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.1833144280968644,
+    "validation_loss": 1.0689646005630493,
+    "validation_accuracy": 0.634703196347032,
+    "validation_f1_macro": 0.5306606392472901
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.0947908875418872,
+    "validation_loss": 0.9768268465995789,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.5910805860805861
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0506021293198191,
+    "validation_loss": 0.9280396103858948,
+    "validation_accuracy": 0.680365296803653,
+    "validation_f1_macro": 0.5913701533244877
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 0.999169804700991,
+    "validation_loss": 0.903730571269989,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.624375070811654
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9537633395776516,
+    "validation_loss": 0.8764855861663818,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.6222269848742044
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.9284529744124994,
+    "validation_loss": 0.870495080947876,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6800145976305553
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.9191909265227434,
+    "validation_loss": 0.8306586742401123,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6281475784972053
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.8589656418416558,
+    "validation_loss": 0.8354954123497009,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7001971862442815
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8755823068502473,
+    "validation_loss": 0.8147370219230652,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.6888091074938482
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8297777030526138,
+    "validation_loss": 0.7851540446281433,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.6843052757402035
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.7822338589807836,
+    "validation_loss": 0.777275562286377,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7143041489780522
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.7837169664662059,
+    "validation_loss": 0.7484090328216553,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.62636908727515
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.7800104937902311,
+    "validation_loss": 0.7660107016563416,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7004440484591072
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.7730191902416509,
+    "validation_loss": 0.7944260835647583,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.698860395932188
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7873704033653911,
+    "validation_loss": 0.7639477252960205,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7032332841288065
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7558516823663944,
+    "validation_loss": 0.7414798736572266,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7004440484591072
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7153960161092805,
+    "validation_loss": 0.7372803688049316,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7106842540796777
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7558937465272298,
+    "validation_loss": 0.7276725172996521,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.6967894385370114
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7580708134465102,
+    "validation_loss": 0.7531004548072815,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7047526663942533
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.745811897807005,
+    "validation_loss": 0.7062485218048096,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7125710434661245
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.7463469519847776,
+    "validation_loss": 0.7009532451629639,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7095670971936119
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.7205249856158,
+    "validation_loss": 0.7257617712020874,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7043211959834462
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.7136954007352271,
+    "validation_loss": 0.7064869999885559,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.6991019731795038
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.6904271637521139,
+    "validation_loss": 0.6817423701286316,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7013489796588387
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.7016249146403336,
+    "validation_loss": 0.7140704989433289,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7183729324463023
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.7061529159545898,
+    "validation_loss": 0.6985487341880798,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7151389502609014
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.6901032234110483,
+    "validation_loss": 0.7164368033409119,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7228159409058066
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.6837317685528499,
+    "validation_loss": 0.6936579346656799,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7228092332549204
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.687675036308242,
+    "validation_loss": 0.6586704850196838,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7046216551479709
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6865368398224435,
+    "validation_loss": 0.6734499335289001,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7203484733365098
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6697868941033759,
+    "validation_loss": 0.7118483781814575,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7191380349608197
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.6508736701273337,
+    "validation_loss": 0.6616345643997192,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7231272175643826
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.6505100578069687,
+    "validation_loss": 0.6596866250038147,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7151389502609014
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6638081597118843,
+    "validation_loss": 0.6713008284568787,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7279401737565508
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6587427185802925,
+    "validation_loss": 0.655285656452179,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7182102678099487
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6385705103961433,
+    "validation_loss": 0.6573081016540527,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7231272175643826
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6348427366919633,
+    "validation_loss": 0.651576578617096,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7288319857577331
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6565009624492831,
+    "validation_loss": 0.6397656798362732,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7188321070054815
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6659962752970253,
+    "validation_loss": 0.6891706585884094,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7309385419512768
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6738467928839893,
+    "validation_loss": 0.6647281646728516,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7263824879672347
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6135960897294487,
+    "validation_loss": 0.6542041897773743,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7170297061311794
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.6420483487408336,
+    "validation_loss": 0.6627779006958008,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7339515372302257
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6316513110224794,
+    "validation_loss": 0.6684327721595764,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7322278512955682
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.5976518806887836,
+    "validation_loss": 0.6376972794532776,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.72072080654124
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.6369955467014778,
+    "validation_loss": 0.6139159798622131,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7241977003359705
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.6086341200805292,
+    "validation_loss": 0.6317619681358337,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7226772044924286
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 1,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 1
+  },
+  "epochs_completed": 50,
+  "best_epoch": 50,
+  "early_stopped": false,
+  "hit_max": true,
+  "train_loss_at_best": 0.6737533146288337,
+  "val_loss_at_best": 0.5473684668540955,
+  "accuracy": 0.771689497716895,
+  "f1_macro": 0.7440047141753251,
+  "f1_weighted": 0.780656107342426,
+  "recall": {
+   "0.0": 0.8275862068965517,
+   "1.0": 0.8484848484848485,
+   "2.0": 0.7272727272727273,
+   "3.0": 0.7889908256880734,
+   "4.0": 0.6216216216216216
+  },
+  "pred_counts": {
+   "0.0": 50,
+   "1.0": 30,
+   "2.0": 12,
+   "3.0": 98,
+   "4.0": 29
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6146349499865276,
+    "validation_loss": 1.575058102607727,
+    "validation_accuracy": 0.1324200913242009,
+    "validation_f1_macro": 0.050216450216450215
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.5576762571567442,
+    "validation_loss": 1.4663180112838745,
+    "validation_accuracy": 0.6712328767123288,
+    "validation_f1_macro": 0.42016638018716657
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.429895084078719,
+    "validation_loss": 1.3014626502990723,
+    "validation_accuracy": 0.5114155251141552,
+    "validation_f1_macro": 0.43583686743324107
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.3091918404509382,
+    "validation_loss": 1.1551529169082642,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.5488767922611446
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.22728247904196,
+    "validation_loss": 1.106502652168274,
+    "validation_accuracy": 0.6757990867579908,
+    "validation_f1_macro": 0.5671942031903903
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.1522581679065054,
+    "validation_loss": 1.0021183490753174,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.641220618738691
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0948621368989713,
+    "validation_loss": 0.9995510578155518,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6484724275824066
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 1.039277656776149,
+    "validation_loss": 0.9134731292724609,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.600579379222993
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9888714748184856,
+    "validation_loss": 0.8895769715309143,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6438500618850422
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.9783786477112189,
+    "validation_loss": 0.8380142450332642,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6358597527533143
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.960785810540362,
+    "validation_loss": 0.789691686630249,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.6705080988917306
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.932400470826684,
+    "validation_loss": 0.7533499598503113,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.6678743728192538
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.9086572571498591,
+    "validation_loss": 0.7680419683456421,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7357015533486121
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8951937149210674,
+    "validation_loss": 0.7899214029312134,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.6859978502715481
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.8932757028719274,
+    "validation_loss": 0.7378162145614624,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7412681571302261
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.8621222921987859,
+    "validation_loss": 0.709632158279419,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.7469195430713926
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.8679965152973081,
+    "validation_loss": 0.700313925743103,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7520373878994568
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.856050679596459,
+    "validation_loss": 0.7888535857200623,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.720227319958807
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.8359041402979595,
+    "validation_loss": 0.7383802533149719,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7502831048969638
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.8181704050157128,
+    "validation_loss": 0.6877511143684387,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7477847298339626
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.8081453064592873,
+    "validation_loss": 0.6738131642341614,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7477847298339626
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.8146184283058818,
+    "validation_loss": 0.6859759092330933,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7519691047333475
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7658585833340157,
+    "validation_loss": 0.6654026508331299,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7481259028885503
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.795221415961661,
+    "validation_loss": 0.6880972385406494,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.749113261531172
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.766509844762523,
+    "validation_loss": 0.6547732353210449,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7504815593738889
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.7451754779350467,
+    "validation_loss": 0.641086757183075,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7513249611234765
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.7445048566271619,
+    "validation_loss": 0.6966652870178223,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7606411893521148
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.7651737048858549,
+    "validation_loss": 0.6241215467453003,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7420460368375397
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.7353100442304844,
+    "validation_loss": 0.6544505953788757,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7657700127712094
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.7035701282140685,
+    "validation_loss": 0.6174139380455017,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7498252638827072
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.7469911822458593,
+    "validation_loss": 0.6883834004402161,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7347811350746591
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.7179534958630074,
+    "validation_loss": 0.6455731391906738,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.732234748765767
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.7100564270484738,
+    "validation_loss": 0.6091746687889099,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7317109445711596
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.7199300287700281,
+    "validation_loss": 0.6026328206062317,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7308664449238883
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6912782046853042,
+    "validation_loss": 0.5964751243591309,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7353358103716783
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.699575291174214,
+    "validation_loss": 0.6650922894477844,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7315696649029981
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.663958967458911,
+    "validation_loss": 0.6142030954360962,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7528734567901234
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.7137848516789879,
+    "validation_loss": 0.6241359710693359,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.752736056214317
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6935832740330115,
+    "validation_loss": 0.6146963238716125,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7642682594117683
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6634499815905966,
+    "validation_loss": 0.647922694683075,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7336896086369771
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6981347810204436,
+    "validation_loss": 0.6087270379066467,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.740196849837119
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6437401011949633,
+    "validation_loss": 0.5964922904968262,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7530886671461106
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6792685098764373,
+    "validation_loss": 0.625365674495697,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7529175812892382
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6364530761067461,
+    "validation_loss": 0.5903960466384888,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.743187921582135
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6544558653017369,
+    "validation_loss": 0.6682825684547424,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.757454120777634
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.6508243360170504,
+    "validation_loss": 0.5978586077690125,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7392423117547007
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6495987329541183,
+    "validation_loss": 0.6385520100593567,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7542611570673247
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.6470511155884441,
+    "validation_loss": 0.6651841998100281,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7437327104909329
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.6453332679300774,
+    "validation_loss": 0.5932057499885559,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7604378496705634
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.6737533146288337,
+    "validation_loss": 0.5473684668540955,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7440047141753251
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 2,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 2
+  },
+  "epochs_completed": 50,
+  "best_epoch": 50,
+  "early_stopped": false,
+  "hit_max": true,
+  "train_loss_at_best": 0.6226128528757793,
+  "val_loss_at_best": 0.6078133583068848,
+  "accuracy": 0.7899543378995434,
+  "f1_macro": 0.7605265274426602,
+  "f1_weighted": 0.8021877366164152,
+  "recall": {
+   "0.0": 0.8620689655172413,
+   "1.0": 0.7878787878787878,
+   "2.0": 0.9090909090909091,
+   "3.0": 0.7981651376146789,
+   "4.0": 0.6756756756756757
+  },
+  "pred_counts": {
+   "0.0": 51,
+   "1.0": 27,
+   "2.0": 17,
+   "3.0": 93,
+   "4.0": 31
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6174468180028403,
+    "validation_loss": 1.546391487121582,
+    "validation_accuracy": 0.4977168949771689,
+    "validation_f1_macro": 0.1329268292682927
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.58593586596047,
+    "validation_loss": 1.4984139204025269,
+    "validation_accuracy": 0.634703196347032,
+    "validation_f1_macro": 0.32264243870355924
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.5026960140321313,
+    "validation_loss": 1.3888118267059326,
+    "validation_accuracy": 0.6301369863013698,
+    "validation_f1_macro": 0.4463465487656915
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.366911018767008,
+    "validation_loss": 1.223652720451355,
+    "validation_accuracy": 0.6666666666666666,
+    "validation_f1_macro": 0.48780167888527026
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.242981534178664,
+    "validation_loss": 1.0726312398910522,
+    "validation_accuracy": 0.6894977168949772,
+    "validation_f1_macro": 0.5486778629605984
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.162133698056384,
+    "validation_loss": 1.0369216203689575,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.5994216476773437
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0751376282878038,
+    "validation_loss": 0.9657370448112488,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.6168606528606528
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 1.0233616610852683,
+    "validation_loss": 0.9052940607070923,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.6206169633347545
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9568443654513941,
+    "validation_loss": 0.8481383919715881,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.6569653120795127
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.965624666795498,
+    "validation_loss": 0.8241379857063293,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.6993045121283534
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.9236650459650086,
+    "validation_loss": 0.7953732013702393,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.6642163006401307
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.9139396679110643,
+    "validation_loss": 0.8159263730049133,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7289687504443403
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8704477693976426,
+    "validation_loss": 0.7864435911178589,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7298061488176467
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8610671446090792,
+    "validation_loss": 0.7811137437820435,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7461521382117662
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.8468370241362874,
+    "validation_loss": 0.7946519255638123,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7148572825024437
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.8112003149055853,
+    "validation_loss": 0.759267270565033,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7438968232086511
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.8073849859761029,
+    "validation_loss": 0.7926062941551208,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7186664392134682
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.816777443740426,
+    "validation_loss": 0.7469311952590942,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7219948399680252
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7756974529929277,
+    "validation_loss": 0.7867042422294617,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7173738004352781
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7760032975092167,
+    "validation_loss": 0.742599606513977,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7142244295830055
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7517595632774073,
+    "validation_loss": 0.7031355500221252,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7442643678160918
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7545412931500411,
+    "validation_loss": 0.6887046694755554,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.7562175932096297
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7387897721151027,
+    "validation_loss": 0.7003820538520813,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7543517355165935
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.735975246603896,
+    "validation_loss": 0.7061943411827087,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7328107884559498
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.7385040648826738,
+    "validation_loss": 0.6894986033439636,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7331689455756886
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.7225778640770331,
+    "validation_loss": 0.6931126713752747,
+    "validation_accuracy": 0.7853881278538812,
+    "validation_f1_macro": 0.7605941350168008
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.7251168932856583,
+    "validation_loss": 0.6966921091079712,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7558511749080189
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.6874374543748251,
+    "validation_loss": 0.6818851232528687,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7543517355165935
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.6856663837665464,
+    "validation_loss": 0.6972566246986389,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7455299037752814
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.7368745691165691,
+    "validation_loss": 0.6822320222854614,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7516292198922165
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.6798700966486116,
+    "validation_loss": 0.6768004298210144,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7219549374996355
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.6843307061893183,
+    "validation_loss": 0.6862162947654724,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7359209675538737
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.6932616771721258,
+    "validation_loss": 0.6602873206138611,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.745427030596522
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6615989964182784,
+    "validation_loss": 0.6448985934257507,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7500281260358693
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6791574282617103,
+    "validation_loss": 0.6585924029350281,
+    "validation_accuracy": 0.7853881278538812,
+    "validation_f1_macro": 0.7664376696199262
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.657524651507052,
+    "validation_loss": 0.6542724967002869,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.7626256130648112
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.6647858172655106,
+    "validation_loss": 0.6971855759620667,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.753978494623656
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6512645035982132,
+    "validation_loss": 0.666414737701416,
+    "validation_accuracy": 0.7853881278538812,
+    "validation_f1_macro": 0.7705946470426236
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6842886165874761,
+    "validation_loss": 0.637172281742096,
+    "validation_accuracy": 0.7853881278538812,
+    "validation_f1_macro": 0.7539143744415602
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6806678484852721,
+    "validation_loss": 0.6234135627746582,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7615367535963815
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.616909401082411,
+    "validation_loss": 0.6653591394424438,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.755696817874246
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6306327792202554,
+    "validation_loss": 0.6591380834579468,
+    "validation_accuracy": 0.7853881278538812,
+    "validation_f1_macro": 0.7742082670405406
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.613016491982995,
+    "validation_loss": 0.6415198445320129,
+    "validation_accuracy": 0.7945205479452054,
+    "validation_f1_macro": 0.7785763733179202
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6375173973601039,
+    "validation_loss": 0.6755697727203369,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.762229243198948
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6442778532097979,
+    "validation_loss": 0.6340303421020508,
+    "validation_accuracy": 0.7808219178082192,
+    "validation_f1_macro": 0.7672196170102448
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.6433215337555583,
+    "validation_loss": 0.6413596272468567,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7579256303497852
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6104130672245491,
+    "validation_loss": 0.691887617111206,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7639407057578872
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.6267564820080269,
+    "validation_loss": 0.6122168302536011,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7422206267279872
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.6220824849314806,
+    "validation_loss": 0.6251174807548523,
+    "validation_accuracy": 0.7945205479452054,
+    "validation_f1_macro": 0.7765977910787982
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.6226128528757793,
+    "validation_loss": 0.6078133583068848,
+    "validation_accuracy": 0.7899543378995434,
+    "validation_f1_macro": 0.7605265274426602
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 3,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 3
+  },
+  "epochs_completed": 50,
+  "best_epoch": 47,
+  "early_stopped": false,
+  "hit_max": true,
+  "train_loss_at_best": 0.6082611465599479,
+  "val_loss_at_best": 0.7235217690467834,
+  "accuracy": 0.730593607305936,
+  "f1_macro": 0.714038944252953,
+  "f1_weighted": 0.7389451956635594,
+  "recall": {
+   "0.0": 0.7931034482758621,
+   "1.0": 0.8181818181818182,
+   "2.0": 0.5454545454545454,
+   "3.0": 0.7431192660550459,
+   "4.0": 0.6216216216216216
+  },
+  "pred_counts": {
+   "0.0": 48,
+   "1.0": 29,
+   "2.0": 6,
+   "3.0": 98,
+   "4.0": 38
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6220814803751504,
+    "validation_loss": 1.5875742435455322,
+    "validation_accuracy": 0.1689497716894977,
+    "validation_f1_macro": 0.058039215686274515
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.5936871476289702,
+    "validation_loss": 1.5093015432357788,
+    "validation_accuracy": 0.5296803652968036,
+    "validation_f1_macro": 0.24350741457124436
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.4826284995893153,
+    "validation_loss": 1.3824553489685059,
+    "validation_accuracy": 0.6529680365296804,
+    "validation_f1_macro": 0.4183523781204423
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.3528232690764637,
+    "validation_loss": 1.1993681192398071,
+    "validation_accuracy": 0.6210045662100456,
+    "validation_f1_macro": 0.37834817436637735
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.2376714392406185,
+    "validation_loss": 1.1297868490219116,
+    "validation_accuracy": 0.680365296803653,
+    "validation_f1_macro": 0.5830366044722973
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.1285978991810868,
+    "validation_loss": 1.022017002105713,
+    "validation_accuracy": 0.6757990867579908,
+    "validation_f1_macro": 0.5483004580309334
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0585624633765802,
+    "validation_loss": 0.9644467234611511,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6055969275814801
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 0.9813799683640643,
+    "validation_loss": 0.9254334568977356,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.6203979879120456
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9255651540872527,
+    "validation_loss": 0.8911700248718262,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6447277343192168
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.8835772624829921,
+    "validation_loss": 0.8481254577636719,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6407901155929054
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.8595382840168185,
+    "validation_loss": 0.8351685404777527,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.6622263514454271
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.8846968368786138,
+    "validation_loss": 0.8464142680168152,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6477420680141102
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8019941928910046,
+    "validation_loss": 0.804776668548584,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.659608964024234
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8261272681922447,
+    "validation_loss": 0.8410500884056091,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.6681086611998096
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.7871945776590487,
+    "validation_loss": 0.7961083054542542,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6631208004473412
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.7758532391815651,
+    "validation_loss": 0.7845243811607361,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.671621371001291
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.7620533690219973,
+    "validation_loss": 0.7771141529083252,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6743542188258713
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.7353853663293327,
+    "validation_loss": 0.7856449484825134,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6742213392095463
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7261699917839795,
+    "validation_loss": 0.7977001070976257,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6888408000002927
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7368398461400009,
+    "validation_loss": 0.7927899360656738,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.692141508174205
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7344355328780848,
+    "validation_loss": 0.7827250957489014,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6867708679747256
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.6790950581794832,
+    "validation_loss": 0.7775569558143616,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6765388587616696
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.6984621402693958,
+    "validation_loss": 0.7734518051147461,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6877153551218909
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.6878889508363677,
+    "validation_loss": 0.7655962705612183,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7023463479684678
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.6925383333752795,
+    "validation_loss": 0.757779061794281,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.6825144295123554
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.6952342565466718,
+    "validation_loss": 0.7567448616027832,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6996234126404904
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.6490353898304265,
+    "validation_loss": 0.7367990016937256,
+    "validation_accuracy": 0.6894977168949772,
+    "validation_f1_macro": 0.6604909727865581
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.6474557097365217,
+    "validation_loss": 0.7714822888374329,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.6954321333750253
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.6501478698195481,
+    "validation_loss": 0.7639806866645813,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.7140885958427835
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.6764153576478725,
+    "validation_loss": 0.7494249939918518,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6996234126404904
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.6499752293272716,
+    "validation_loss": 0.7761468291282654,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.708692259472666
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.6555054213215665,
+    "validation_loss": 0.7684013247489929,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6801009433656993
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.6549598781800852,
+    "validation_loss": 0.7674687504768372,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6881563966843104
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6319407271902736,
+    "validation_loss": 0.7449424266815186,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.693226179799515
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6287453588916034,
+    "validation_loss": 0.7504529356956482,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.693226179799515
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.6294544270125831,
+    "validation_loss": 0.7591035962104797,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.7009088674911956
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.6553425941525436,
+    "validation_loss": 0.768791139125824,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.694211786459072
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6051805818953165,
+    "validation_loss": 0.7681089639663696,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.7162450669231648
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6314349792352537,
+    "validation_loss": 0.7359892725944519,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6996234126404904
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6244277274463235,
+    "validation_loss": 0.7540702223777771,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7048316602752086
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6361175441160435,
+    "validation_loss": 0.7612143754959106,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.7021036366124976
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6066937213990746,
+    "validation_loss": 0.7710081934928894,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7149340559925526
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6121673482220348,
+    "validation_loss": 0.7481098771095276,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7149526054029889
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6385712423702565,
+    "validation_loss": 0.7499731183052063,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.6983818499613152
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6101498901844025,
+    "validation_loss": 0.734427809715271,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6948720961647448
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.5905797506977872,
+    "validation_loss": 0.7421669363975525,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7123222741391517
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6082611465599479,
+    "validation_loss": 0.7235217690467834,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.714038944252953
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.5902958159766546,
+    "validation_loss": 0.779229998588562,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7216398602451733
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.5838393728907515,
+    "validation_loss": 0.7277742624282837,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7205624757048593
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.5906014115345187,
+    "validation_loss": 0.7636589407920837,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7135170432604537
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 7,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 7
+  },
+  "epochs_completed": 44,
+  "best_epoch": 32,
+  "early_stopped": true,
+  "hit_max": false,
+  "train_loss_at_best": 0.661002103875323,
+  "val_loss_at_best": 0.6984432935714722,
+  "accuracy": 0.771689497716895,
+  "f1_macro": 0.7524499679210638,
+  "f1_weighted": 0.7826279144428353,
+  "recall": {
+   "0.0": 0.8620689655172413,
+   "1.0": 0.7878787878787878,
+   "2.0": 0.6363636363636364,
+   "3.0": 0.7614678899082569,
+   "4.0": 0.7567567567567568
+  },
+  "pred_counts": {
+   "0.0": 51,
+   "1.0": 28,
+   "2.0": 8,
+   "3.0": 91,
+   "4.0": 41
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.62655437574154,
+    "validation_loss": 1.589923620223999,
+    "validation_accuracy": 0.5296803652968036,
+    "validation_f1_macro": 0.20888953657825704
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.5845980702376947,
+    "validation_loss": 1.548748254776001,
+    "validation_accuracy": 0.2785388127853881,
+    "validation_f1_macro": 0.21267281105990782
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.468249797821045,
+    "validation_loss": 1.371727466583252,
+    "validation_accuracy": 0.680365296803653,
+    "validation_f1_macro": 0.5202160299834718
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.32800172596443,
+    "validation_loss": 1.2302584648132324,
+    "validation_accuracy": 0.6666666666666666,
+    "validation_f1_macro": 0.5269408837466517
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.1880390469620867,
+    "validation_loss": 1.1226760149002075,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.5486220093958027
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.0946302021422036,
+    "validation_loss": 1.04401433467865,
+    "validation_accuracy": 0.6894977168949772,
+    "validation_f1_macro": 0.5714749984315202
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.030611061468357,
+    "validation_loss": 1.023421287536621,
+    "validation_accuracy": 0.6712328767123288,
+    "validation_f1_macro": 0.5742135730750795
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 0.9878716337971571,
+    "validation_loss": 0.9709606170654297,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.5975834311588126
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9323210745322995,
+    "validation_loss": 0.9185066819190979,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6833981316972961
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.917416339967309,
+    "validation_loss": 0.8980910181999207,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.6809404013460666
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.8819241145762001,
+    "validation_loss": 0.8560897707939148,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.727852427060365
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.8653858725617571,
+    "validation_loss": 0.8336981534957886,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7134181842050694
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8303249274812093,
+    "validation_loss": 0.8161752820014954,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7285636610973818
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8463927158495275,
+    "validation_loss": 0.7870453000068665,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7248277326555709
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.8173226692327639,
+    "validation_loss": 0.8049059510231018,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7151836250752119
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.8112960501414973,
+    "validation_loss": 0.8151602745056152,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7007828498472742
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.7827643555838887,
+    "validation_loss": 0.8070281744003296,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7055984574326514
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.8010357509299022,
+    "validation_loss": 0.7747320532798767,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7083783373985632
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7957447483772184,
+    "validation_loss": 0.7730648517608643,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7056989471814845
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7606793692926082,
+    "validation_loss": 0.756865918636322,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.724972322759208
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7359354161634678,
+    "validation_loss": 0.7705332636833191,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7109910765242468
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7506408742288264,
+    "validation_loss": 0.7497557997703552,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7229705455948657
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7357381807594765,
+    "validation_loss": 0.7553285956382751,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7082830552595245
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.7082513803389014,
+    "validation_loss": 0.75914466381073,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7139834784486362
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.7012432530158903,
+    "validation_loss": 0.7224891185760498,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7516536481181431
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.702656633970214,
+    "validation_loss": 0.7347502708435059,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7160051749244645
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.699109113071023,
+    "validation_loss": 0.7329419851303101,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7380830487817869
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.6665385103807216,
+    "validation_loss": 0.7400558590888977,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7070388770928432
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.6689512576998734,
+    "validation_loss": 0.7360281944274902,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7391559722160727
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.6843016118538089,
+    "validation_loss": 0.7508696913719177,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7082830552595245
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.6739339893910943,
+    "validation_loss": 0.7405543923377991,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7156167475000182
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.661002103875323,
+    "validation_loss": 0.6984432935714722,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7524499679210638
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.6655780258702069,
+    "validation_loss": 0.7283544540405273,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7389977447779943
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.658410701082974,
+    "validation_loss": 0.7165646553039551,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7410359336394337
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6328801788934847,
+    "validation_loss": 0.7204789519309998,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7420996007457028
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.634215030001431,
+    "validation_loss": 0.7153851389884949,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7344943048421657
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.6196055564938522,
+    "validation_loss": 0.7263249158859253,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7404189709107568
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6232949244539913,
+    "validation_loss": 0.7163978219032288,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7450726024380072
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.627199369232829,
+    "validation_loss": 0.7059804201126099,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7450726024380072
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6092390396246096,
+    "validation_loss": 0.703453004360199,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7410359336394337
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6367075861227222,
+    "validation_loss": 0.7127879858016968,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7450726024380072
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6207717867886148,
+    "validation_loss": 0.7062743902206421,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7450726024380072
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.599103124403372,
+    "validation_loss": 0.7054443955421448,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7410359336394337
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.5925045471365858,
+    "validation_loss": 0.7284175157546997,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7344943048421657
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 42,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 42
+  },
+  "epochs_completed": 50,
+  "best_epoch": 38,
+  "early_stopped": true,
+  "hit_max": false,
+  "train_loss_at_best": 0.6739114813688325,
+  "val_loss_at_best": 0.6592805981636047,
+  "accuracy": 0.7671232876712328,
+  "f1_macro": 0.7474258785478297,
+  "f1_weighted": 0.7740214080920172,
+  "recall": {
+   "0.0": 0.7586206896551724,
+   "1.0": 0.9393939393939394,
+   "2.0": 0.9090909090909091,
+   "3.0": 0.7706422018348624,
+   "4.0": 0.5675675675675675
+  },
+  "pred_counts": {
+   "0.0": 46,
+   "1.0": 32,
+   "2.0": 16,
+   "3.0": 96,
+   "4.0": 29
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6413315621817983,
+    "validation_loss": 1.5951603651046753,
+    "validation_accuracy": 0.2054794520547945,
+    "validation_f1_macro": 0.09564573659061848
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.6051898903963042,
+    "validation_loss": 1.5471464395523071,
+    "validation_accuracy": 0.1506849315068493,
+    "validation_f1_macro": 0.05238095238095238
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.541287974613469,
+    "validation_loss": 1.4482790231704712,
+    "validation_accuracy": 0.3744292237442922,
+    "validation_f1_macro": 0.3273361566811023
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.4257210987370188,
+    "validation_loss": 1.3203811645507812,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.5610404919891623
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.2903572379088983,
+    "validation_loss": 1.1790156364440918,
+    "validation_accuracy": 0.6712328767123288,
+    "validation_f1_macro": 0.5303037900262373
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.192301835955643,
+    "validation_loss": 1.1542149782180786,
+    "validation_accuracy": 0.6301369863013698,
+    "validation_f1_macro": 0.5829572914647542
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0858885224272565,
+    "validation_loss": 0.9872011542320251,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6546160686293118
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 1.0201251594031728,
+    "validation_loss": 0.8829815983772278,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.6462817282590345
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9495046894724776,
+    "validation_loss": 0.8631280064582825,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.6743014067193173
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.9342777976175634,
+    "validation_loss": 0.9005348086357117,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.689469101097008
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.9066276433991223,
+    "validation_loss": 0.8241670727729797,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.678359222040761
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.87182052833278,
+    "validation_loss": 0.8306285738945007,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7570054945054945
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8655777348250877,
+    "validation_loss": 0.7712150812149048,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7153631638631638
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.8454430975565096,
+    "validation_loss": 0.7842550277709961,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7185530009472144
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.8387555247399865,
+    "validation_loss": 0.797715425491333,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7263994447222295
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.8214820347181181,
+    "validation_loss": 0.7511270642280579,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.714060392728248
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.8219952263483187,
+    "validation_loss": 0.8021854162216187,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7480525030525029
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.7887809269311952,
+    "validation_loss": 0.7537360787391663,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.747718705511105
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7926808195870098,
+    "validation_loss": 0.7402254343032837,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7275013872674709
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7622463841263841,
+    "validation_loss": 0.7453633546829224,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7481171465278779
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.775284816579121,
+    "validation_loss": 0.7456088662147522,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7482920706416814
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7533575325477414,
+    "validation_loss": 0.7158185243606567,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7533803044226147
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7445236517161857,
+    "validation_loss": 0.7244470119476318,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.761157481482884
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.777114295377964,
+    "validation_loss": 0.7160983085632324,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7669865422393788
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.7019447509835406,
+    "validation_loss": 0.7080726623535156,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7259406795109891
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.7030961208227204,
+    "validation_loss": 0.7158434391021729,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7512252258700489
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.7424630958859514,
+    "validation_loss": 0.7906358242034912,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7316288331342096
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.7209375667862776,
+    "validation_loss": 0.7476458549499512,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7402686202686202
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.69572512696429,
+    "validation_loss": 0.7600176334381104,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7432034632034632
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.7271860489031163,
+    "validation_loss": 0.6967012882232666,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.748410452041456
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.7023421089823653,
+    "validation_loss": 0.6869670152664185,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7708608058608057
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.7001944823962886,
+    "validation_loss": 0.6629666686058044,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7455364218244153
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.6928130410066465,
+    "validation_loss": 0.7037253975868225,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7682012832528295
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6658707871669676,
+    "validation_loss": 0.7267974615097046,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7537267670601004
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6562810256713774,
+    "validation_loss": 0.7219017148017883,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7664590697119064
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.6679604137089195,
+    "validation_loss": 0.7252770662307739,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.757229532482369
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.6854233596383071,
+    "validation_loss": 0.7343176007270813,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.754186366259151
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6739114813688325,
+    "validation_loss": 0.6592805981636047,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7474258785478297
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6822220225159715,
+    "validation_loss": 0.6769421696662903,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7536289821289821
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6873945408478016,
+    "validation_loss": 0.7250295877456665,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7680280026990554
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6764432002858418,
+    "validation_loss": 0.7391007542610168,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7396409969111502
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6339026725873714,
+    "validation_loss": 0.6608799695968628,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7407714164293111
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6590220651975492,
+    "validation_loss": 0.6673893332481384,
+    "validation_accuracy": 0.776255707762557,
+    "validation_f1_macro": 0.7676702673517622
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6262539588823551,
+    "validation_loss": 0.6594874262809753,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.771904365904366
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6377055292449346,
+    "validation_loss": 0.729316771030426,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7633931438869814
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.6469743251800537,
+    "validation_loss": 0.6833896636962891,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7558843937791307
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6347538745257912,
+    "validation_loss": 0.6690486669540405,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7504796029796029
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.6262150170599542,
+    "validation_loss": 0.7333135008811951,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7590531975184693
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.6454702768383956,
+    "validation_loss": 0.6962722539901733,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7579154497562188
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.652672092362148,
+    "validation_loss": 0.7580622434616089,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7515952999180847
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 123,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 123
+  },
+  "epochs_completed": 49,
+  "best_epoch": 37,
+  "early_stopped": true,
+  "hit_max": false,
+  "train_loss_at_best": 0.7013607054221921,
+  "val_loss_at_best": 0.6157943606376648,
+  "accuracy": 0.7625570776255708,
+  "f1_macro": 0.7396509109600558,
+  "f1_weighted": 0.773059563038045,
+  "recall": {
+   "0.0": 0.896551724137931,
+   "1.0": 0.8787878787878788,
+   "2.0": 0.7272727272727273,
+   "3.0": 0.7431192660550459,
+   "4.0": 0.6216216216216216
+  },
+  "pred_counts": {
+   "0.0": 53,
+   "1.0": 29,
+   "2.0": 13,
+   "3.0": 90,
+   "4.0": 34
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.635941365870034,
+    "validation_loss": 1.5491065979003906,
+    "validation_accuracy": 0.4977168949771689,
+    "validation_f1_macro": 0.1329268292682927
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.60081900619879,
+    "validation_loss": 1.565390706062317,
+    "validation_accuracy": 0.2054794520547945,
+    "validation_f1_macro": 0.13772619984264362
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.558504796609646,
+    "validation_loss": 1.463072657585144,
+    "validation_accuracy": 0.3515981735159817,
+    "validation_f1_macro": 0.34616743982970455
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.4567168573053872,
+    "validation_loss": 1.3390640020370483,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.5221375511647885
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.3068824832032366,
+    "validation_loss": 1.1395671367645264,
+    "validation_accuracy": 0.6438356164383562,
+    "validation_f1_macro": 0.38348620210442724
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.2041963615068576,
+    "validation_loss": 1.1338005065917969,
+    "validation_accuracy": 0.639269406392694,
+    "validation_f1_macro": 0.6026584562317626
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.1291042668063467,
+    "validation_loss": 1.0177698135375977,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.6605897494132789
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 1.0272816027083047,
+    "validation_loss": 0.9249964356422424,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.6695071782095134
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.9894306848688823,
+    "validation_loss": 0.8962472677230835,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6633021248201514
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.9454721168773931,
+    "validation_loss": 0.8671642541885376,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6884058429093891
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.9224555572358574,
+    "validation_loss": 0.8263513445854187,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.6688240031872784
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.902957145760699,
+    "validation_loss": 0.8039170503616333,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.6904180841904947
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8497625836511937,
+    "validation_loss": 0.7779726386070251,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7200793198284486
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.835496270075077,
+    "validation_loss": 0.8103938102722168,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6977670831675085
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.8223561884426489,
+    "validation_loss": 0.7685360908508301,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7046996399222168
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.811010141925114,
+    "validation_loss": 0.7409086227416992,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7090524723509385
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.7999161881644551,
+    "validation_loss": 0.7301066517829895,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7120104277129728
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.7460179510639935,
+    "validation_loss": 0.7436771988868713,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7203665088299235
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7962587337668349,
+    "validation_loss": 0.7397180795669556,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7236053752266847
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7818930701511663,
+    "validation_loss": 0.7297154068946838,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7188234593548022
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7805201520280141,
+    "validation_loss": 0.7355466485023499,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7253273117532787
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7384932738978688,
+    "validation_loss": 0.6986443996429443,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7195532673450402
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7141552843698641,
+    "validation_loss": 0.6742230653762817,
+    "validation_accuracy": 0.7671232876712328,
+    "validation_f1_macro": 0.7405008288847317
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.7321291711272263,
+    "validation_loss": 0.6987788081169128,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7281183285945602
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.6965155884986971,
+    "validation_loss": 0.7463865876197815,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7170696172579023
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.6961427418197074,
+    "validation_loss": 0.6954517364501953,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7281183285945602
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.7121596714345421,
+    "validation_loss": 0.6809113621711731,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7431894328050214
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.7088243794150468,
+    "validation_loss": 0.6739152669906616,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.725214910428325
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.7163037883072365,
+    "validation_loss": 0.675110399723053,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7234408217962469
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.6662919306900443,
+    "validation_loss": 0.6735612750053406,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7343850277506163
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.7074004760602626,
+    "validation_loss": 0.6266469955444336,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7166526021057865
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.6581337015803267,
+    "validation_loss": 0.7085683345794678,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7259083059953252
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.7113121275494738,
+    "validation_loss": 0.6482834219932556,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7338258176370583
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6737250231388139,
+    "validation_loss": 0.6657728552818298,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7259177036668325
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6514683213175797,
+    "validation_loss": 0.6538190245628357,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7388930908490514
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.688026835278767,
+    "validation_loss": 0.6805294752120972,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7273117482908698
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.7013607054221921,
+    "validation_loss": 0.6157943606376648,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7396509109600558
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6459702645860067,
+    "validation_loss": 0.6483727693557739,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7292628517453911
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.6444681490339884,
+    "validation_loss": 0.638465166091919,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7423711371070605
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6560473761907438,
+    "validation_loss": 0.6490275263786316,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.733175554993892
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6408006159997568,
+    "validation_loss": 0.6614896059036255,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7379692009517402
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6322115252657634,
+    "validation_loss": 0.6705759167671204,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7258992769847068
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6636554475237684,
+    "validation_loss": 0.6184088587760925,
+    "validation_accuracy": 0.771689497716895,
+    "validation_f1_macro": 0.7516816985948285
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.6606775471350042,
+    "validation_loss": 0.6265592575073242,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7323776838957562
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6194041851090222,
+    "validation_loss": 0.6486321091651917,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.734953370421454
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.6432370637975088,
+    "validation_loss": 0.6816199421882629,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7333772212502159
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6377307420096746,
+    "validation_loss": 0.6541978120803833,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7303845381526105
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.6303909973400396,
+    "validation_loss": 0.6276454329490662,
+    "validation_accuracy": 0.7534246575342466,
+    "validation_f1_macro": 0.7408801972754376
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.61785961215089,
+    "validation_loss": 0.6480539441108704,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7393257736451248
+   }
+  ]
+ },
+ {
+  "limit": 50,
+  "seed": 2024,
+  "hyperparameters": {
+   "epochs": 50,
+   "batch_size": 16,
+   "learning_rate": 0.001,
+   "weight_decay": 0.0001,
+   "dropout": 0.3,
+   "projection_dim": 64,
+   "fusion_dim": 12,
+   "validation_fraction": 0.25,
+   "optimizer": "Adam",
+   "class_weighting": true,
+   "early_stopping_patience": 12,
+   "seed": 2024
+  },
+  "epochs_completed": 50,
+  "best_epoch": 41,
+  "early_stopped": false,
+  "hit_max": true,
+  "train_loss_at_best": 0.6359976275665004,
+  "val_loss_at_best": 0.6597247123718262,
+  "accuracy": 0.7579908675799086,
+  "f1_macro": 0.7372566114501599,
+  "f1_weighted": 0.7695326473497044,
+  "recall": {
+   "0.0": 0.7586206896551724,
+   "1.0": 0.8181818181818182,
+   "2.0": 0.7272727272727273,
+   "3.0": 0.7798165137614679,
+   "4.0": 0.6486486486486487
+  },
+  "pred_counts": {
+   "0.0": 48,
+   "1.0": 29,
+   "2.0": 10,
+   "3.0": 95,
+   "4.0": 37
+  },
+  "collapsed": false,
+  "history": [
+   {
+    "epoch": 1.0,
+    "training_loss": 1.6225681711987752,
+    "validation_loss": 1.511662483215332,
+    "validation_accuracy": 0.4977168949771689,
+    "validation_f1_macro": 0.1329268292682927
+   },
+   {
+    "epoch": 2.0,
+    "training_loss": 1.569906612721885,
+    "validation_loss": 1.5234297513961792,
+    "validation_accuracy": 0.4703196347031963,
+    "validation_f1_macro": 0.2245614035087719
+   },
+   {
+    "epoch": 3.0,
+    "training_loss": 1.4499594322065028,
+    "validation_loss": 1.3053251504898071,
+    "validation_accuracy": 0.6894977168949772,
+    "validation_f1_macro": 0.4256609173759133
+   },
+   {
+    "epoch": 4.0,
+    "training_loss": 1.3003037121237777,
+    "validation_loss": 1.198732852935791,
+    "validation_accuracy": 0.634703196347032,
+    "validation_f1_macro": 0.440913835023424
+   },
+   {
+    "epoch": 5.0,
+    "training_loss": 1.2116903895285072,
+    "validation_loss": 1.080140471458435,
+    "validation_accuracy": 0.684931506849315,
+    "validation_f1_macro": 0.5687377787336579
+   },
+   {
+    "epoch": 6.0,
+    "training_loss": 1.100791285677654,
+    "validation_loss": 1.0347250699996948,
+    "validation_accuracy": 0.6940639269406392,
+    "validation_f1_macro": 0.6177055752611009
+   },
+   {
+    "epoch": 7.0,
+    "training_loss": 1.0073844906760425,
+    "validation_loss": 0.937835693359375,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6832595686116318
+   },
+   {
+    "epoch": 8.0,
+    "training_loss": 0.938835183294808,
+    "validation_loss": 0.9243969917297363,
+    "validation_accuracy": 0.7031963470319634,
+    "validation_f1_macro": 0.6607865027822933
+   },
+   {
+    "epoch": 9.0,
+    "training_loss": 0.8966184040395225,
+    "validation_loss": 0.8663153052330017,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.6923081349030715
+   },
+   {
+    "epoch": 10.0,
+    "training_loss": 0.8589439442971858,
+    "validation_loss": 0.8406924605369568,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7221737682971412
+   },
+   {
+    "epoch": 11.0,
+    "training_loss": 0.85784370870125,
+    "validation_loss": 0.8226235508918762,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7167905712472775
+   },
+   {
+    "epoch": 12.0,
+    "training_loss": 0.8242414804493509,
+    "validation_loss": 0.8313245177268982,
+    "validation_accuracy": 0.6986301369863014,
+    "validation_f1_macro": 0.6600682809617517
+   },
+   {
+    "epoch": 13.0,
+    "training_loss": 0.8054670357122654,
+    "validation_loss": 0.7694411873817444,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7182388720082254
+   },
+   {
+    "epoch": 14.0,
+    "training_loss": 0.78004708450015,
+    "validation_loss": 0.8055667281150818,
+    "validation_accuracy": 0.7077625570776256,
+    "validation_f1_macro": 0.6869753086419753
+   },
+   {
+    "epoch": 15.0,
+    "training_loss": 0.7828267034960956,
+    "validation_loss": 0.7524726390838623,
+    "validation_accuracy": 0.7397260273972602,
+    "validation_f1_macro": 0.7179090965205553
+   },
+   {
+    "epoch": 16.0,
+    "training_loss": 0.7784952884767113,
+    "validation_loss": 0.743364155292511,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7206798541605313
+   },
+   {
+    "epoch": 17.0,
+    "training_loss": 0.7457227278046492,
+    "validation_loss": 0.7856912016868591,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.7093765406198898
+   },
+   {
+    "epoch": 18.0,
+    "training_loss": 0.7468449764135408,
+    "validation_loss": 0.7758557200431824,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.7032627437525654
+   },
+   {
+    "epoch": 19.0,
+    "training_loss": 0.7238862507226991,
+    "validation_loss": 0.7738156914710999,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.707996859273741
+   },
+   {
+    "epoch": 20.0,
+    "training_loss": 0.7094513577658955,
+    "validation_loss": 0.7316175699234009,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7231832224960381
+   },
+   {
+    "epoch": 21.0,
+    "training_loss": 0.7254922986030579,
+    "validation_loss": 0.7338428497314453,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7052727011184234
+   },
+   {
+    "epoch": 22.0,
+    "training_loss": 0.7048084808558952,
+    "validation_loss": 0.7339456677436829,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.7166191562713704
+   },
+   {
+    "epoch": 23.0,
+    "training_loss": 0.7217072335685172,
+    "validation_loss": 0.7096080183982849,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7232759567659875
+   },
+   {
+    "epoch": 24.0,
+    "training_loss": 0.694426145495438,
+    "validation_loss": 0.7198649644851685,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7219951080353464
+   },
+   {
+    "epoch": 25.0,
+    "training_loss": 0.685888970043601,
+    "validation_loss": 0.7141242027282715,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.709853949870283
+   },
+   {
+    "epoch": 26.0,
+    "training_loss": 0.6784876305155638,
+    "validation_loss": 0.717924952507019,
+    "validation_accuracy": 0.7168949771689498,
+    "validation_f1_macro": 0.7044832586768071
+   },
+   {
+    "epoch": 27.0,
+    "training_loss": 0.6826280130118858,
+    "validation_loss": 0.7048062682151794,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7313134228007573
+   },
+   {
+    "epoch": 28.0,
+    "training_loss": 0.6632811921398815,
+    "validation_loss": 0.7309033274650574,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7225617660101258
+   },
+   {
+    "epoch": 29.0,
+    "training_loss": 0.6795863043971178,
+    "validation_loss": 0.6975738406181335,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7100132842068326
+   },
+   {
+    "epoch": 30.0,
+    "training_loss": 0.6437340684053374,
+    "validation_loss": 0.665772020816803,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7328823093731366
+   },
+   {
+    "epoch": 31.0,
+    "training_loss": 0.7039524542122353,
+    "validation_loss": 0.719477117061615,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7113093026126989
+   },
+   {
+    "epoch": 32.0,
+    "training_loss": 0.6489482645581408,
+    "validation_loss": 0.7060401439666748,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.720639788504223
+   },
+   {
+    "epoch": 33.0,
+    "training_loss": 0.6433238772357383,
+    "validation_loss": 0.7341402769088745,
+    "validation_accuracy": 0.7123287671232876,
+    "validation_f1_macro": 0.6899070486620005
+   },
+   {
+    "epoch": 34.0,
+    "training_loss": 0.6565981808958984,
+    "validation_loss": 0.6739374399185181,
+    "validation_accuracy": 0.7625570776255708,
+    "validation_f1_macro": 0.7401719424176123
+   },
+   {
+    "epoch": 35.0,
+    "training_loss": 0.6548550739520933,
+    "validation_loss": 0.7086837291717529,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.7100132842068326
+   },
+   {
+    "epoch": 36.0,
+    "training_loss": 0.602200011291155,
+    "validation_loss": 0.6941585540771484,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7234462506071674
+   },
+   {
+    "epoch": 37.0,
+    "training_loss": 0.63544079007172,
+    "validation_loss": 0.7004677653312683,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.720639788504223
+   },
+   {
+    "epoch": 38.0,
+    "training_loss": 0.6624184563392546,
+    "validation_loss": 0.7010037899017334,
+    "validation_accuracy": 0.726027397260274,
+    "validation_f1_macro": 0.717861527690879
+   },
+   {
+    "epoch": 39.0,
+    "training_loss": 0.5985031287844588,
+    "validation_loss": 0.6666479110717773,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7315324401710119
+   },
+   {
+    "epoch": 40.0,
+    "training_loss": 0.6056617698291453,
+    "validation_loss": 0.6940096020698547,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7360844329718171
+   },
+   {
+    "epoch": 41.0,
+    "training_loss": 0.6359976275665004,
+    "validation_loss": 0.6597247123718262,
+    "validation_accuracy": 0.7579908675799086,
+    "validation_f1_macro": 0.7372566114501599
+   },
+   {
+    "epoch": 42.0,
+    "training_loss": 0.6047317712772183,
+    "validation_loss": 0.6979882717132568,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7124974170733906
+   },
+   {
+    "epoch": 43.0,
+    "training_loss": 0.6072182982433133,
+    "validation_loss": 0.7281728386878967,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.7046568261714561
+   },
+   {
+    "epoch": 44.0,
+    "training_loss": 0.5998183230801326,
+    "validation_loss": 0.6768959760665894,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7169811762776586
+   },
+   {
+    "epoch": 45.0,
+    "training_loss": 0.6171028475935866,
+    "validation_loss": 0.6866343021392822,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.7085102123329972
+   },
+   {
+    "epoch": 46.0,
+    "training_loss": 0.5848244028847392,
+    "validation_loss": 0.678470253944397,
+    "validation_accuracy": 0.7351598173515982,
+    "validation_f1_macro": 0.7195757575757575
+   },
+   {
+    "epoch": 47.0,
+    "training_loss": 0.6237764242218762,
+    "validation_loss": 0.684725821018219,
+    "validation_accuracy": 0.7442922374429224,
+    "validation_f1_macro": 0.7301780448040185
+   },
+   {
+    "epoch": 48.0,
+    "training_loss": 0.5977335670372335,
+    "validation_loss": 0.6758159399032593,
+    "validation_accuracy": 0.730593607305936,
+    "validation_f1_macro": 0.70888995215311
+   },
+   {
+    "epoch": 49.0,
+    "training_loss": 0.6079468349131142,
+    "validation_loss": 0.6623165607452393,
+    "validation_accuracy": 0.7488584474885844,
+    "validation_f1_macro": 0.7466358704586553
+   },
+   {
+    "epoch": 50.0,
+    "training_loss": 0.6073242571295762,
+    "validation_loss": 0.7163891792297363,
+    "validation_accuracy": 0.7214611872146118,
+    "validation_f1_macro": 0.7070384085443462
+   }
+  ]
+ }
+]

diagnostics/2026-10-16-brca-training-collapse/results/epochs_50.log

@@ -0,0 +1,9 @@
+limit=50 seed=0 done: best=49/50 acc=0.763 f1m=0.724
+limit=50 seed=1 done: best=50/50 acc=0.772 f1m=0.744
+limit=50 seed=2 done: best=50/50 acc=0.790 f1m=0.761
+limit=50 seed=3 done: best=47/50 acc=0.731 f1m=0.714
+limit=50 seed=7 done: best=32/44 acc=0.772 f1m=0.752
+limit=50 seed=42 done: best=38/50 acc=0.767 f1m=0.747
+limit=50 seed=123 done: best=37/49 acc=0.763 f1m=0.740
+limit=50 seed=2024 done: best=41/50 acc=0.758 f1m=0.737
+limit=50 ALL DONE

diagnostics/2026-10-16-brca-training-collapse/results/paper_config_sgd.txt

@@ -0,0 +1,4 @@
+paper: SGD lr=0.01 wd=0.01 200ep     acc=0.498 f1_macro=0.133 f1_w=0.331 epochs=17 pred=[0, 0, 0, 219, 0] prob std=0.0000
+paper, no early stop                 acc=0.498 f1_macro=0.133 f1_w=0.331 epochs=200 pred=[0, 0, 0, 219, 0] prob std=0.0000
+
+[exited with code 0]

diagnostics/2026-10-16-brca-training-collapse/results/seed_runs.txt

@@ -0,0 +1,63 @@
+weight_decay=0.000 seed=0 accuracy=0.763 f1_macro=0.724 f1_weighted=0.769 epochs=50 classes_predicted=5 prob_std=0.265 recall_0.0=0.793 recall_1.0=0.818 recall_2.0=0.545 recall_3.0=0.807 recall_4.0=0.622
+weight_decay=0.000 seed=1 accuracy=0.772 f1_macro=0.744 f1_weighted=0.781 epochs=50 classes_predicted=5 prob_std=0.266 recall_0.0=0.828 recall_1.0=0.848 recall_2.0=0.727 recall_3.0=0.789 recall_4.0=0.622
+weight_decay=0.000 seed=2 accuracy=0.790 f1_macro=0.761 f1_weighted=0.802 epochs=50 classes_predicted=5 prob_std=0.278 recall_0.0=0.862 recall_1.0=0.788 recall_2.0=0.909 recall_3.0=0.798 recall_4.0=0.676
+weight_decay=0.000 seed=3 accuracy=0.731 f1_macro=0.714 f1_weighted=0.739 epochs=50 classes_predicted=5 prob_std=0.271 recall_0.0=0.793 recall_1.0=0.818 recall_2.0=0.545 recall_3.0=0.743 recall_4.0=0.622
+weight_decay=0.000 seed=7 accuracy=0.772 f1_macro=0.752 f1_weighted=0.783 epochs=44 classes_predicted=5 prob_std=0.259 recall_0.0=0.862 recall_1.0=0.788 recall_2.0=0.636 recall_3.0=0.761 recall_4.0=0.757
+weight_decay=0.000 seed=42 accuracy=0.767 f1_macro=0.747 f1_weighted=0.774 epochs=50 classes_predicted=5 prob_std=0.270 recall_0.0=0.759 recall_1.0=0.939 recall_2.0=0.909 recall_3.0=0.771 recall_4.0=0.568
+weight_decay=0.000 seed=123 accuracy=0.763 f1_macro=0.740 f1_weighted=0.773 epochs=49 classes_predicted=5 prob_std=0.267 recall_0.0=0.897 recall_1.0=0.879 recall_2.0=0.727 recall_3.0=0.743 recall_4.0=0.622
+weight_decay=0.000 seed=2024 accuracy=0.758 f1_macro=0.737 f1_weighted=0.770 epochs=50 classes_predicted=5 prob_std=0.276 recall_0.0=0.759 recall_1.0=0.818 recall_2.0=0.727 recall_3.0=0.780 recall_4.0=0.649
+weight_decay=0.010 seed=0 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=13 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=1 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=15 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=2 accuracy=0.626 f1_macro=0.539 f1_weighted=0.611 epochs=50 classes_predicted=5 prob_std=0.142 recall_0.0=0.103 recall_1.0=0.879 recall_2.0=0.636 recall_3.0=0.817 recall_4.0=0.243
+weight_decay=0.010 seed=3 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=23 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=7 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=22 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=42 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=31 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=123 accuracy=0.498 f1_macro=0.133 f1_weighted=0.331 epochs=17 classes_predicted=1 prob_std=0.000 recall_0.0=0.000 recall_1.0=0.000 recall_2.0=0.000 recall_3.0=1.000 recall_4.0=0.000
+weight_decay=0.010 seed=2024 accuracy=0.658 f1_macro=0.480 f1_weighted=0.565 epochs=50 classes_predicted=4 prob_std=0.134 recall_0.0=0.000 recall_1.0=0.848 recall_2.0=0.545 recall_3.0=0.982 recall_4.0=0.081
+
+SUMMARY (mean ± std over seeds)
+weight_decay            0.0001  0.0100
+accuracy          mean   0.764   0.534
+                  std    0.017   0.067
+                  min    0.731   0.498
+                  max    0.790   0.658
+f1_macro          mean   0.740   0.227
+                  std    0.015   0.175
+                  min    0.714   0.133
+                  max    0.761   0.539
+f1_weighted       mean   0.774   0.395
+                  std    0.018   0.120
+                  min    0.739   0.331
+                  max    0.802   0.611
+epochs            mean  49.125  27.625
+                  std    2.100  14.889
+                  min   44.000  13.000
+                  max   50.000  50.000
+classes_predicted mean   5.000   1.875
+                  std    0.000   1.642
+                  min    5.000   1.000
+                  max    5.000   5.000
+recall_0.0        mean   0.819   0.013
+                  std    0.051   0.037
+                  min    0.759   0.000
+                  max    0.897   0.103
+recall_1.0        mean   0.837   0.216
+                  std    0.051   0.400
+                  min    0.788   0.000
+                  max    0.939   0.879
+recall_2.0        mean   0.716   0.148
+                  std    0.141   0.275
+                  min    0.545   0.000
+                  max    0.909   0.636
+recall_3.0        mean   0.774   0.975
+                  std    0.024   0.064
+                  min    0.743   0.817
+                  max    0.807   1.000
+recall_4.0        mean   0.642   0.041
+                  std    0.055   0.087
+                  min    0.568   0.000
+                  max    0.757   0.243
+
+collapsed runs (1 class predicted): [[0.01, 0.0], [0.01, 1.0], [0.01, 3.0], [0.01, 7.0], [0.01, 42.0], [0.01, 123.0]]
+
+[exited with code 0]

diagnostics/2026-10-16-brca-training-collapse/results/weight_decay_sweep.txt

@@ -0,0 +1,7 @@
+app default (Adam, wd=0.01)    acc=0.498 f1_macro=0.133 epochs= 31 pred counts=[0, 0, 0, 219, 0] prob std across samples=0.0000
+wd=0                           acc=0.749 f1_macro=0.741 epochs= 44 pred counts=[46, 29, 15, 91, 38] prob std across samples=0.2643
+wd=1e-4                        acc=0.767 f1_macro=0.747 epochs= 50 pred counts=[46, 32, 16, 96, 29] prob std across samples=0.2704
+wd=0, no class weighting       acc=0.772 f1_macro=0.692 epochs= 50 pred counts=[31, 33, 5, 112, 38] prob std across samples=0.2582
+wd=0.01, lr=0.01               acc=0.498 f1_macro=0.133 epochs= 21 pred counts=[0, 0, 0, 219, 0] prob std across samples=0.0000
+
+[exited with code 0]

diagnostics/2026-10-16-brca-training-collapse/scripts/diagnose.py

@@ -0,0 +1,177 @@
+"""Read-only diagnostics for a BioLM-NET training run. Does not modify project code;
+it wraps library functions from the outside to observe what they do."""
+from __future__ import annotations
+
+import pickle
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import torch
+
+OUT = Path(__file__).parent
+DATASET = sys.argv[1] if len(sys.argv) > 1 else "BRCA"
+CLASS_WEIGHTING = (sys.argv[2] if len(sys.argv) > 2 else "1") == "1"
+EPOCHS = int(sys.argv[3]) if len(sys.argv) > 3 else 50
+
+import app  # noqa: E402
+import biolmnet.training as training  # noqa: E402
+from biolmnet.data import read_csv, upstream_example_sources  # noqa: E402
+
+pd.set_option("display.width", 200)
+np.set_printoptions(precision=3, suppress=True, linewidth=160)
+
+
+def section(title):
+    print(f"\n{'=' * 8} {title} {'=' * 8}")
+
+
+# ---------- raw files ----------
+section(f"RAW FILES ({DATASET})")
+for key, url in upstream_example_sources(DATASET).items():
+    frame = read_csv(url)
+    print(f"{key:14s} shape={frame.shape}  first cols={list(frame.columns[:3])}")
+    if key == "labels":
+        print("  raw label counts:", frame.iloc[:, 0].value_counts().to_dict())
+
+# ---------- workspace ----------
+cache = OUT / f"workspace_{DATASET}.pkl"
+if cache.exists():
+    workspace = pickle.loads(cache.read_bytes())
+else:
+    result = app.prepare_workspace(
+        "BioLM-NET examples", DATASET, "", None, None, None, None, None, None, None,
+        True, "Auto — bulk or single-cell based on source", progress=lambda *a, **k: None,
+    )
+    workspace = result[0]
+    if workspace is None:
+        print(result[1]); sys.exit(1)
+    cache.write_bytes(pickle.dumps(workspace))
+
+section("WORKSPACE")
+print("warnings:", workspace.warnings)
+print("samples:", len(workspace.labels), " classes:", workspace.label_names)
+print("full class counts:", dict(zip(workspace.label_names, np.bincount(workspace.labels).tolist())))
+for name, values in (("gene", workspace.gene_expression), ("dna", workspace.dna_methylation)):
+    print(f"{name}: shape={values.shape} min={values.min():.3g} max={values.max():.3g} "
+          f"mean={values.mean():.3g} zero-var cols={(values.std(0) < 1e-8).sum()}")
+
+section("MASKS")
+for name, branch in (("gene", workspace.gene_branch), ("dna", workspace.dna_branch)):
+    m = branch.biological_mask
+    p = branch.pathway_mask
+    print(f"{name}: bio mask {m.shape} density={(m > 0).mean():.4%} nnz={(m > 0).sum():,} "
+          f"PDI={branch.pdi_edges} PPI={branch.ppi_edges}")
+    print(f"   inputs with 0 outgoing edges: {(m.sum(1) == 0).sum()}/{m.shape[0]}   "
+          f"hidden genes in-degree median={np.median((m > 0).sum(0)):.0f} max={(m > 0).sum(0).max()}")
+    print(f"   nonzero mask values: min={m[m > 0].min():.3f} median={np.median(m[m > 0]):.3f} max={m.max():.3f}")
+    print(f"   pathway mask {p.shape}: hidden genes in no pathway={(p.sum(1) == 0).sum()}/{p.shape[0]}  "
+          f"genes/pathway median={np.median(p.sum(0)):.0f}")
+
+# ---------- instrumented training ----------
+hp = training.Hyperparameters(epochs=EPOCHS, class_weighting=CLASS_WEIGHTING)
+section(f"TRAINING  {hp}")
+train_index, validation_index = training.split_train_validation(workspace.labels, hp.validation_fraction, hp.seed)
+names = workspace.label_names
+print("training class counts:  ", dict(zip(names, np.bincount(workspace.labels[train_index], minlength=len(names)).tolist())))
+print("validation class counts:", dict(zip(names, np.bincount(workspace.labels[validation_index], minlength=len(names)).tolist())))
+
+captured = {"criterion_weights": [], "evals": [], "loss_calls": 0}
+orig_ce = training.nn.CrossEntropyLoss
+
+
+class SpyCE(orig_ce):
+    def forward(self, logits, target):
+        captured["loss_calls"] += 1
+        captured["seen_weight"] = None if self.weight is None else self.weight.detach().cpu().numpy()
+        return super().forward(logits, target)
+
+
+def make_spy(*args, **kwargs):
+    loss = SpyCE(*args, **kwargs)
+    captured["criterion_weights"].append(None if loss.weight is None else loss.weight.cpu().numpy())
+    return loss
+
+
+training.nn.CrossEntropyLoss = make_spy
+orig_eval = training._evaluate
+
+
+def spy_eval(model, gene, dna, labels, device):
+    out = orig_eval(model, gene, dna, labels, device)
+    captured["evals"].append((out[1], out[2]))
+    return out
+
+
+training._evaluate = spy_eval
+orig_make = training._make_model
+
+
+def spy_make(*a, **k):
+    model = orig_make(*a, **k)
+    captured["initial_state"] = {k2: v.detach().clone() for k2, v in model.state_dict().items()}
+    return model
+
+
+training._make_model = spy_make
+
+result = training.train(workspace, hp)
+bundle = result.bundle
+
+section("CLASS WEIGHTS")
+print("weights passed to CrossEntropyLoss:", captured["criterion_weights"])
+print("weight attribute seen inside loss.forward:", captured.get("seen_weight"))
+print("loss.forward calls:", captured["loss_calls"])
+
+section("HISTORY")
+print(pd.DataFrame(bundle.history).to_string(index=False))
+
+section("PER-EPOCH VALIDATION PREDICTION COUNTS")
+for i, (probs, preds) in enumerate(captured["evals"][:-1]):
+    counts = np.bincount(preds, minlength=len(names))
+    print(f"epoch {i + 1:3d}: counts={counts.tolist()}  mean max-prob={probs.max(1).mean():.3f}")
+
+section("FINAL (best checkpoint) VALIDATION")
+probs, preds = captured["evals"][-1]
+y_val = workspace.labels[validation_index]
+print("prediction counts:", dict(zip(names, np.bincount(preds, minlength=len(names)).tolist())))
+print("average predicted probability per class:", dict(zip(names, probs.mean(0).round(3).tolist())))
+print("training class prior:", dict(zip(names, (np.bincount(workspace.labels[train_index], minlength=len(names)) / len(train_index)).round(3).tolist())))
+print("std of each class prob across samples:", probs.std(0).round(4).tolist())
+print("max-prob: min/median/max =", np.percentile(probs.max(1), [0, 50, 100]).round(3).tolist())
+sorted_p = np.sort(probs, 1)
+print("margin top1-top2: median =", np.median(sorted_p[:, -1] - sorted_p[:, -2]).round(4))
+print("first 8 validation rows (prob vectors):\n", probs[:8])
+print("metrics:", {k: v for k, v in bundle.metrics.items() if k != "device"})
+print("confusion:\n", result.confusion)
+
+section("LOGITS / ACTIVATIONS (best checkpoint, validation rows)")
+model = bundle.model.eval()
+gene = torch.from_numpy(((workspace.gene_expression[validation_index] - bundle.gene_mean) / bundle.gene_scale).astype(np.float32))
+dna = torch.from_numpy(((workspace.dna_methylation[validation_index] - bundle.dna_mean) / bundle.dna_scale).astype(np.float32))
+with torch.no_grad():
+    stats = {}
+    for name, branch, x in (("gene", model.gene_branch, gene), ("dna", model.dna_branch, dna)):
+        pre = branch.biological(x)
+        hidden = torch.relu(pre)
+        pathways = branch.pathway_attention(hidden)
+        proj = torch.sigmoid(branch.projection(pathways))
+        dead = (hidden.max(0).values == 0).float().mean().item()
+        print(f"{name}: bio pre-act mean={pre.mean():.3g} std={pre.std():.3g}  dead ReLU units={dead:.1%}  "
+              f"pathway std across samples={pathways.std(0).mean():.3g}  "
+              f"sigmoid saturated(<.02|>.98)={((proj < .02) | (proj > .98)).float().mean():.1%}  "
+              f"proj std across samples={proj.std(0).mean():.3g}")
+        stats[name] = proj
+    fused = torch.tanh(model.fusion(torch.cat([stats["gene"], stats["dna"]], 1)))
+    logits = model.output(fused)
+    print(f"fusion tanh saturated(|x|>.98)={(fused.abs() > .98).float().mean():.1%}  fused std across samples={fused.std(0).mean():.3g}")
+    print("logits mean per class:", logits.mean(0).numpy(), " logits std across samples per class:", logits.std(0).numpy())
+
+section("PARAMETER CHANGE FROM INIT")
+for key, value in bundle.model.state_dict().items():
+    init = captured["initial_state"].get(key)
+    if init is None or not value.is_floating_point() or key.endswith("mask") or key.endswith("embeddings"):
+        continue
+    delta = (value - init.cpu()).norm() / (init.norm() + 1e-12)
+    print(f"{key:45s} rel change={delta:.4f}")

diagnostics/2026-10-16-brca-training-collapse/scripts/epochs.py

@@ -0,0 +1,71 @@
+"""Diagnostic only: BRCA, weight_decay=1e-4, one epoch limit across 8 seeds.
+Usage: epochs.py <epoch_limit>. Writes epochs_<limit>.json next to this file.
+Project code is used unchanged; only Hyperparameters fields are overridden."""
+import json
+import pickle
+import sys
+from dataclasses import asdict, replace
+from pathlib import Path
+
+import numpy as np
+from sklearn.metrics import recall_score
+
+import biolmnet.training as training
+
+HERE = Path(__file__).parent
+LIMIT = int(sys.argv[1])
+SEEDS = [0, 1, 2, 3, 7, 42, 123, 2024]
+TOLERANCE = 1e-5  # mirrors train(): improvement must exceed 1e-5
+
+workspace = pickle.loads((HERE / "workspace_BRCA.pkl").read_bytes())
+names = workspace.label_names
+
+
+def replay(history, patience_limit):
+    """Re-run train()'s checkpoint/early-stop bookkeeping on the recorded history."""
+    best_loss, best_epoch, patience, stopped = float("inf"), None, 0, False
+    for row in history:
+        if row["validation_loss"] < best_loss - TOLERANCE:
+            best_loss, best_epoch, patience = row["validation_loss"], int(row["epoch"]), 0
+        else:
+            patience += 1
+            if patience >= patience_limit:
+                stopped = True
+                break
+    return best_epoch, stopped
+
+
+runs = []
+for seed in SEEDS:
+    hp = replace(
+        training.Hyperparameters(),
+        epochs=LIMIT, weight_decay=1e-4, seed=seed,
+        optimizer="Adam", learning_rate=0.001,
+    )
+    result = training.train(workspace, hp)
+    bundle = result.bundle
+    history = bundle.history
+    best_epoch, early_stopped = replay(history, hp.early_stopping_patience)
+    best_row = history[best_epoch - 1]
+    vp = result.validation_predictions
+    recalls = recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0)
+    counts = vp["predicted"].value_counts().reindex(names, fill_value=0).astype(int).tolist()
+    m = bundle.metrics
+    # Sanity: the restored checkpoint's re-evaluated loss must match the best history row.
+    assert abs(m["validation_loss"] - best_row["validation_loss"]) < 1e-4, (seed, m["validation_loss"], best_row)
+    run = {
+        "limit": LIMIT, "seed": seed, "hyperparameters": asdict(hp),
+        "epochs_completed": m["epochs_completed"], "best_epoch": best_epoch,
+        "early_stopped": early_stopped, "hit_max": m["epochs_completed"] == LIMIT and not early_stopped,
+        "train_loss_at_best": best_row["training_loss"], "val_loss_at_best": best_row["validation_loss"],
+        "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"],
+        "recall": dict(zip(names, map(float, recalls))), "pred_counts": dict(zip(names, counts)),
+        "collapsed": int((np.array(counts) > 0).sum()) == 1,
+        "history": history,
+    }
+    runs.append(run)
+    print(f"limit={LIMIT} seed={seed} done: best={best_epoch}/{m['epochs_completed']} "
+          f"acc={m['accuracy']:.3f} f1m={m['f1_macro']:.3f}", flush=True)
+
+(HERE / f"epochs_{LIMIT}.json").write_text(json.dumps(runs, indent=1))
+print(f"limit={LIMIT} ALL DONE", flush=True)

diagnostics/2026-10-16-brca-training-collapse/scripts/epochs_summary.py

@@ -0,0 +1,95 @@
+"""Summarise epochs_{50,100,200}.json produced by epochs.py."""
+import json
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+HERE = Path(__file__).parent
+LIMITS = [50, 100, 200]
+data = {limit: json.loads((HERE / f"epochs_{limit}.json").read_text()) for limit in LIMITS}
+names = list(data[50][0]["recall"])
+pd.set_option("display.width", 250)
+pd.set_option("display.max_columns", 40)
+
+print("=" * 20, "PER RUN")
+rows = []
+for limit in LIMITS:
+    for r in data[limit]:
+        h = pd.DataFrame(r["history"])
+        best = r["best_epoch"]
+        after = h[h.epoch > best]
+        rows.append({
+            "limit": limit, "seed": r["seed"], "done": r["epochs_completed"], "best": best,
+            "early_stop": r["early_stopped"], "hit_max": r["hit_max"],
+            "trL@best": r["train_loss_at_best"], "vaL@best": r["val_loss_at_best"],
+            "acc": r["accuracy"], "f1_mac": r["f1_macro"], "f1_w": r["f1_weighted"],
+            **{f"rec_{n}": r["recall"][n] for n in names},
+            "pred_counts": "/".join(str(r["pred_counts"][n]) for n in names),
+            "collapsed": r["collapsed"],
+            # overfitting signals after the best checkpoint
+            "trL_drop_after_best": (r["train_loss_at_best"] - h.training_loss.iloc[-1]) if len(after) else 0.0,
+            "vaL_rise_after_best": (h.validation_loss.iloc[-1] - r["val_loss_at_best"]) if len(after) else 0.0,
+            "vaL_max_after_best": (after.validation_loss.max() - r["val_loss_at_best"]) if len(after) else 0.0,
+            "f1_last_epoch": h.validation_f1_macro.iloc[-1],
+        })
+runs = pd.DataFrame(rows)
+print(runs.round(3).to_string(index=False))
+runs.to_csv(HERE / "epochs_runs.csv", index=False)
+
+print("\n" + "=" * 20, "SUMMARY PER EPOCH LIMIT (8 seeds)")
+summary = []
+for limit, g in runs.groupby("limit"):
+    summary.append({
+        "limit": limit,
+        "acc": f"{g.acc.mean():.3f} ± {g.acc.std():.3f}",
+        "f1_macro": f"{g.f1_mac.mean():.3f} ± {g.f1_mac.std():.3f}",
+        "f1_weighted": f"{g.f1_w.mean():.3f} ± {g.f1_w.std():.3f}",
+        "f1_macro min/max": f"{g.f1_mac.min():.3f} / {g.f1_mac.max():.3f}",
+        "mean best epoch": f"{g.best.mean():.1f}",
+        "best epoch range": f"{g.best.min()}–{g.best.max()}",
+        "hit max": int(g.hit_max.sum()), "early stopped": int(g.early_stop.sum()),
+        "collapsed": int(g.collapsed.sum()),
+        **{f"recall {n}": f"{g[f'rec_{n}'].mean():.3f}" for n in names},
+    })
+print(pd.DataFrame(summary).set_index("limit").T.to_string())
+
+print("\n" + "=" * 20, "PAIRED CHANGE vs 50 EPOCHS (same seed)")
+base = runs[runs.limit == 50].set_index("seed")
+for limit in (100, 200):
+    other = runs[runs.limit == limit].set_index("seed")
+    for metric in ("acc", "f1_mac", "f1_w"):
+        diff = other[metric] - base[metric]
+        print(f"{limit:3d} vs 50  {metric:7s} mean Δ={diff.mean():+.4f}  std={diff.std():.4f}  "
+              f"better/equal/worse = {(diff > 1e-9).sum()}/{(diff.abs() <= 1e-9).sum()}/{(diff < -1e-9).sum()}  "
+              f"per-seed={diff.round(3).tolist()}")
+other = runs[runs.limit == 200].set_index("seed"); mid = runs[runs.limit == 100].set_index("seed")
+diff = other.f1_mac - mid.f1_mac
+print(f"200 vs 100 f1_mac  mean Δ={diff.mean():+.4f} better/equal/worse = "
+      f"{(diff > 1e-9).sum()}/{(diff.abs() <= 1e-9).sum()}/{(diff < -1e-9).sum()}")
+
+print("\n" + "=" * 20, "OVERFITTING SIGNALS (after the best checkpoint)")
+for limit, g in runs.groupby("limit"):
+    trained_past = g[g.done > g.best]
+    print(f"limit {limit}: runs that trained past best={len(trained_past)}  "
+          f"mean epochs after best={(g.done - g.best).mean():.1f}  "
+          f"mean train-loss drop after best={trained_past.trL_drop_after_best.mean():.3f}  "
+          f"mean val-loss rise (final-best)={trained_past.vaL_rise_after_best.mean():.3f}  "
+          f"runs where train↓ and val↑ after best="
+          f"{int(((trained_past.trL_drop_after_best > 0) & (trained_past.vaL_rise_after_best > 0)).sum())}")
+
+print("\n" + "=" * 20, "MEAN LOSS CURVES (200-epoch runs; mean over seeds still running at that epoch)")
+curves = pd.concat([pd.DataFrame(r["history"]).assign(seed=r["seed"]) for r in data[200]])
+checkpoints = [1, 5, 10, 20, 30, 40, 50, 60, 80, 100, 125, 150, 175, 200]
+table = curves[curves.epoch.isin(checkpoints)].groupby("epoch").agg(
+    runs=("seed", "count"), train_loss=("training_loss", "mean"),
+    val_loss=("validation_loss", "mean"), val_acc=("validation_accuracy", "mean"),
+    val_f1_macro=("validation_f1_macro", "mean"))
+print(table.round(3).to_string())
+
+print("\n" + "=" * 20, "DETERMINISM: shared epochs identical across limits for the same seed?")
+for seed in [r["seed"] for r in data[50]]:
+    hs = {L: pd.DataFrame(next(r for r in data[L] if r["seed"] == seed)["history"]) for L in LIMITS}
+    n = min(len(h) for h in hs.values())
+    same = all(np.allclose(hs[L].validation_loss.iloc[:n], hs[50].validation_loss.iloc[:n], atol=1e-6) for L in LIMITS)
+    print(f"seed {seed}: first {n} epochs identical across 50/100/200 = {same}")

diagnostics/2026-10-16-brca-training-collapse/scripts/seeds.py

@@ -0,0 +1,44 @@
+"""Seed-stability check on BRCA: weight_decay=1e-4 vs the current default 0.01.
+Each seed changes the stratified split, weight init and batch order together.
+Uses project code as-is; only Hyperparameters are overridden."""
+import pickle
+from dataclasses import replace
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from sklearn.metrics import recall_score
+
+import biolmnet.training as training
+
+workspace = pickle.loads((Path(__file__).parent / "workspace_BRCA.pkl").read_bytes())
+names = workspace.label_names
+seeds = [0, 1, 2, 3, 7, 42, 123, 2024]
+rows = []
+for weight_decay in (1e-4, 0.01):
+    for seed in seeds:
+        hp = replace(training.Hyperparameters(), weight_decay=weight_decay, seed=seed)
+        result = training.train(workspace, hp)
+        m = result.bundle.metrics
+        vp = result.validation_predictions
+        probs = vp[[f"P({n})" for n in names]].to_numpy()
+        recalls = recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0)
+        row = {
+            "weight_decay": weight_decay, "seed": seed,
+            "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"],
+            "epochs": m["epochs_completed"],
+            "classes_predicted": int(vp["predicted"].nunique()),
+            "prob_std": float(probs.std(0).mean()),
+            **{f"recall_{n}": r for n, r in zip(names, recalls)},
+        }
+        rows.append(row)
+        print(" ".join(f"{k}={v:.3f}" if isinstance(v, float) else f"{k}={v}" for k, v in row.items()), flush=True)
+
+frame = pd.DataFrame(rows)
+frame.to_csv(Path(__file__).parent / "seed_runs.csv", index=False)
+print("\nSUMMARY (mean ± std over seeds)")
+metrics = ["accuracy", "f1_macro", "f1_weighted", "epochs", "classes_predicted"] + [f"recall_{n}" for n in names]
+summary = frame.groupby("weight_decay")[metrics].agg(["mean", "std", "min", "max"]).T
+pd.set_option("display.width", 200)
+print(summary.round(3).to_string())
+print("\ncollapsed runs (1 class predicted):", frame[frame.classes_predicted == 1][["weight_decay", "seed"]].values.tolist())

diagnostics/2026-10-16-brca-training-collapse/scripts/sweep.py

@@ -0,0 +1,27 @@
+"""Train BRCA under a few optimizer settings and report whether the model collapses."""
+import pickle
+import sys
+from dataclasses import replace
+from pathlib import Path
+
+import numpy as np
+
+import biolmnet.training as training
+
+workspace = pickle.loads((Path(__file__).parent / "workspace_BRCA.pkl").read_bytes())
+base = training.Hyperparameters(epochs=int(sys.argv[1]) if len(sys.argv) > 1 else 50)
+configs = {
+    "app default (Adam, wd=0.01)": base,
+    "wd=0": replace(base, weight_decay=0.0),
+    "wd=1e-4": replace(base, weight_decay=1e-4),
+    "wd=0, no class weighting": replace(base, weight_decay=0.0, class_weighting=False),
+    "wd=0.01, lr=0.01": replace(base, learning_rate=0.01),
+}
+for name, hp in configs.items():
+    result = training.train(workspace, hp)
+    vp = result.validation_predictions
+    probs = vp[[c for c in vp if c.startswith("P(")]].to_numpy()
+    m = result.bundle.metrics
+    counts = vp["predicted"].value_counts().reindex(workspace.label_names, fill_value=0).tolist()
+    print(f"{name:30s} acc={m['accuracy']:.3f} f1_macro={m['f1_macro']:.3f} epochs={m['epochs_completed']:3d} "
+          f"pred counts={counts} prob std across samples={probs.std(0).mean():.4f}", flush=True)

diagnostics/2026-10-16-brca-training-collapse/upstream/BioLMNET.ipynb

@@ -0,0 +1,947 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "634f46b7-f788-499d-b014-7c28905eeff6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "os.environ['PYTHONHASHSEED'] = '0'\n",
+    "os.environ['TF_DETERMINISTIC_OPS'] = '1'\n",
+    "os.environ['CUDA_VISIBLE_DEVICES'] = ''\n",
+    "import random\n",
+    "import numpy as np\n",
+    "import pandas as pd\n",
+    "import itertools\n",
+    "from sklearn.preprocessing import StandardScaler, LabelEncoder\n",
+    "from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\n",
+    "from sklearn.model_selection import train_test_split, StratifiedKFold\n",
+    "import matplotlib.pyplot as plt\n",
+    "import tensorflow as tf\n",
+    "from pathlib import Path\n",
+    "from tensorflow.keras import layers, Model\n",
+    "from tensorflow.keras.layers import Lambda, Dense, Dropout, Input, concatenate, Flatten, BatchNormalization, Activation\n",
+    "from tensorflow.keras import backend as K, regularizers, initializers\n",
+    "from tensorflow.keras.initializers import LecunUniform, GlorotUniform\n",
+    "from tensorflow.keras.regularizers import l2\n",
+    "from joblib import Parallel, delayed\n",
+    "from tqdm import tqdm\n",
+    "from tqdm_joblib import tqdm_joblib\n",
+    "from huggingface_hub import hf_hub_download\n",
+    "tf.config.threading.set_intra_op_parallelism_threads(1)\n",
+    "tf.config.threading.set_inter_op_parallelism_threads(1)\n",
+    "tf.random.set_seed(42)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "535e884e-58a5-434e-95df-0735956321b5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from huggingface_hub import hf_hub_download\n",
+    "\n",
+    "FNAME = \"embedding_associations_cell_type_tissue_drug_pathway_openai_large.parquet\"  \n",
+    "parquet_path = hf_hub_download(\n",
+    "    repo_id   = \"honicky/genept-composable-embeddings\",\n",
+    "    filename  = FNAME,\n",
+    "    repo_type = \"model\"\n",
+    ")\n",
+    "emb = pd.read_parquet(parquet_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1a4d2523-d9a0-466f-bc58-8c815370e9e2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "df_gene = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"Gene_Expression.csv\")\n",
+    "df_dna = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"DNA_Methylation.csv\")\n",
+    " \n",
+    "true_label = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"label.csv\")\n",
+    "\n",
+    "scaler = StandardScaler()\n",
+    "le = LabelEncoder()\n",
+    "le.fit(true_label)\n",
+    "y = le.transform(true_label)\n",
+    "\n",
+    "ge_genes = list(df_gene.columns)\n",
+    "\n",
+    "df_pdi = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"PDI\"/ \"PDI.csv\",sep=',')\n",
+    "df_ge_pdi = df_pdi[df_pdi.Target.isin(ge_genes)]\n",
+    "unique_tfs = df_ge_pdi['TF'].unique()\n",
+    "unique_targets = df_ge_pdi['Target'].unique()\n",
+    "\n",
+    "dna_genes =list(df_dna.columns) \n",
+    "df_dna_pdi = df_pdi[df_pdi.Target.isin(dna_genes)]\n",
+    "\n",
+    "dna_unique_tfs = df_dna_pdi['TF'].unique()\n",
+    "dna_unique_targets = df_dna_pdi['Target'].unique()\n",
+    "\n",
+    "df_ppi = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"PPI\"/ \"PPI.csv\",sep=',')\n",
+    "df_ppi = df_ppi.drop(['Unnamed: 0'], axis =1)\n",
+    "df_ppi['combined_score'] = df_ppi['combined_score'] /1000\n",
+    "threshold = df_ppi['combined_score'].quantile(0.9)\n",
+    "df_ppi = df_ppi[df_ppi['combined_score'] >= threshold]\n",
+    "df_ge_ppi = df_ppi[df_ppi['protein1'].isin(ge_genes) | df_ppi['protein2'].isin(ge_genes)]\n",
+    "unique_protein1 = df_ge_ppi['protein1'].unique()\n",
+    "unique_protein2 = df_ge_ppi['protein2'].unique()\n",
+    "\n",
+    "df_dna_ppi = df_ppi[df_ppi['protein1'].isin(dna_genes) | df_ppi['protein2'].isin(dna_genes)]\n",
+    "dna_unique_protein1 = df_dna_ppi['protein1'].unique()\n",
+    "dna_unique_protein2 = df_dna_ppi['protein2'].unique()\n",
+    "\n",
+    "ge_pdi_ppi_genes = list(set(unique_targets) | set(unique_protein2))\n",
+    "dna_pdi_ppi_genes = list(set(dna_unique_targets) | set(dna_unique_protein2))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e868db94-f1d2-413a-b36f-028cb0d7499d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "genept_gene_emb = emb.loc[emb.index.isin(ge_pdi_ppi_genes)]\n",
+    "ge_pdi_ppi_genept = genept_gene_emb.index"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dfd04c7c-6119-4f60-836e-43d2c2060299",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "genept_dna_emb = emb.loc[emb.index.isin(dna_pdi_ppi_genes)]\n",
+    "dna_pdi_ppi_genept = genept_dna_emb.index"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "537876a2-01e5-4a8f-92df-621469614756",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_ge_mask_matrix (extra_neurons):\n",
+    "    de_gene_to_idx = {gene: i for i, gene in enumerate(ge_genes)}\n",
+    "    target_to_idx = {target: j for j, target in enumerate(unique_targets)}\n",
+    "    protein2_to_idx = {protein2: k for k, protein2 in enumerate(unique_protein2)}\n",
+    "\n",
+    "    pdi_mask_matrix = np.zeros((len(ge_genes), len(unique_targets)), dtype=int)\n",
+    "\n",
+    "    for _, row in df_ge_pdi.iterrows():\n",
+    "        tfa = row[\"TF\"]\n",
+    "        target = row[\"Target\"]\n",
+    "        if tfa in de_gene_to_idx and target in target_to_idx:\n",
+    "            tf_idx = de_gene_to_idx[tfa]\n",
+    "            target_idx = target_to_idx[target]\n",
+    "            pdi_mask_matrix[tf_idx, target_idx] = 1\n",
+    "    \n",
+    "    ppi_mask_matrix = np.zeros((len(ge_genes), len(unique_protein2)))\n",
+    "\n",
+    "    for _, row in df_ge_ppi.iterrows():\n",
+    "        protein1 = row[\"protein1\"]\n",
+    "        protein2 = row[\"protein2\"]\n",
+    "        combined_score = row[\"combined_score\"]\n",
+    "\n",
+    "        if protein1 in de_gene_to_idx and protein2 in protein2_to_idx:\n",
+    "            protein1_idx = de_gene_to_idx[protein1]\n",
+    "            protein2_idx = protein2_to_idx[protein2]\n",
+    "            ppi_mask_matrix[protein1_idx, protein2_idx] = combined_score \n",
+    "            \n",
+    "    pdi_mask_df=     pd.DataFrame(pdi_mask_matrix)\n",
+    "    ppi_mask_df=     pd.DataFrame(ppi_mask_matrix)\n",
+    "    pdi_mask_df.columns = unique_targets\n",
+    "    ppi_mask_df.columns = unique_protein2\n",
+    "    pdi_ppi_mask_df = pd.concat([pdi_mask_df, ppi_mask_df], axis=1)\n",
+    "    pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, ~pdi_ppi_mask_df.columns.duplicated(keep='last')]\n",
+    "    col_mask = pdi_ppi_mask_df.columns.isin(ge_pdi_ppi_genept)\n",
+    "    sub_pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, col_mask]\n",
+    "    \n",
+    "    extra_columns = np.ones((len(ge_genes), extra_neurons), dtype=int)\n",
+    "    adjusted_mask_matrix = np.hstack([sub_pdi_ppi_mask_df, extra_columns])\n",
+    "    new_mask_matrix = tf.constant(adjusted_mask_matrix, dtype=tf.float32)\n",
+    "    \n",
+    "    return new_mask_matrix\n",
+    "\n",
+    "def create_dna_mask_matrix (extra_neurons):\n",
+    "    de_gene_to_idx = {gene: i for i, gene in enumerate(dna_genes)}\n",
+    "    target_to_idx = {target: j for j, target in enumerate(dna_unique_targets)}\n",
+    "    protein2_to_idx = {protein2: k for k, protein2 in enumerate(dna_unique_protein2)}\n",
+    "\n",
+    "\n",
+    "    pdi_mask_matrix = np.zeros((len(dna_genes), len(dna_unique_targets)), dtype=int)\n",
+    "\n",
+    "\n",
+    "    for _, row in df_dna_pdi.iterrows():\n",
+    "        tfa = row[\"TF\"]\n",
+    "        target = row[\"Target\"]\n",
+    "        if tfa in de_gene_to_idx and target in target_to_idx:\n",
+    "            tf_idx = de_gene_to_idx[tfa]\n",
+    "            target_idx = target_to_idx[target]\n",
+    "            pdi_mask_matrix[tf_idx, target_idx] = 1\n",
+    "    \n",
+    "    ppi_mask_matrix = np.zeros((len(dna_genes), len(dna_unique_protein2)))\n",
+    "\n",
+    "    for _, row in df_dna_ppi.iterrows():\n",
+    "        protein1 = row[\"protein1\"]\n",
+    "        protein2 = row[\"protein2\"]\n",
+    "        combined_score = row[\"combined_score\"]\n",
+    "\n",
+    "        if protein1 in de_gene_to_idx and protein2 in protein2_to_idx:\n",
+    "            protein1_idx = de_gene_to_idx[protein1]\n",
+    "            protein2_idx = protein2_to_idx[protein2]\n",
+    "            ppi_mask_matrix[protein1_idx, protein2_idx] = combined_score \n",
+    "            \n",
+    "    pdi_mask_df=     pd.DataFrame(pdi_mask_matrix)\n",
+    "    ppi_mask_df=     pd.DataFrame(ppi_mask_matrix)\n",
+    "    pdi_mask_df.columns = dna_unique_targets\n",
+    "    ppi_mask_df.columns = dna_unique_protein2\n",
+    "    pdi_ppi_mask_df = pd.concat([pdi_mask_df, ppi_mask_df], axis=1)\n",
+    "    pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, ~pdi_ppi_mask_df.columns.duplicated(keep='last')]\n",
+    "    col_mask = pdi_ppi_mask_df.columns.isin(dna_pdi_ppi_genept)\n",
+    "    sub_pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, col_mask]\n",
+    "\n",
+    "    \n",
+    "    extra_columns = np.ones((len(dna_genes), extra_neurons), dtype=int)\n",
+    "    adjusted_mask_matrix = np.hstack([sub_pdi_ppi_mask_df, extra_columns])\n",
+    "    new_mask_matrix = tf.constant(adjusted_mask_matrix, dtype=tf.float32)\n",
+    "    \n",
+    "    return new_mask_matrix"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e5161905-ef7c-4943-aa4d-46a2dc84f498",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ge_target_to_kegg = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"ge_target_to_KEGG_significant.csv\",sep=',')\n",
+    "dna_target_to_kegg = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"dna_target_to_KEGG_significant.csv\",sep=',')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "235fc5c0-0c09-47c7-aa1a-36d708372be1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_ge_pathway_mask_matrix ():\n",
+    "    genes = ge_pdi_ppi_genept\n",
+    "    pathways = ge_target_to_kegg[\"PathwayID\"].unique().tolist()\n",
+    "\n",
+    "    gene_to_idx = {gene: i for i, gene in enumerate(genes)}\n",
+    "    pathway_to_idx = {pw: j for j, pw in enumerate(pathways)}\n",
+    "\n",
+    "    mask_matrix = np.zeros((len(genes), len(pathways)), dtype=int)\n",
+    "\n",
+    "    for _, row in ge_target_to_kegg.iterrows():\n",
+    "        gene = row[\"SYMBOL\"]\n",
+    "        pw = row[\"PathwayID\"]\n",
+    "        if gene in gene_to_idx and pw in pathway_to_idx:\n",
+    "            i = gene_to_idx[gene]\n",
+    "            j = pathway_to_idx[pw]\n",
+    "            mask_matrix[i, j] = 1\n",
+    "    mask_df = pd.DataFrame(mask_matrix, index=genes, columns=pathways)\n",
+    "    return mask_df\n",
+    "\n",
+    "def create_dna_pathway_mask_matrix ():\n",
+    "    genes = dna_pdi_ppi_genept\n",
+    "    pathways = dna_target_to_kegg[\"PathwayID\"].unique().tolist()\n",
+    "\n",
+    "    gene_to_idx = {gene: i for i, gene in enumerate(genes)}\n",
+    "    pathway_to_idx = {pw: j for j, pw in enumerate(pathways)}\n",
+    "\n",
+    "    mask_matrix = np.zeros((len(genes), len(pathways)), dtype=int)\n",
+    "\n",
+    "    for _, row in dna_target_to_kegg.iterrows():\n",
+    "        gene = row[\"SYMBOL\"]\n",
+    "        pw = row[\"PathwayID\"]\n",
+    "        if gene in gene_to_idx and pw in pathway_to_idx:\n",
+    "            i = gene_to_idx[gene]\n",
+    "            j = pathway_to_idx[pw]\n",
+    "            mask_matrix[i, j] = 1\n",
+    "    mask_df = pd.DataFrame(mask_matrix, index=genes, columns=pathways)\n",
+    "    return mask_df\n",
+    "    \n",
+    "mask_gene_pathway  = create_ge_pathway_mask_matrix ()\n",
+    "mask_dna_pathway = create_dna_pathway_mask_matrix()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dbfe572a-d773-4306-94a1-69ab748f924c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class MaskedLayer(layers.Layer):\n",
+    "    def __init__(self, mask_matrix, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "        self.mask_matrix = K.constant(mask_matrix, dtype='float32')\n",
+    "\n",
+    "    def call(self, inputs):\n",
+    "        return K.dot(inputs, self.mask_matrix)\n",
+    "\n",
+    "\n",
+    "class AttentionPathwayLayer(layers.Layer):\n",
+    "    def __init__(self, genept_embs, pathway_mask, init_seed=42, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "\n",
+    "        genept_np = genept_embs.values if hasattr(genept_embs, \"values\") else np.asarray(genept_embs)\n",
+    "        n_genes, D_embed = genept_np.shape\n",
+    "\n",
+    "        pw_np = pathway_mask.values if hasattr(pathway_mask, \"values\") else np.asarray(pathway_mask)\n",
+    "        n_pathways = pw_np.shape[1]\n",
+    "\n",
+    "        self.G = K.constant(genept_np, dtype=\"float32\")\n",
+    "        self.P = K.constant(pw_np, dtype=\"float32\")\n",
+    "\n",
+    "        glorot = initializers.GlorotUniform(seed=init_seed)\n",
+    "        self.W = self.add_weight(\n",
+    "            shape=(n_pathways, D_embed),\n",
+    "            initializer=glorot,\n",
+    "            trainable=True,\n",
+    "            name=\"W_query\"\n",
+    "        )\n",
+    "\n",
+    "    def call(self, x):\n",
+    "        G_T      = K.transpose(self.G)            \n",
+    "        S        = K.dot(self.W, G_T)             \n",
+    "        S_T      = K.transpose(S)                 \n",
+    "        S_masked = S_T * self.P                   \n",
+    "        alpha    = K.softmax(S_masked, axis=0)    \n",
+    "        x_sq     = K.squeeze(x, axis=-1)          \n",
+    "        out_mat  = K.dot(x_sq, alpha)             \n",
+    "        return K.expand_dims(out_mat, axis=-1)    \n",
+    "\n",
+    "    def compute_output_shape(self, input_shape):\n",
+    "        batch_size, n_genes, _ = input_shape\n",
+    "        n_pathways = K.int_shape(self.P)[1]\n",
+    "        return (batch_size, n_pathways, 1)\n",
+    "    \n",
+    "    def get_attention(self):\n",
+    "        G_T      = K.transpose(self.G)            \n",
+    "        S        = K.dot(self.W, G_T)             \n",
+    "        S_T      = K.transpose(S)                 \n",
+    "        S_masked = S_T * self.P                   \n",
+    "        return K.softmax(S_masked, axis=0)      \n",
+    "\n",
+    "\n",
+    "def create_model(\n",
+    "    num_ex_neuron,\n",
+    "    layer_one,\n",
+    "    activation_function_one,\n",
+    "    activation_function_two,\n",
+    "    activation_function_three,\n",
+    "    dropout_p,\n",
+    "    learning_rate,\n",
+    "    l2_reg,\n",
+    "    optimizer_choice,\n",
+    "    genept_gene_emb,\n",
+    "    mask_gene_pathway,\n",
+    "    genept_dna_emb,\n",
+    "    mask_dna_pathway\n",
+    "):\n",
+    "\n",
+    "    new_ge_mask_matrix = create_ge_mask_matrix(num_ex_neuron)\n",
+    "    new_dna_mask_matrix = create_dna_mask_matrix(num_ex_neuron)\n",
+    "    \n",
+    "    inputA = Input(shape=(df_gene.shape[1],), name=\"gene_input\")\n",
+    "    x = layers.Dense(\n",
+    "        new_ge_mask_matrix.shape[0],\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42)\n",
+    "    )(inputA)\n",
+    "    x = MaskedLayer(new_ge_mask_matrix)(x)\n",
+    "    x = Dropout(dropout_p)(x)\n",
+    "    x_feat = Lambda(lambda t: K.expand_dims(t, -1), name=\"gene_add_dim\")(x)\n",
+    "\n",
+    "    gene_path_layer = AttentionPathwayLayer(\n",
+    "    genept_embs  = genept_gene_emb,\n",
+    "    pathway_mask = mask_gene_pathway,\n",
+    "    name         = \"gene_path_attn\"\n",
+    "    )\n",
+    "    x_path = gene_path_layer(x_feat)\n",
+    "    x_path = Flatten()(x_path)\n",
+    "\n",
+    "    x_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42),\n",
+    "        name=\"gene_proj\"\n",
+    "    )(x_path)\n",
+    "\n",
+    "    inputB = Input(shape=(df_dna.shape[1],), name=\"dna_input\")\n",
+    "    y = layers.Dense(\n",
+    "        new_dna_mask_matrix.shape[0],\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42)\n",
+    "    )(inputB)\n",
+    "    y = MaskedLayer(new_dna_mask_matrix)(y)\n",
+    "    y = Dropout(dropout_p)(y)\n",
+    "    y_feat = Lambda(lambda t: K.expand_dims(t, -1), name=\"dna_add_dim\")(y)\n",
+    "\n",
+    "    dna_path_layer = AttentionPathwayLayer(\n",
+    "    genept_embs  = genept_dna_emb,\n",
+    "    pathway_mask = mask_dna_pathway,\n",
+    "    name         = \"dna_path_attn\"\n",
+    "    )\n",
+    "    y_path = dna_path_layer(y_feat)\n",
+    "    y_path = Flatten()(y_path)\n",
+    "\n",
+    "    y_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42),\n",
+    "        name=\"dna_proj\"\n",
+    "    )(y_path)\n",
+    "\n",
+    "    merged = concatenate([x_proj, y_proj], name=\"merge_xy\")\n",
+    "    h = layers.Dense(\n",
+    "        12,\n",
+    "        activation=activation_function_three,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.GlorotUniform(42),\n",
+    "        name=\"fusion_dense\"\n",
+    "    )(merged)\n",
+    "    h = Dropout(dropout_p, name=\"drop_fusion\")(h)\n",
+    "    out = layers.Dense(\n",
+    "        4,\n",
+    "        activation=\"softmax\",\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.GlorotUniform(42),\n",
+    "        name=\"output\"\n",
+    "    )(h)\n",
+    "\n",
+    "    model = Model(inputs=[inputA, inputB], outputs=out, name=\"PathwayModel\")\n",
+    "\n",
+    "    if optimizer_choice == 'adam':\n",
+    "        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
+    "    elif optimizer_choice == 'sgd':\n",
+    "        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate)\n",
+    "    model.compile(\n",
+    "        optimizer=optimizer,\n",
+    "        loss='sparse_categorical_crossentropy',\n",
+    "        metrics=['accuracy']\n",
+    "    )\n",
+    "    return model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "be89b57c-5f26-4b58-be38-7d4c797d4c6d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def evaluate_model_cv(X_train, X_val, y_train, y_val, params):\n",
+    "    model = create_model(\n",
+    "        num_ex_neuron=params['num_ex_neuron'],\n",
+    "        layer_one=params['layer_one'],\n",
+    "        activation_function_one=params['activation_function_one'],\n",
+    "        activation_function_two=params['activation_function_two'],\n",
+    "        activation_function_three=params['activation_function_three'],\n",
+    "        dropout_p=params['dropout_p'],\n",
+    "        learning_rate=params['learning_rate'],\n",
+    "        l2_reg=params['l2_reg'],\n",
+    "        optimizer_choice=params['optimizer'],\n",
+    "        genept_gene_emb       = genept_gene_emb,\n",
+    "        mask_gene_pathway     = mask_gene_pathway,\n",
+    "        genept_dna_emb        = genept_dna_emb,\n",
+    "        mask_dna_pathway      = mask_dna_pathway\n",
+    "    )\n",
+    "    \n",
+    "    history = model.fit(\n",
+    "        X_train, y_train,\n",
+    "        validation_data=(X_val, y_val),\n",
+    "        epochs=params['epoch'],\n",
+    "        batch_size=params['batch_size'],\n",
+    "        shuffle = False,\n",
+    "        verbose=0\n",
+    "    )\n",
+    "    \n",
+    "    y_pred = model.predict(X_val).argmax(1)\n",
+    "    \n",
+    "\n",
+    "    f1_weighted = f1_score(y_val, y_pred, average='weighted')\n",
+    "    f1_macro = f1_score(y_val, y_pred, average='macro')\n",
+    "    f1_micro = f1_score(y_val, y_pred, average='micro')\n",
+    "\n",
+    "    precision_weighted = precision_score(y_val, y_pred, average='weighted')\n",
+    "    precision_macro = precision_score(y_val, y_pred, average='macro')\n",
+    "    precision_micro = precision_score(y_val, y_pred, average='micro')\n",
+    "    \n",
+    "\n",
+    "    recall_weighted = recall_score(y_val, y_pred, average='weighted')\n",
+    "    recall_macro = recall_score(y_val, y_pred, average='macro')\n",
+    "    recall_micro = recall_score(y_val, y_pred, average='micro')\n",
+    "    \n",
+    "\n",
+    "    accuracy = accuracy_score(y_val, y_pred)\n",
+    "\n",
+    "    return {\n",
+    "        'history': history.history,\n",
+    "        'f1_weighted': f1_weighted,\n",
+    "        'f1_macro': f1_macro,\n",
+    "        'f1_micro': f1_micro,\n",
+    "        'precision_weighted': precision_weighted,\n",
+    "        'precision_macro': precision_macro,\n",
+    "        'precision_micro': precision_micro,\n",
+    "        'recall_weighted': recall_weighted,\n",
+    "        'recall_macro': recall_macro,\n",
+    "        'recall_micro': recall_micro,\n",
+    "        'accuracy': accuracy\n",
+    "    }\n",
+    "\n",
+    "def evaluate_combination(param_combination,X_gene_train,X_dna_train,y_train,n_splits=5):\n",
+    "    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)\n",
+    "    \n",
+    "    f1_weighted_list = []\n",
+    "    f1_macro_list = []\n",
+    "    f1_micro_list = []\n",
+    "    \n",
+    "    precision_weighted_list = []\n",
+    "    precision_macro_list = []\n",
+    "    precision_micro_list = []\n",
+    "    \n",
+    "    recall_weighted_list = []\n",
+    "    recall_macro_list = []\n",
+    "    recall_micro_list = []\n",
+    "    \n",
+    "    accuracy_list = []\n",
+    "    all_histories = []\n",
+    "    \n",
+    "    for train_idx, val_idx in skf.split(X_gene_train, y_train):\n",
+    "        Xg_tr, Xg_val = X_gene_train[train_idx], X_gene_train[val_idx]\n",
+    "        Xd_tr, Xd_val = X_dna_train[train_idx], X_dna_train[val_idx]\n",
+    "        y_tr, y_val   = y_train[train_idx], y_train[val_idx]\n",
+    "    \n",
+    "        scores = evaluate_model_cv([Xg_tr, Xd_tr], [Xg_val, Xd_val], y_tr, y_val, param_combination)\n",
+    "        \n",
+    "        history_dict = scores.pop('history')\n",
+    "        all_histories.append(history_dict)\n",
+    "        \n",
+    "        f1_weighted_list.append(scores['f1_weighted'])\n",
+    "        f1_macro_list.append(scores['f1_macro'])\n",
+    "        f1_micro_list.append(scores['f1_micro'])\n",
+    "        \n",
+    "        precision_weighted_list.append(scores['precision_weighted'])\n",
+    "        precision_macro_list.append(scores['precision_macro'])\n",
+    "        precision_micro_list.append(scores['precision_micro'])\n",
+    "        \n",
+    "        recall_weighted_list.append(scores['recall_weighted'])\n",
+    "        recall_macro_list.append(scores['recall_macro'])\n",
+    "        recall_micro_list.append(scores['recall_micro'])\n",
+    "        \n",
+    "        accuracy_list.append(scores['accuracy'])\n",
+    "    \n",
+    "    results = {\n",
+    "        'f1_weighted_mean': np.mean(f1_weighted_list),\n",
+    "        'f1_weighted_std': np.std(f1_weighted_list),\n",
+    "        'f1_macro_mean': np.mean(f1_macro_list),\n",
+    "        'f1_macro_std': np.std(f1_macro_list),\n",
+    "        'f1_micro_mean': np.mean(f1_micro_list),\n",
+    "        'f1_micro_std': np.std(f1_micro_list),\n",
+    "        'precision_weighted_mean': np.mean(precision_weighted_list),\n",
+    "        'precision_weighted_std': np.std(precision_weighted_list),\n",
+    "        'precision_macro_mean': np.mean(precision_macro_list),\n",
+    "        'precision_macro_std': np.std(precision_macro_list),\n",
+    "        'precision_micro_mean': np.mean(precision_micro_list),\n",
+    "        'precision_micro_std': np.std(precision_micro_list),\n",
+    "        'recall_weighted_mean': np.mean(recall_weighted_list),\n",
+    "        'recall_weighted_std': np.std(recall_weighted_list),\n",
+    "        'recall_macro_mean': np.mean(recall_macro_list),\n",
+    "        'recall_macro_std': np.std(recall_macro_list),\n",
+    "        'recall_micro_mean': np.mean(recall_micro_list),\n",
+    "        'recall_micro_std': np.std(recall_micro_list),\n",
+    "        'accuracy_mean': np.mean(accuracy_list),\n",
+    "        'accuracy_std': np.std(accuracy_list)\n",
+    "    }\n",
+    "    \n",
+    "    return param_combination, results, all_histories\n",
+    "\n",
+    "\n",
+    "params_grid = {\n",
+    "     'num_ex_neuron': [0],\n",
+    "     'layer_one': [64],\n",
+    "     'activation_function_one': ['relu'],\n",
+    "     'activation_function_two': ['sigmoid'],\n",
+    "     'activation_function_three': ['tanh'],\n",
+    "     'dropout_p': [0.3],\n",
+    "     'learning_rate': [0.01],\n",
+    "    'l2_reg': [0.01],\n",
+    "    'epoch': [200],\n",
+    "    'batch_size':[16],\n",
+    "    'optimizer': ['sgd']\n",
+    " }\n",
+    "\n",
+    "param_combinations = list(itertools.product(\n",
+    "    params_grid['num_ex_neuron'],\n",
+    "    params_grid['layer_one'],\n",
+    "    params_grid['activation_function_one'],\n",
+    "    params_grid['activation_function_two'],\n",
+    "    params_grid['activation_function_three'],\n",
+    "    params_grid['dropout_p'],\n",
+    "    params_grid['learning_rate'],\n",
+    "    params_grid['l2_reg'],\n",
+    "    params_grid['epoch'],\n",
+    "    params_grid['batch_size'],\n",
+    "    params_grid['optimizer']\n",
+    "    \n",
+    "))\n",
+    "\n",
+    "\n",
+    "param_dicts = []\n",
+    "for comb in param_combinations:\n",
+    "    param_dicts.append({\n",
+    "        'num_ex_neuron': comb[0],\n",
+    "        'layer_one': comb[1],\n",
+    "        'activation_function_one': comb[2],\n",
+    "        'activation_function_two': comb[3],\n",
+    "        'activation_function_three': comb[4],\n",
+    "        'dropout_p': comb[5],\n",
+    "        'learning_rate': comb[6],\n",
+    "        'l2_reg': comb[7],\n",
+    "        'epoch': comb[8],\n",
+    "        'batch_size': comb[9],\n",
+    "        'optimizer': comb[10]\n",
+    "    })\n",
+    "\n",
+    "\n",
+    "X_gene_train, X_gene_test, X_dna_train, X_dna_test, y_train, y_test = train_test_split(df_gene, df_dna, y, test_size=0.2, stratify=y, random_state=42)\n",
+    "scaler = StandardScaler()\n",
+    "X_gene_train=(np.array(X_gene_train).astype(np.float32))\n",
+    "scaler.fit(X_gene_train)\n",
+    "X_gene_train=scaler.transform(X_gene_train)\n",
+    "\n",
+    "X_gene_test=(np.array(X_gene_test).astype(np.float32))\n",
+    "scaler.fit(X_gene_test)\n",
+    "X_gene_test=scaler.transform(X_gene_test)\n",
+    "\n",
+    "X_dna_train=(np.array(X_dna_train).astype(np.float32))\n",
+    "scaler.fit(X_dna_train)\n",
+    "X_dna_train=scaler.transform(X_dna_train)\n",
+    "\n",
+    "X_dna_test=(np.array(X_dna_test).astype(np.float32))\n",
+    "scaler.fit(X_dna_test)\n",
+    "X_dna_test=scaler.transform(X_dna_test)\n",
+    "\n",
+    "with tqdm_joblib(\n",
+    "        tqdm(desc=\"Evaluating hyperparameter combinations\",\n",
+    "             total=len(param_dicts))\n",
+    "     ) as progress_bar:\n",
+    "    results = Parallel(n_jobs=-1)(\n",
+    "        delayed(evaluate_combination)(\n",
+    "            param, X_gene_train, X_dna_train, y_train, n_splits=5\n",
+    "        )\n",
+    "        for param in param_dicts\n",
+    "    )\n",
+    "\n",
+    "results_list = []\n",
+    "histories_dict = {}\n",
+    "best_param_weighted = None\n",
+    "best_weighted_scores = None\n",
+    "best_avg_f1_weighted = -np.inf\n",
+    "\n",
+    "best_param_macro = None\n",
+    "best_macro_scores = None\n",
+    "best_avg_f1_macro = -np.inf\n",
+    "\n",
+    "best_param_micro = None\n",
+    "best_micro_scores = None\n",
+    "best_avg_f1_micro = -np.inf\n",
+    "\n",
+    "for param_dict, scores_dict, histories in results:\n",
+    "    key = str(param_dict)\n",
+    "    histories_dict[key] = histories\n",
+    "\n",
+    "    row = {}\n",
+    "    row.update(param_dict)\n",
+    "    row.update(scores_dict)\n",
+    "    results_list.append(row)\n",
+    "\n",
+    "    if scores_dict['f1_weighted_mean'] > best_avg_f1_weighted:\n",
+    "        best_avg_f1_weighted = scores_dict['f1_weighted_mean']\n",
+    "        best_param_weighted  = param_dict\n",
+    "        best_weighted_scores = scores_dict\n",
+    "\n",
+    "    if scores_dict['f1_macro_mean'] > best_avg_f1_macro:\n",
+    "        best_avg_f1_macro = scores_dict['f1_macro_mean']\n",
+    "        best_param_macro  = param_dict\n",
+    "        best_macro_scores = scores_dict\n",
+    "\n",
+    "    if scores_dict['f1_micro_mean'] > best_avg_f1_micro:\n",
+    "        best_avg_f1_micro = scores_dict['f1_micro_mean']\n",
+    "        best_param_micro  = param_dict\n",
+    "        best_micro_scores = scores_dict\n",
+    "\n",
+    "results_df = pd.DataFrame(results_list)\n",
+    "best_key = str(best_param_macro)\n",
+    "best_histories = histories_dict[best_key]  \n",
+    "\n",
+    "\n",
+    "hyperparam_cols = list(param_dict.keys())\n",
+    "\n",
+    "\n",
+    "weighted_cols = hyperparam_cols + ['f1_weighted_mean', 'f1_weighted_std',\n",
+    "                                     'precision_weighted_mean', 'precision_weighted_std',\n",
+    "                                     'recall_weighted_mean', 'recall_weighted_std',\n",
+    "                                     'accuracy_mean', 'accuracy_std']\n",
+    "results_df_weighted = results_df[weighted_cols]\n",
+    "\n",
+    "macro_cols = hyperparam_cols + ['f1_macro_mean', 'f1_macro_std',\n",
+    "                                  'precision_macro_mean', 'precision_macro_std',\n",
+    "                                  'recall_macro_mean', 'recall_macro_std',\n",
+    "                                  'accuracy_mean', 'accuracy_std']\n",
+    "results_df_macro = results_df[macro_cols]\n",
+    "\n",
+    "micro_cols = hyperparam_cols + ['f1_micro_mean', 'f1_micro_std',\n",
+    "                                  'precision_micro_mean', 'precision_micro_std',\n",
+    "                                  'recall_micro_mean', 'recall_micro_std',\n",
+    "                                  'accuracy_mean', 'accuracy_std']\n",
+    "results_df_micro = results_df[micro_cols]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2576c647-0fa9-44ff-b26b-b152a8f96389",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_model_test(\n",
+    "    num_ex_neuron,\n",
+    "    layer_one,\n",
+    "    activation_function_one,\n",
+    "    activation_function_two,\n",
+    "    activation_function_three,\n",
+    "    dropout_p,\n",
+    "    learning_rate,\n",
+    "    l2_reg,\n",
+    "    optimizer_choice\n",
+    "):\n",
+    "\n",
+    "    new_ge_mask_matrix = create_ge_mask_matrix(num_ex_neuron)    \n",
+    "    new_dna_mask_matrix = create_dna_mask_matrix(num_ex_neuron)  \n",
+    "\n",
+    "    inputA = Input(shape=(df_gene.shape[1],), name=\"gene_input\") \n",
+    "    inputB = Input(shape=(df_dna.shape[1],), name=\"dna_input\")   \n",
+    "\n",
+    "    x1 = layers.Dense(\n",
+    "        new_ge_mask_matrix.shape[0],               \n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=l2(l2_reg)\n",
+    "    )(inputA)                                        \n",
+    "\n",
+    "\n",
+    "    x1_masked = MaskedLayer(new_ge_mask_matrix)(x1) \n",
+    "    x1_masked = Dropout(dropout_p)(x1_masked)       \n",
+    "\n",
+    "    x1_feat = layers.Lambda(lambda t: K.expand_dims(t, -1), name=\"gene_add_dim\")(x1_masked)\n",
+    "\n",
+    "    gene_path_layer = AttentionPathwayLayer(\n",
+    "        genept_embs  = genept_gene_emb,    \n",
+    "        pathway_mask = mask_gene_pathway,  \n",
+    "        name         = \"gene_path_attn\"\n",
+    "    )\n",
+    "    x1_path = gene_path_layer(x1_feat)     \n",
+    "\n",
+    "    x1_path = Flatten(name=\"gene_flat_path\")(x1_path)  \n",
+    "\n",
+    "    x1_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"gene_proj\"\n",
+    "    )(x1_path)                                   \n",
+    "\n",
+    "    y1 = layers.Dense(\n",
+    "        new_dna_mask_matrix.shape[0],               \n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=l2(l2_reg)\n",
+    "    )(inputB)                                     \n",
+    "\n",
+    "    y1_masked = MaskedLayer(new_dna_mask_matrix)(y1)  \n",
+    "    y1_masked = Dropout(dropout_p)(y1_masked)         \n",
+    "\n",
+    "    y1_feat = layers.Lambda(lambda t: K.expand_dims(t, -1), name=\"dna_add_dim\")(y1_masked)\n",
+    "\n",
+    "    dna_path_layer = AttentionPathwayLayer(\n",
+    "        genept_embs  = genept_dna_emb,      \n",
+    "        pathway_mask = mask_dna_pathway,    \n",
+    "        name         = \"dna_path_attn\"\n",
+    "    )\n",
+    "    y1_path = dna_path_layer(y1_feat) \n",
+    "    y1_path = Flatten(name=\"dna_flat_path\")(y1_path)   \n",
+    "\n",
+    "    y1_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"dna_proj\"\n",
+    "    )(y1_path)                                    \n",
+    "\n",
+    "    merged = concatenate([x1_proj, y1_proj], name=\"merge_xy\")  \n",
+    "\n",
+    "    h = layers.Dense(\n",
+    "        12,\n",
+    "        activation=activation_function_three,\n",
+    "        kernel_initializer=GlorotUniform(seed=42),\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"fusion_dense\"\n",
+    "    )(merged)                                        \n",
+    "    h = Dropout(dropout_p, name=\"drop_fusion\")(h)    \n",
+    "\n",
+    "    outputs = layers.Dense(\n",
+    "        4,\n",
+    "        activation=\"softmax\",\n",
+    "        kernel_initializer=GlorotUniform(seed=42),\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"output\"\n",
+    "    )(h)                                             \n",
+    "\n",
+    "    model = Model(inputs=[inputA, inputB], outputs=outputs, name=\"PathwayModel_Test\")\n",
+    "\n",
+    "    if optimizer_choice == 'adam':\n",
+    "        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
+    "    else:\n",
+    "        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate)\n",
+    "\n",
+    "    model.compile(\n",
+    "        optimizer=optimizer,\n",
+    "        loss='sparse_categorical_crossentropy',\n",
+    "        metrics=['accuracy']\n",
+    "    )\n",
+    "    return model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fcda02d6-b3bd-4028-ba48-b0095d80245a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def evaluate_test_model(best_param, avg_type, save_csv_path=None):\n",
+    "   \n",
+    "    per_run_results = []\n",
+    "\n",
+    "    for run_i in range(1,11):\n",
+    "       \n",
+    "        model = create_model_test(\n",
+    "            num_ex_neuron           = best_param['num_ex_neuron'],\n",
+    "            layer_one               = best_param['layer_one'],\n",
+    "            activation_function_one = best_param['activation_function_one'],\n",
+    "            activation_function_two = best_param['activation_function_two'],\n",
+    "            activation_function_three = best_param['activation_function_three'],\n",
+    "            dropout_p               = best_param['dropout_p'],\n",
+    "            learning_rate           = best_param['learning_rate'],\n",
+    "            l2_reg                  = best_param['l2_reg'],\n",
+    "            optimizer_choice        = best_param['optimizer']\n",
+    "        )\n",
+    "        model.fit(\n",
+    "            [X_gene_train, X_dna_train],\n",
+    "            y_train,\n",
+    "            epochs=best_param['epoch'],\n",
+    "            batch_size=best_param['batch_size'],\n",
+    "            shuffle=False,\n",
+    "            verbose=0\n",
+    "        )\n",
+    "        y_pred = model.predict([X_gene_test, X_dna_test]).argmax(axis=1)\n",
+    "\n",
+    "        f1_val      = f1_score(y_test, y_pred, average=avg_type)\n",
+    "        precision_v = precision_score(y_test, y_pred, average=avg_type)\n",
+    "        recall_v    = recall_score(y_test, y_pred, average=avg_type)\n",
+    "        acc_v       = accuracy_score(y_test, y_pred)\n",
+    "\n",
+    "        per_run_results.append({\n",
+    "            'run_idx': run_i,\n",
+    "            f'f1_{avg_type}':       f1_val,\n",
+    "            f'precision_{avg_type}': precision_v,\n",
+    "            f'recall_{avg_type}':    recall_v,\n",
+    "            'accuracy':              acc_v\n",
+    "        })\n",
+    "\n",
+    "    df_runs = pd.DataFrame(per_run_results)\n",
+    "    summary = {\n",
+    "        f'f1_{avg_type}_mean':        df_runs[f'f1_{avg_type}'].mean(),\n",
+    "        f'f1_{avg_type}_std':         df_runs[f'f1_{avg_type}'].std(),\n",
+    "        f'precision_{avg_type}_mean': df_runs[f'precision_{avg_type}'].mean(),\n",
+    "        f'precision_{avg_type}_std':  df_runs[f'precision_{avg_type}'].std(),\n",
+    "        f'recall_{avg_type}_mean':    df_runs[f'recall_{avg_type}'].mean(),\n",
+    "        f'recall_{avg_type}_std':     df_runs[f'recall_{avg_type}'].std(),\n",
+    "        'accuracy_mean':              df_runs['accuracy'].mean(),\n",
+    "        'accuracy_std':               df_runs['accuracy'].std()\n",
+    "    }\n",
+    "\n",
+    "    df_runs[f'f1_{avg_type}_std']        = pd.NA\n",
+    "    df_runs[f'precision_{avg_type}_std'] = pd.NA\n",
+    "    df_runs[f'recall_{avg_type}_std']    = pd.NA\n",
+    "    df_runs['accuracy_std']              = pd.NA\n",
+    "\n",
+    "    summary_row = {\n",
+    "        'run_idx': 'mean',\n",
+    "        f'f1_{avg_type}':        summary[f'f1_{avg_type}_mean'],\n",
+    "        f'f1_{avg_type}_std':    summary[f'f1_{avg_type}_std'],\n",
+    "        f'precision_{avg_type}': summary[f'precision_{avg_type}_mean'],\n",
+    "        f'precision_{avg_type}_std': summary[f'precision_{avg_type}_std'],\n",
+    "        f'recall_{avg_type}':    summary[f'recall_{avg_type}_mean'],\n",
+    "        f'recall_{avg_type}_std': summary[f'recall_{avg_type}_std'],\n",
+    "        'accuracy':              summary['accuracy_mean'],\n",
+    "        'accuracy_std':          summary['accuracy_std']\n",
+    "    }\n",
+    "    df_summary = pd.DataFrame([summary_row])\n",
+    "\n",
+    "    df_runs = pd.concat([df_runs, df_summary], ignore_index=True)\n",
+    "\n",
+    "    if save_csv_path is not None:\n",
+    "        df_runs.to_csv(save_csv_path, index=False)\n",
+    "\n",
+    "    return summary, df_runs\n",
+    "\n",
+    "summary_macro, df_macro_runs = evaluate_test_model(\n",
+    "    best_param_macro,\n",
+    "    avg_type=\"macro\")"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "colo2",
+   "language": "python",
+   "name": "colo2"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

diagnostics/2026-10-16-brca-training-collapse/upstream/BioLMNET_Weighted_Loss.ipynb

@@ -0,0 +1,940 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "634f46b7-f788-499d-b014-7c28905eeff6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "os.environ['PYTHONHASHSEED'] = '0'\n",
+    "os.environ['TF_DETERMINISTIC_OPS'] = '1'\n",
+    "os.environ['CUDA_VISIBLE_DEVICES'] = ''\n",
+    "import random\n",
+    "import numpy as np\n",
+    "import pandas as pd\n",
+    "import re\n",
+    "import glob\n",
+    "from collections import defaultdict\n",
+    "from sklearn.preprocessing import StandardScaler, LabelEncoder\n",
+    "from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\n",
+    "from sklearn.model_selection import train_test_split, StratifiedKFold\n",
+    "import matplotlib.pyplot as plt\n",
+    "import tensorflow as tf\n",
+    "from tensorflow.keras import layers, Model\n",
+    "from tensorflow.keras.layers import Dense, Dropout, Input, concatenate\n",
+    "import itertools\n",
+    "from joblib import Parallel, delayed\n",
+    "from tqdm import tqdm\n",
+    "from tqdm_joblib import tqdm_joblib\n",
+    "from sklearn.utils.class_weight import compute_class_weight\n",
+    "tf.config.threading.set_intra_op_parallelism_threads(1)\n",
+    "tf.config.threading.set_inter_op_parallelism_threads(1)\n",
+    "tf.random.set_seed(42)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "535e884e-58a5-434e-95df-0735956321b5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from huggingface_hub import hf_hub_download\n",
+    "import pandas as pd\n",
+    "\n",
+    "FNAME = \"embedding_associations_cell_type_tissue_drug_pathway_openai_large.parquet\"  \n",
+    "\n",
+    "parquet_path = hf_hub_download(\n",
+    "    repo_id   = \"honicky/genept-composable-embeddings\",\n",
+    "    filename  = FNAME,\n",
+    "    repo_type = \"model\"\n",
+    ")\n",
+    "emb = pd.read_parquet(parquet_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1a4d2523-d9a0-466f-bc58-8c815370e9e2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "df_gene = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"Gene_Expression.csv\")\n",
+    "df_dna = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"DNA_Methylation.csv\")\n",
+    " \n",
+    "true_label = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"label.csv\")\n",
+    "\n",
+    "scaler = StandardScaler()\n",
+    "le = LabelEncoder()\n",
+    "le.fit(true_label)\n",
+    "y = le.transform(true_label)\n",
+    "\n",
+    "ge_genes = list(df_gene.columns)\n",
+    "\n",
+    "df_pdi = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"PDI\"/ \"PDI.csv\",sep=',')\n",
+    "df_ge_pdi = df_pdi[df_pdi.Target.isin(ge_genes)]\n",
+    "unique_tfs = df_ge_pdi['TF'].unique()\n",
+    "unique_targets = df_ge_pdi['Target'].unique()\n",
+    "\n",
+    "dna_genes =list(df_dna.columns) \n",
+    "df_dna_pdi = df_pdi[df_pdi.Target.isin(dna_genes)]\n",
+    "\n",
+    "dna_unique_tfs = df_dna_pdi['TF'].unique()\n",
+    "dna_unique_targets = df_dna_pdi['Target'].unique()\n",
+    "\n",
+    "df_ppi = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"PPI\"/ \"PPI.csv\",sep=',')\n",
+    "df_ppi = df_ppi.drop(['Unnamed: 0'], axis =1)\n",
+    "df_ppi['combined_score'] = df_ppi['combined_score'] /1000\n",
+    "threshold = df_ppi['combined_score'].quantile(0.9)\n",
+    "df_ppi = df_ppi[df_ppi['combined_score'] >= threshold]\n",
+    "df_ge_ppi = df_ppi[df_ppi['protein1'].isin(ge_genes) | df_ppi['protein2'].isin(ge_genes)]\n",
+    "unique_protein1 = df_ge_ppi['protein1'].unique()\n",
+    "unique_protein2 = df_ge_ppi['protein2'].unique()\n",
+    "\n",
+    "df_dna_ppi = df_ppi[df_ppi['protein1'].isin(dna_genes) | df_ppi['protein2'].isin(dna_genes)]\n",
+    "dna_unique_protein1 = df_dna_ppi['protein1'].unique()\n",
+    "dna_unique_protein2 = df_dna_ppi['protein2'].unique()\n",
+    "\n",
+    "ge_pdi_ppi_genes = list(set(unique_targets) | set(unique_protein2))\n",
+    "dna_pdi_ppi_genes = list(set(dna_unique_targets) | set(dna_unique_protein2))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e868db94-f1d2-413a-b36f-028cb0d7499d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "genept_gene_emb = emb.loc[emb.index.isin(ge_pdi_ppi_genes)]\n",
+    "ge_pdi_ppi_genept = genept_gene_emb.index"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dfd04c7c-6119-4f60-836e-43d2c2060299",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "genept_dna_emb = emb.loc[emb.index.isin(dna_pdi_ppi_genes)]\n",
+    "dna_pdi_ppi_genept = genept_dna_emb.index"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "537876a2-01e5-4a8f-92df-621469614756",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_ge_mask_matrix (extra_neurons):\n",
+    "    de_gene_to_idx = {gene: i for i, gene in enumerate(ge_genes)}\n",
+    "    target_to_idx = {target: j for j, target in enumerate(unique_targets)}\n",
+    "    protein2_to_idx = {protein2: k for k, protein2 in enumerate(unique_protein2)}\n",
+    "\n",
+    "    pdi_mask_matrix = np.zeros((len(ge_genes), len(unique_targets)), dtype=int)\n",
+    "\n",
+    "    for _, row in df_ge_pdi.iterrows():\n",
+    "        tfa = row[\"TF\"]\n",
+    "        target = row[\"Target\"]\n",
+    "        if tfa in de_gene_to_idx and target in target_to_idx:\n",
+    "            tf_idx = de_gene_to_idx[tfa]\n",
+    "            target_idx = target_to_idx[target]\n",
+    "            pdi_mask_matrix[tf_idx, target_idx] = 1\n",
+    "    \n",
+    "    ppi_mask_matrix = np.zeros((len(ge_genes), len(unique_protein2)))\n",
+    "\n",
+    "    for _, row in df_ge_ppi.iterrows():\n",
+    "        protein1 = row[\"protein1\"]\n",
+    "        protein2 = row[\"protein2\"]\n",
+    "        combined_score = row[\"combined_score\"]\n",
+    "\n",
+    "        if protein1 in de_gene_to_idx and protein2 in protein2_to_idx:\n",
+    "            protein1_idx = de_gene_to_idx[protein1]\n",
+    "            protein2_idx = protein2_to_idx[protein2]\n",
+    "            ppi_mask_matrix[protein1_idx, protein2_idx] = combined_score \n",
+    "            \n",
+    "    pdi_mask_df=     pd.DataFrame(pdi_mask_matrix)\n",
+    "    ppi_mask_df=     pd.DataFrame(ppi_mask_matrix)\n",
+    "    pdi_mask_df.columns = unique_targets\n",
+    "    ppi_mask_df.columns = unique_protein2\n",
+    "    pdi_ppi_mask_df = pd.concat([pdi_mask_df, ppi_mask_df], axis=1)\n",
+    "    pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, ~pdi_ppi_mask_df.columns.duplicated(keep='last')]\n",
+    "    col_mask = pdi_ppi_mask_df.columns.isin(ge_pdi_ppi_genept)\n",
+    "    sub_pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, col_mask]\n",
+    "    \n",
+    "    extra_columns = np.ones((len(ge_genes), extra_neurons), dtype=int)\n",
+    "    adjusted_mask_matrix = np.hstack([sub_pdi_ppi_mask_df, extra_columns])\n",
+    "    new_mask_matrix = tf.constant(adjusted_mask_matrix, dtype=tf.float32)\n",
+    "    \n",
+    "    return new_mask_matrix\n",
+    "\n",
+    "def create_dna_mask_matrix (extra_neurons):\n",
+    "    de_gene_to_idx = {gene: i for i, gene in enumerate(dna_genes)}\n",
+    "    target_to_idx = {target: j for j, target in enumerate(dna_unique_targets)}\n",
+    "    protein2_to_idx = {protein2: k for k, protein2 in enumerate(dna_unique_protein2)}\n",
+    "\n",
+    "\n",
+    "    pdi_mask_matrix = np.zeros((len(dna_genes), len(dna_unique_targets)), dtype=int)\n",
+    "\n",
+    "\n",
+    "    for _, row in df_dna_pdi.iterrows():\n",
+    "        tfa = row[\"TF\"]\n",
+    "        target = row[\"Target\"]\n",
+    "        if tfa in de_gene_to_idx and target in target_to_idx:\n",
+    "            tf_idx = de_gene_to_idx[tfa]\n",
+    "            target_idx = target_to_idx[target]\n",
+    "            pdi_mask_matrix[tf_idx, target_idx] = 1\n",
+    "    \n",
+    "    ppi_mask_matrix = np.zeros((len(dna_genes), len(dna_unique_protein2)))\n",
+    "\n",
+    "    for _, row in df_dna_ppi.iterrows():\n",
+    "        protein1 = row[\"protein1\"]\n",
+    "        protein2 = row[\"protein2\"]\n",
+    "        combined_score = row[\"combined_score\"]\n",
+    "\n",
+    "        if protein1 in de_gene_to_idx and protein2 in protein2_to_idx:\n",
+    "            protein1_idx = de_gene_to_idx[protein1]\n",
+    "            protein2_idx = protein2_to_idx[protein2]\n",
+    "            ppi_mask_matrix[protein1_idx, protein2_idx] = combined_score \n",
+    "            \n",
+    "    pdi_mask_df=     pd.DataFrame(pdi_mask_matrix)\n",
+    "    ppi_mask_df=     pd.DataFrame(ppi_mask_matrix)\n",
+    "    pdi_mask_df.columns = dna_unique_targets\n",
+    "    ppi_mask_df.columns = dna_unique_protein2\n",
+    "    pdi_ppi_mask_df = pd.concat([pdi_mask_df, ppi_mask_df], axis=1)\n",
+    "    pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, ~pdi_ppi_mask_df.columns.duplicated(keep='last')]\n",
+    "    col_mask = pdi_ppi_mask_df.columns.isin(dna_pdi_ppi_genept)\n",
+    "    sub_pdi_ppi_mask_df = pdi_ppi_mask_df.loc[:, col_mask]\n",
+    "\n",
+    "    \n",
+    "    extra_columns = np.ones((len(dna_genes), extra_neurons), dtype=int)\n",
+    "    adjusted_mask_matrix = np.hstack([sub_pdi_ppi_mask_df, extra_columns])\n",
+    "    new_mask_matrix = tf.constant(adjusted_mask_matrix, dtype=tf.float32)\n",
+    "    \n",
+    "    return new_mask_matrix"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e5161905-ef7c-4943-aa4d-46a2dc84f498",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ge_target_to_kegg = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"ge_target_to_KEGG_significant.csv\",sep=',')\n",
+    "dna_target_to_kegg = pd.read_csv(Path(__file__).parent / \"Dataset\"/\"scTrioseq2\"/ \"dna_target_to_KEGG_significant.csv\",sep=',')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "235fc5c0-0c09-47c7-aa1a-36d708372be1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_ge_pathway_mask_matrix ():\n",
+    "    genes = ge_pdi_ppi_genept\n",
+    "    pathways = ge_target_to_kegg[\"PathwayID\"].unique().tolist()\n",
+    "\n",
+    "    gene_to_idx = {gene: i for i, gene in enumerate(genes)}\n",
+    "    pathway_to_idx = {pw: j for j, pw in enumerate(pathways)}\n",
+    "\n",
+    "    mask_matrix = np.zeros((len(genes), len(pathways)), dtype=int)\n",
+    "\n",
+    "    for _, row in ge_target_to_kegg.iterrows():\n",
+    "        gene = row[\"SYMBOL\"]\n",
+    "        pw = row[\"PathwayID\"]\n",
+    "        if gene in gene_to_idx and pw in pathway_to_idx:\n",
+    "            i = gene_to_idx[gene]\n",
+    "            j = pathway_to_idx[pw]\n",
+    "            mask_matrix[i, j] = 1\n",
+    "    mask_df = pd.DataFrame(mask_matrix, index=genes, columns=pathways)\n",
+    "    return mask_df\n",
+    "\n",
+    "def create_dna_pathway_mask_matrix ():\n",
+    "    genes = dna_pdi_ppi_genept\n",
+    "    pathways = dna_target_to_kegg[\"PathwayID\"].unique().tolist()\n",
+    "\n",
+    "    gene_to_idx = {gene: i for i, gene in enumerate(genes)}\n",
+    "    pathway_to_idx = {pw: j for j, pw in enumerate(pathways)}\n",
+    "\n",
+    "    mask_matrix = np.zeros((len(genes), len(pathways)), dtype=int)\n",
+    "\n",
+    "    for _, row in dna_target_to_kegg.iterrows():\n",
+    "        gene = row[\"SYMBOL\"]\n",
+    "        pw = row[\"PathwayID\"]\n",
+    "        if gene in gene_to_idx and pw in pathway_to_idx:\n",
+    "            i = gene_to_idx[gene]\n",
+    "            j = pathway_to_idx[pw]\n",
+    "            mask_matrix[i, j] = 1\n",
+    "    mask_df = pd.DataFrame(mask_matrix, index=genes, columns=pathways)\n",
+    "    return mask_df\n",
+    "    \n",
+    "mask_gene_pathway  = create_ge_pathway_mask_matrix ()\n",
+    "mask_dna_pathway = create_dna_pathway_mask_matrix()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dbfe572a-d773-4306-94a1-69ab748f924c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class MaskedLayer(layers.Layer):\n",
+    "    def __init__(self, mask_matrix, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "        self.mask_matrix = K.constant(mask_matrix, dtype='float32')\n",
+    "\n",
+    "    def call(self, inputs):\n",
+    "        return K.dot(inputs, self.mask_matrix)\n",
+    "\n",
+    "\n",
+    "class AttentionPathwayLayer(layers.Layer):\n",
+    "    def __init__(self, genept_embs, pathway_mask, init_seed=42, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "\n",
+    "        genept_np = genept_embs.values if hasattr(genept_embs, \"values\") else np.asarray(genept_embs)\n",
+    "        n_genes, D_embed = genept_np.shape\n",
+    "\n",
+    "        pw_np = pathway_mask.values if hasattr(pathway_mask, \"values\") else np.asarray(pathway_mask)\n",
+    "        n_pathways = pw_np.shape[1]\n",
+    "\n",
+    "        self.G = K.constant(genept_np, dtype=\"float32\")\n",
+    "        self.P = K.constant(pw_np, dtype=\"float32\")\n",
+    "\n",
+    "        glorot = initializers.GlorotUniform(seed=init_seed)\n",
+    "        self.W = self.add_weight(\n",
+    "            shape=(n_pathways, D_embed),\n",
+    "            initializer=glorot,\n",
+    "            trainable=True,\n",
+    "            name=\"W_query\"\n",
+    "        )\n",
+    "\n",
+    "    def call(self, x):\n",
+    "        G_T      = K.transpose(self.G)            \n",
+    "        S        = K.dot(self.W, G_T)             \n",
+    "        S_T      = K.transpose(S)                 \n",
+    "        S_masked = S_T * self.P                   \n",
+    "        alpha    = K.softmax(S_masked, axis=0)    \n",
+    "        x_sq     = K.squeeze(x, axis=-1)          \n",
+    "        out_mat  = K.dot(x_sq, alpha)             \n",
+    "        return K.expand_dims(out_mat, axis=-1)    \n",
+    "\n",
+    "    def compute_output_shape(self, input_shape):\n",
+    "        batch_size, n_genes, _ = input_shape\n",
+    "        n_pathways = K.int_shape(self.P)[1]\n",
+    "        return (batch_size, n_pathways, 1)\n",
+    "    \n",
+    "    def get_attention(self):\n",
+    "        G_T      = K.transpose(self.G)            \n",
+    "        S        = K.dot(self.W, G_T)             \n",
+    "        S_T      = K.transpose(S)                 \n",
+    "        S_masked = S_T * self.P                   \n",
+    "        return K.softmax(S_masked, axis=0)      \n",
+    "\n",
+    "\n",
+    "def create_model(\n",
+    "    num_ex_neuron,\n",
+    "    layer_one,\n",
+    "    activation_function_one,\n",
+    "    activation_function_two,\n",
+    "    activation_function_three,\n",
+    "    dropout_p,\n",
+    "    learning_rate,\n",
+    "    l2_reg,\n",
+    "    optimizer_choice,\n",
+    "    genept_gene_emb,\n",
+    "    mask_gene_pathway,\n",
+    "    genept_dna_emb,\n",
+    "    mask_dna_pathway\n",
+    "):\n",
+    "\n",
+    "    new_ge_mask_matrix = create_ge_mask_matrix(num_ex_neuron)\n",
+    "    new_dna_mask_matrix = create_dna_mask_matrix(num_ex_neuron)\n",
+    "    \n",
+    "    inputA = Input(shape=(df_gene.shape[1],), name=\"gene_input\")\n",
+    "    x = layers.Dense(\n",
+    "        new_ge_mask_matrix.shape[0],\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42)\n",
+    "    )(inputA)\n",
+    "    x = MaskedLayer(new_ge_mask_matrix)(x)\n",
+    "    x = Dropout(dropout_p)(x)\n",
+    "    x_feat = Lambda(lambda t: K.expand_dims(t, -1), name=\"gene_add_dim\")(x)\n",
+    "\n",
+    "    gene_path_layer = AttentionPathwayLayer(\n",
+    "    genept_embs  = genept_gene_emb,\n",
+    "    pathway_mask = mask_gene_pathway,\n",
+    "    name         = \"gene_path_attn\"\n",
+    "    )\n",
+    "    x_path = gene_path_layer(x_feat)\n",
+    "    x_path = Flatten()(x_path)\n",
+    "\n",
+    "    x_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42),\n",
+    "        name=\"gene_proj\"\n",
+    "    )(x_path)\n",
+    "\n",
+    "    inputB = Input(shape=(df_dna.shape[1],), name=\"dna_input\")\n",
+    "    y = layers.Dense(\n",
+    "        new_dna_mask_matrix.shape[0],\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42)\n",
+    "    )(inputB)\n",
+    "    y = MaskedLayer(new_dna_mask_matrix)(y)\n",
+    "    y = Dropout(dropout_p)(y)\n",
+    "    y_feat = Lambda(lambda t: K.expand_dims(t, -1), name=\"dna_add_dim\")(y)\n",
+    "\n",
+    "    dna_path_layer = AttentionPathwayLayer(\n",
+    "    genept_embs  = genept_dna_emb,\n",
+    "    pathway_mask = mask_dna_pathway,\n",
+    "    name         = \"dna_path_attn\"\n",
+    "    )\n",
+    "    y_path = dna_path_layer(y_feat)\n",
+    "    y_path = Flatten()(y_path)\n",
+    "\n",
+    "    y_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.LecunUniform(42),\n",
+    "        name=\"dna_proj\"\n",
+    "    )(y_path)\n",
+    "\n",
+    "    merged = concatenate([x_proj, y_proj], name=\"merge_xy\")\n",
+    "    h = layers.Dense(\n",
+    "        12,\n",
+    "        activation=activation_function_three,\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.GlorotUniform(42),\n",
+    "        name=\"fusion_dense\"\n",
+    "    )(merged)\n",
+    "    h = Dropout(dropout_p, name=\"drop_fusion\")(h)\n",
+    "    out = layers.Dense(\n",
+    "        4,\n",
+    "        activation=\"softmax\",\n",
+    "        kernel_regularizer=regularizers.l2(l2_reg),\n",
+    "        kernel_initializer=initializers.GlorotUniform(42),\n",
+    "        name=\"output\"\n",
+    "    )(h)\n",
+    "\n",
+    "    model = Model(inputs=[inputA, inputB], outputs=out, name=\"PathwayModel\")\n",
+    "\n",
+    "    if optimizer_choice == 'adam':\n",
+    "        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
+    "    elif optimizer_choice == 'sgd':\n",
+    "        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate)\n",
+    "    model.compile(\n",
+    "        optimizer=optimizer,\n",
+    "        loss='sparse_categorical_crossentropy',\n",
+    "        metrics=['accuracy']\n",
+    "    )\n",
+    "    return model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "be89b57c-5f26-4b58-be38-7d4c797d4c6d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def evaluate_model_cv(X_train, X_val, y_train, y_val, params):\n",
+    "    model = create_model(\n",
+    "        params['num_ex_neuron'], params['layer_one'],\n",
+    "        params['activation_function_one'], params['activation_function_two'], params['activation_function_three'],\n",
+    "        params['dropout_p'], params['learning_rate'], params['l2_reg'], params['optimizer'],\n",
+    "        genept_gene_emb, mask_gene_pathway, genept_dna_emb, mask_dna_pathway\n",
+    "    )\n",
+    "\n",
+    "    classes = np.unique(y_train)\n",
+    "    weights = compute_class_weight('balanced', classes=classes, y=y_train)\n",
+    "    class_weight = dict(zip(classes, weights))\n",
+    "\n",
+    "    history = model.fit(\n",
+    "        X_train, y_train,\n",
+    "        validation_data=(X_val, y_val),\n",
+    "        epochs=params['epoch'], batch_size=params['batch_size'],\n",
+    "        shuffle=False, verbose=0,\n",
+    "        class_weight=class_weight\n",
+    "    )\n",
+    "\n",
+    "    y_pred = model.predict(X_val).argmax(1)\n",
+    "    return {\n",
+    "        'history': history.history,\n",
+    "        'f1_weighted': f1_score(y_val, y_pred, average='weighted'),\n",
+    "        'f1_macro':    f1_score(y_val, y_pred, average='macro'),\n",
+    "        'f1_micro':    f1_score(y_val, y_pred, average='micro'),\n",
+    "        'precision_weighted': precision_score(y_val, y_pred, average='weighted'),\n",
+    "        'precision_macro':    precision_score(y_val, y_pred, average='macro'),\n",
+    "        'precision_micro':    precision_score(y_val, y_pred, average='micro'),\n",
+    "        'recall_weighted':    recall_score(y_val, y_pred, average='weighted'),\n",
+    "        'recall_macro':       recall_score(y_val, y_pred, average='macro'),\n",
+    "        'recall_micro':       recall_score(y_val, y_pred, average='micro'),\n",
+    "        'accuracy': accuracy_score(y_val, y_pred)\n",
+    "    }\n",
+    "\n",
+    "def evaluate_combination(param_combination,X_gene_train,X_dna_train,y_train,n_splits=5):\n",
+    "    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=42)\n",
+    "    \n",
+    "    f1_weighted_list = []\n",
+    "    f1_macro_list = []\n",
+    "    f1_micro_list = []\n",
+    "    \n",
+    "    precision_weighted_list = []\n",
+    "    precision_macro_list = []\n",
+    "    precision_micro_list = []\n",
+    "    \n",
+    "    recall_weighted_list = []\n",
+    "    recall_macro_list = []\n",
+    "    recall_micro_list = []\n",
+    "    \n",
+    "    accuracy_list = []\n",
+    "    all_histories = []\n",
+    "    \n",
+    "    for train_idx, val_idx in skf.split(X_gene_train, y_train):\n",
+    "        Xg_tr, Xg_val = X_gene_train[train_idx], X_gene_train[val_idx]\n",
+    "        Xd_tr, Xd_val = X_dna_train[train_idx], X_dna_train[val_idx]\n",
+    "        y_tr, y_val   = y_train[train_idx], y_train[val_idx]\n",
+    "    \n",
+    "        scores = evaluate_model_cv([Xg_tr, Xd_tr], [Xg_val, Xd_val], y_tr, y_val, param_combination)\n",
+    "        \n",
+    "        history_dict = scores.pop('history')\n",
+    "        all_histories.append(history_dict)\n",
+    "        \n",
+    "        # Append each metric.\n",
+    "        f1_weighted_list.append(scores['f1_weighted'])\n",
+    "        f1_macro_list.append(scores['f1_macro'])\n",
+    "        f1_micro_list.append(scores['f1_micro'])\n",
+    "        \n",
+    "        precision_weighted_list.append(scores['precision_weighted'])\n",
+    "        precision_macro_list.append(scores['precision_macro'])\n",
+    "        precision_micro_list.append(scores['precision_micro'])\n",
+    "        \n",
+    "        recall_weighted_list.append(scores['recall_weighted'])\n",
+    "        recall_macro_list.append(scores['recall_macro'])\n",
+    "        recall_micro_list.append(scores['recall_micro'])\n",
+    "        \n",
+    "        accuracy_list.append(scores['accuracy'])\n",
+    "    \n",
+    "    results = {\n",
+    "        'f1_weighted_mean': np.mean(f1_weighted_list),\n",
+    "        'f1_weighted_std': np.std(f1_weighted_list),\n",
+    "        'f1_macro_mean': np.mean(f1_macro_list),\n",
+    "        'f1_macro_std': np.std(f1_macro_list),\n",
+    "        'f1_micro_mean': np.mean(f1_micro_list),\n",
+    "        'f1_micro_std': np.std(f1_micro_list),\n",
+    "        'precision_weighted_mean': np.mean(precision_weighted_list),\n",
+    "        'precision_weighted_std': np.std(precision_weighted_list),\n",
+    "        'precision_macro_mean': np.mean(precision_macro_list),\n",
+    "        'precision_macro_std': np.std(precision_macro_list),\n",
+    "        'precision_micro_mean': np.mean(precision_micro_list),\n",
+    "        'precision_micro_std': np.std(precision_micro_list),\n",
+    "        'recall_weighted_mean': np.mean(recall_weighted_list),\n",
+    "        'recall_weighted_std': np.std(recall_weighted_list),\n",
+    "        'recall_macro_mean': np.mean(recall_macro_list),\n",
+    "        'recall_macro_std': np.std(recall_macro_list),\n",
+    "        'recall_micro_mean': np.mean(recall_micro_list),\n",
+    "        'recall_micro_std': np.std(recall_micro_list),\n",
+    "        'accuracy_mean': np.mean(accuracy_list),\n",
+    "        'accuracy_std': np.std(accuracy_list)\n",
+    "    }\n",
+    "    \n",
+    "    return param_combination, results, all_histories\n",
+    "\n",
+    "params_grid = {\n",
+    "     'num_ex_neuron': [0],\n",
+    "     'layer_one': [64],\n",
+    "     'activation_function_one': ['relu'],\n",
+    "     'activation_function_two': ['sigmoid'],\n",
+    "     'activation_function_three': ['tanh'],\n",
+    "     'dropout_p': [0.3],\n",
+    "     'learning_rate': [0.01],\n",
+    "    'l2_reg': [0.01],\n",
+    "    'epoch': [200],\n",
+    "    'batch_size':[16],\n",
+    "    'optimizer': ['sgd']\n",
+    " }\n",
+    "\n",
+    "\n",
+    "param_combinations = list(itertools.product(\n",
+    "    params_grid['num_ex_neuron'],\n",
+    "    params_grid['layer_one'],\n",
+    "    params_grid['activation_function_one'],\n",
+    "    params_grid['activation_function_two'],\n",
+    "    params_grid['activation_function_three'],\n",
+    "    params_grid['dropout_p'],\n",
+    "    params_grid['learning_rate'],\n",
+    "    params_grid['l2_reg'],\n",
+    "    params_grid['epoch'],\n",
+    "    params_grid['batch_size'],\n",
+    "    params_grid['optimizer']\n",
+    "    \n",
+    "))\n",
+    "\n",
+    "\n",
+    "param_dicts = []\n",
+    "for comb in param_combinations:\n",
+    "    param_dicts.append({\n",
+    "        'num_ex_neuron': comb[0],\n",
+    "        'layer_one': comb[1],\n",
+    "        'activation_function_one': comb[2],\n",
+    "        'activation_function_two': comb[3],\n",
+    "        'activation_function_three': comb[4],\n",
+    "        'dropout_p': comb[5],\n",
+    "        'learning_rate': comb[6],\n",
+    "        'l2_reg': comb[7],\n",
+    "        'epoch': comb[8],\n",
+    "        'batch_size': comb[9],\n",
+    "        'optimizer': comb[10]\n",
+    "    })\n",
+    "\n",
+    "\n",
+    "X_gene_train, X_gene_test, X_dna_train, X_dna_test, y_train, y_test = train_test_split(df_gene, df_dna, y, test_size=0.2, stratify=y, random_state=42)\n",
+    "scaler = StandardScaler()\n",
+    "X_gene_train=(np.array(X_gene_train).astype(np.float32))\n",
+    "scaler.fit(X_gene_train)\n",
+    "X_gene_train=scaler.transform(X_gene_train)\n",
+    "\n",
+    "X_gene_test=(np.array(X_gene_test).astype(np.float32))\n",
+    "scaler.fit(X_gene_test)\n",
+    "X_gene_test=scaler.transform(X_gene_test)\n",
+    "\n",
+    "X_dna_train=(np.array(X_dna_train).astype(np.float32))\n",
+    "scaler.fit(X_dna_train)\n",
+    "X_dna_train=scaler.transform(X_dna_train)\n",
+    "\n",
+    "X_dna_test=(np.array(X_dna_test).astype(np.float32))\n",
+    "scaler.fit(X_dna_test)\n",
+    "X_dna_test=scaler.transform(X_dna_test)\n",
+    "\n",
+    "\n",
+    "with tqdm_joblib(\n",
+    "        tqdm(desc=\"Evaluating hyperparameter combinations\",\n",
+    "             total=len(param_dicts))\n",
+    "     ) as progress_bar:\n",
+    "    results = Parallel(n_jobs=-1)(\n",
+    "        delayed(evaluate_combination)(\n",
+    "            param, X_gene_train, X_dna_train, y_train, n_splits=5\n",
+    "        )\n",
+    "        for param in param_dicts\n",
+    "    )\n",
+    "\n",
+    "results_list = []\n",
+    "histories_dict = {}\n",
+    "best_param_weighted = None\n",
+    "best_weighted_scores = None\n",
+    "best_avg_f1_weighted = -np.inf\n",
+    "\n",
+    "best_param_macro = None\n",
+    "best_macro_scores = None\n",
+    "best_avg_f1_macro = -np.inf\n",
+    "\n",
+    "best_param_micro = None\n",
+    "best_micro_scores = None\n",
+    "best_avg_f1_micro = -np.inf\n",
+    "\n",
+    "for param_dict, scores_dict, histories in results:\n",
+    "    key = str(param_dict)\n",
+    "    histories_dict[key] = histories\n",
+    "\n",
+    "    row = {}\n",
+    "    row.update(param_dict)\n",
+    "    row.update(scores_dict)\n",
+    "    results_list.append(row)\n",
+    "\n",
+    "    if scores_dict['f1_weighted_mean'] > best_avg_f1_weighted:\n",
+    "        best_avg_f1_weighted = scores_dict['f1_weighted_mean']\n",
+    "        best_param_weighted  = param_dict\n",
+    "        best_weighted_scores = scores_dict\n",
+    "\n",
+    "    if scores_dict['f1_macro_mean'] > best_avg_f1_macro:\n",
+    "        best_avg_f1_macro = scores_dict['f1_macro_mean']\n",
+    "        best_param_macro  = param_dict\n",
+    "        best_macro_scores = scores_dict\n",
+    "\n",
+    "    if scores_dict['f1_micro_mean'] > best_avg_f1_micro:\n",
+    "        best_avg_f1_micro = scores_dict['f1_micro_mean']\n",
+    "        best_param_micro  = param_dict\n",
+    "        best_micro_scores = scores_dict\n",
+    "\n",
+    "results_df = pd.DataFrame(results_list)\n",
+    "best_key = str(best_param_macro)\n",
+    "best_histories = histories_dict[best_key]  \n",
+    "\n",
+    "\n",
+    "hyperparam_cols = list(param_dict.keys())\n",
+    "\n",
+    "\n",
+    "weighted_cols = hyperparam_cols + ['f1_weighted_mean', 'f1_weighted_std',\n",
+    "                                     'precision_weighted_mean', 'precision_weighted_std',\n",
+    "                                     'recall_weighted_mean', 'recall_weighted_std',\n",
+    "                                     'accuracy_mean', 'accuracy_std']\n",
+    "results_df_weighted = results_df[weighted_cols]\n",
+    "\n",
+    "macro_cols = hyperparam_cols + ['f1_macro_mean', 'f1_macro_std',\n",
+    "                                  'precision_macro_mean', 'precision_macro_std',\n",
+    "                                  'recall_macro_mean', 'recall_macro_std',\n",
+    "                                  'accuracy_mean', 'accuracy_std']\n",
+    "results_df_macro = results_df[macro_cols]\n",
+    "\n",
+    "micro_cols = hyperparam_cols + ['f1_micro_mean', 'f1_micro_std',\n",
+    "                                  'precision_micro_mean', 'precision_micro_std',\n",
+    "                                  'recall_micro_mean', 'recall_micro_std',\n",
+    "                                  'accuracy_mean', 'accuracy_std']\n",
+    "results_df_micro = results_df[micro_cols]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2576c647-0fa9-44ff-b26b-b152a8f96389",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def create_model_test(\n",
+    "    num_ex_neuron,\n",
+    "    layer_one,\n",
+    "    activation_function_one,\n",
+    "    activation_function_two,\n",
+    "    activation_function_three,\n",
+    "    dropout_p,\n",
+    "    learning_rate,\n",
+    "    l2_reg,\n",
+    "    optimizer_choice\n",
+    "):\n",
+    "\n",
+    "    new_ge_mask_matrix = create_ge_mask_matrix(num_ex_neuron)    \n",
+    "    new_dna_mask_matrix = create_dna_mask_matrix(num_ex_neuron)  \n",
+    "\n",
+    "    inputA = Input(shape=(df_gene.shape[1],), name=\"gene_input\") \n",
+    "    inputB = Input(shape=(df_dna.shape[1],), name=\"dna_input\")   \n",
+    "\n",
+    "    x1 = layers.Dense(\n",
+    "        new_ge_mask_matrix.shape[0],               \n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=l2(l2_reg)\n",
+    "    )(inputA)                                        \n",
+    "\n",
+    "\n",
+    "    x1_masked = MaskedLayer(new_ge_mask_matrix)(x1) \n",
+    "    x1_masked = Dropout(dropout_p)(x1_masked)       \n",
+    "\n",
+    "    x1_feat = layers.Lambda(lambda t: K.expand_dims(t, -1), name=\"gene_add_dim\")(x1_masked)\n",
+    "\n",
+    "    gene_path_layer = AttentionPathwayLayer(\n",
+    "        genept_embs  = genept_gene_emb,    \n",
+    "        pathway_mask = mask_gene_pathway,  \n",
+    "        name         = \"gene_path_attn\"\n",
+    "    )\n",
+    "    x1_path = gene_path_layer(x1_feat)     \n",
+    "\n",
+    "    x1_path = Flatten(name=\"gene_flat_path\")(x1_path)  \n",
+    "\n",
+    "    x1_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"gene_proj\"\n",
+    "    )(x1_path)                                   \n",
+    "\n",
+    "    y1 = layers.Dense(\n",
+    "        new_dna_mask_matrix.shape[0],               \n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_one,\n",
+    "        kernel_regularizer=l2(l2_reg)\n",
+    "    )(inputB)                                     \n",
+    "\n",
+    "    y1_masked = MaskedLayer(new_dna_mask_matrix)(y1)  \n",
+    "    y1_masked = Dropout(dropout_p)(y1_masked)         \n",
+    "\n",
+    "    y1_feat = layers.Lambda(lambda t: K.expand_dims(t, -1), name=\"dna_add_dim\")(y1_masked)\n",
+    "\n",
+    "    dna_path_layer = AttentionPathwayLayer(\n",
+    "        genept_embs  = genept_dna_emb,      \n",
+    "        pathway_mask = mask_dna_pathway,    \n",
+    "        name         = \"dna_path_attn\"\n",
+    "    )\n",
+    "    y1_path = dna_path_layer(y1_feat) \n",
+    "    y1_path = Flatten(name=\"dna_flat_path\")(y1_path)   \n",
+    "\n",
+    "    y1_proj = layers.Dense(\n",
+    "        layer_one,\n",
+    "        kernel_initializer=LecunUniform(seed=42),\n",
+    "        activation=activation_function_two,\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"dna_proj\"\n",
+    "    )(y1_path)                                    \n",
+    "\n",
+    "    merged = concatenate([x1_proj, y1_proj], name=\"merge_xy\")  \n",
+    "\n",
+    "    h = layers.Dense(\n",
+    "        12,\n",
+    "        activation=activation_function_three,\n",
+    "        kernel_initializer=GlorotUniform(seed=42),\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"fusion_dense\"\n",
+    "    )(merged)                                        \n",
+    "    h = Dropout(dropout_p, name=\"drop_fusion\")(h)    \n",
+    "\n",
+    "    outputs = layers.Dense(\n",
+    "        4,\n",
+    "        activation=\"softmax\",\n",
+    "        kernel_initializer=GlorotUniform(seed=42),\n",
+    "        kernel_regularizer=l2(l2_reg),\n",
+    "        name=\"output\"\n",
+    "    )(h)                                             \n",
+    "\n",
+    "    model = Model(inputs=[inputA, inputB], outputs=outputs, name=\"PathwayModel_Test\")\n",
+    "\n",
+    "    if optimizer_choice == 'adam':\n",
+    "        optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
+    "    else:\n",
+    "        optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate)\n",
+    "\n",
+    "    model.compile(\n",
+    "        optimizer=optimizer,\n",
+    "        loss='sparse_categorical_crossentropy',\n",
+    "        metrics=['accuracy']\n",
+    "    )\n",
+    "    return model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fcda02d6-b3bd-4028-ba48-b0095d80245a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def _run_once(run_i,\n",
+    "              best_param,\n",
+    "              class_weight,\n",
+    "              avg_type,\n",
+    "              X_gene_train, X_dna_train, y_train,\n",
+    "              X_gene_test,  X_dna_test,  y_test,\n",
+    "              genept_gene_emb, mask_gene_pathway,\n",
+    "              genept_dna_emb,  mask_dna_pathway):\n",
+    "    model = create_model(\n",
+    "        best_param['num_ex_neuron'],\n",
+    "        best_param['layer_one'],\n",
+    "        best_param['activation_function_one'],\n",
+    "        best_param['activation_function_two'],\n",
+    "        best_param['activation_function_three'],\n",
+    "        best_param['dropout_p'],\n",
+    "        best_param['learning_rate'],\n",
+    "        best_param['l2_reg'],\n",
+    "        best_param['optimizer'],\n",
+    "        genept_gene_emb, mask_gene_pathway,\n",
+    "        genept_dna_emb,  mask_dna_pathway\n",
+    "    )\n",
+    "\n",
+    "    model.fit(\n",
+    "        [X_gene_train, X_dna_train], y_train,\n",
+    "        epochs=best_param['epoch'],\n",
+    "        batch_size=best_param['batch_size'],\n",
+    "        shuffle=False,\n",
+    "        verbose=0,\n",
+    "        class_weight=class_weight\n",
+    "    )\n",
+    "   \n",
+    "    y_pred = model.predict([X_gene_test, X_dna_test]).argmax(axis=1)\n",
+    "    return {\n",
+    "        'run_idx': run_i,\n",
+    "        f'f1_{avg_type}':       f1_score(y_test, y_pred, average=avg_type),\n",
+    "        f'precision_{avg_type}': precision_score(y_test, y_pred, average=avg_type),\n",
+    "        f'recall_{avg_type}':    recall_score(y_test, y_pred, average=avg_type),\n",
+    "        'accuracy':              accuracy_score(y_test, y_pred)\n",
+    "    }\n",
+    "\n",
+    "def evaluate_test_model(best_param, avg_type, n_runs=10, save_csv_path=None):\n",
+    "    \n",
+    "    classes = np.unique(y_train)\n",
+    "    weights = compute_class_weight('balanced', classes=classes, y=y_train)\n",
+    "    class_weight = dict(zip(classes, weights))\n",
+    "\n",
+    "    per_run_results = Parallel(\n",
+    "        n_jobs=n_runs,\n",
+    "        backend='multiprocessing'\n",
+    "    )(\n",
+    "        delayed(_run_once)(\n",
+    "            run_i,\n",
+    "            best_param,\n",
+    "            class_weight,\n",
+    "            avg_type,\n",
+    "            X_gene_train, X_dna_train, y_train,\n",
+    "            X_gene_test,  X_dna_test,  y_test,\n",
+    "            genept_gene_emb, mask_gene_pathway,\n",
+    "            genept_dna_emb,  mask_dna_pathway\n",
+    "        )\n",
+    "        for run_i in range(1, n_runs+1)\n",
+    "    )\n",
+    "\n",
+    "    df_runs = pd.DataFrame(per_run_results)\n",
+    "    \n",
+    "    summary = {\n",
+    "        f'f1_{avg_type}_mean':        df_runs[f'f1_{avg_type}'].mean(),\n",
+    "        f'f1_{avg_type}_std':         df_runs[f'f1_{avg_type}'].std(),\n",
+    "        f'precision_{avg_type}_mean': df_runs[f'precision_{avg_type}'].mean(),\n",
+    "        f'precision_{avg_type}_std':  df_runs[f'precision_{avg_type}'].std(),\n",
+    "        f'recall_{avg_type}_mean':    df_runs[f'recall_{avg_type}'].mean(),\n",
+    "        f'recall_{avg_type}_std':     df_runs[f'recall_{avg_type}'].std(),\n",
+    "        'accuracy_mean':              df_runs['accuracy'].mean(),\n",
+    "        'accuracy_std':               df_runs['accuracy'].std()\n",
+    "    }\n",
+    "\n",
+    "    if save_csv_path:\n",
+    "        df_runs.to_csv(save_csv_path, index=False)\n",
+    "\n",
+    "    return summary, df_runs\n",
+    "\n",
+    "\n",
+    "summary_macro, df_macro_runs = evaluate_test_model(\n",
+    "    best_param_macro,\n",
+    "    avg_type=\"macro\"\n",
+    ")\n",
+    "summary_weighted, df_weighted_runs = evaluate_test_model(\n",
+    "    best_param_weighted,\n",
+    "    avg_type=\"weighted\"\n",
+    ")\n",
+    "summary_micro, df_micro_runs = evaluate_test_model(\n",
+    "    best_param_micro,\n",
+    "    avg_type=\"micro\"\n",
+    ")\n"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "colo2",
+   "language": "python",
+   "name": "colo2"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.18"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

diagnostics/2026-10-30-lab-hyperparameter-grid/README.md

@@ -0,0 +1,22 @@
+# Lab hyperparameter grid on BRCA (2026-10-30)
+
+These runs test the hyperparameter grid Jubair Rifat suggested on 2026-10-17 (lr 0.01, L2 0.01, dropout 0.3, 200 epochs, SGD with momentum 0.03 or Adam, layer_one 64/128, batch 16/32, activations relu/sigmoid/tanh, single-cell GenePT file). The Workbench's own `train()` was used unchanged.
+
+## Result
+
+- Of 216 configurations, 185 collapsed to a single class.
+- All 108 SGD configurations failed (106 collapsed), and none with a sigmoid activation worked.
+- The 5 best configurations were re-run on 8 seeds. Configuration **D** (Adam, batch 32, projection 64, tanh/relu/tanh) was the most stable, with macro F1 0.749 ± 0.018 and no collapses. It is now the Workbench default.
+- `num_ex_neuron` (extra hidden neurons) isn't supported by the Workbench and was fixed at 0.
+- Only BRCA has been tested.
+
+## Files
+
+- `scripts/build_workspaces.py <cache_dir>`: builds the BRCA workspace with the bulk and single-cell GenePT files.
+- `scripts/grid.py <workspace.pkl> <out.jsonl> <shard> <n_shards>`: runs the grid; it is resumable.
+- `scripts/grid_summary.py <grid.jsonl>`: summarises a grid run.
+- `scripts/seed_check.py <workspace.pkl> <out.jsonl> <shard> <n_shards>`: runs the 8-seed check of the top configurations against the 2026-10-16 baseline.
+- `results/grid_singlecell_seed42.jsonl` and `results/grid_summary.txt`: grid results, seed 42.
+- `results/seed_check_singlecell.jsonl` and `results/seed_check_summary.txt`: 8-seed results.
+
+Run the scripts from the repository root with `PYTHONPATH=$PWD .venv/bin/python ...`.

diagnostics/2026-10-30-lab-hyperparameter-grid/results/grid_shard0.log

@@ -0,0 +1,55 @@
+[0] 0 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 4 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 8 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 12 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[0] 16 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 20 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 24 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 28 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 32 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 36 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 40 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 44 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 48 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 52 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 56 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 60 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 64 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 68 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 72 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 76 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 80 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 84 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[0] 88 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 92 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 96 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 100 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 104 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 108 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 112 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 116 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 120 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 124 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 128 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 132 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 136 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 140 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 144 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 148 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 152 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[0] 156 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 160 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 164 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 168 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 172 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 176 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 180 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 184 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 188 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 192 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 196 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 200 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 204 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 208 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[0] 212 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+SHARD 0 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/grid_shard1.log

@@ -0,0 +1,55 @@
+[1] 1 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.403 classes=4
+[1] 5 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 9 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.442 classes=3
+[1] 13 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 17 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 21 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 25 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.690 classes=5
+[1] 29 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 33 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.657 classes=5
+[1] 37 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 41 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 45 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 49 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 53 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 57 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 61 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 65 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 69 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 73 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.723 classes=5
+[1] 77 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.545 classes=4
+[1] 81 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.741 classes=5
+[1] 85 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 89 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 93 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 97 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.733 classes=5
+[1] 101 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.411 classes=3
+[1] 105 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.752 classes=5
+[1] 109 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 113 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 117 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.789 classes=5
+[1] 121 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 125 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 129 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 133 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 137 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 141 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.701 classes=5
+[1] 145 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 149 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 153 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 157 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 161 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 165 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 169 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 173 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 177 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 181 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.728 classes=5
+[1] 185 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 189 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.727 classes=5
+[1] 193 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 197 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 201 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[1] 205 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.712 classes=5
+[1] 209 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.400 classes=3
+[1] 213 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 16, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.733 classes=5
+SHARD 1 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/grid_shard2.log

@@ -0,0 +1,55 @@
+[2] 2 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 6 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 10 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 14 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 18 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 22 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 26 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 30 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 34 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 38 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 42 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 46 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 50 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 54 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 58 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 62 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 66 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 70 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 74 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 78 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 82 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 86 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 90 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 94 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 98 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 102 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 106 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.047 classes=1
+[2] 110 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 114 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 118 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 122 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 126 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 130 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 134 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 138 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 142 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 146 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 150 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 154 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 158 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 162 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 166 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 170 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 174 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 178 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 182 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=2
+[2] 186 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 190 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+[2] 194 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 198 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 202 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 206 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=2
+[2] 210 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.133 classes=1
+[2] 214 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'SGD', 'seed': 42} -> f1m=0.058 classes=1
+SHARD 2 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/grid_shard3.log

@@ -0,0 +1,55 @@
+[3] 3 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 7 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 11 {'projection_dim': 64, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.737 classes=5
+[3] 15 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 19 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.052 classes=1
+[3] 23 {'projection_dim': 64, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 27 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 31 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 35 {'projection_dim': 64, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.713 classes=5
+[3] 39 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.052 classes=1
+[3] 43 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 47 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 51 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 55 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 59 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.052 classes=1
+[3] 63 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 67 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 71 {'projection_dim': 64, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 75 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 79 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 83 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.778 classes=5
+[3] 87 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 91 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.052 classes=1
+[3] 95 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 99 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.789 classes=5
+[3] 103 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 107 {'projection_dim': 64, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.757 classes=5
+[3] 111 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 115 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.047 classes=1
+[3] 119 {'projection_dim': 128, 'act1': 'relu', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.756 classes=5
+[3] 123 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 127 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 131 {'projection_dim': 128, 'act1': 'relu', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 135 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 139 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.047 classes=1
+[3] 143 {'projection_dim': 128, 'act1': 'relu', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.691 classes=5
+[3] 147 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 151 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 155 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.052 classes=1
+[3] 159 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 163 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 167 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.058 classes=1
+[3] 171 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 175 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.058 classes=1
+[3] 179 {'projection_dim': 128, 'act1': 'sigmoid', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 183 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.789 classes=5
+[3] 187 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.047 classes=1
+[3] 191 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'relu', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.773 classes=5
+[3] 195 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 199 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.133 classes=1
+[3] 203 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'sigmoid', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.058 classes=1
+[3] 207 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'relu', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.725 classes=5
+[3] 211 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'sigmoid', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.444 classes=3
+[3] 215 {'projection_dim': 128, 'act1': 'tanh', 'act2': 'tanh', 'act3': 'tanh', 'batch_size': 32, 'optimizer': 'Adam', 'seed': 42} -> f1m=0.696 classes=5
+SHARD 3 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/grid_summary.txt

@@ -0,0 +1,63 @@
+configs: 216  collapsed (1 class): 185  all 5 classes predicted: 23  macro F1 >= 0.70: 19
+
+TOP 20 by macro F1
+optimizer  batch_size  projection_dim           acts   acc  f1_macro  f1_w  min_recall  best  epochs            pred
+     Adam          32              64 tanh/tanh/relu 0.795     0.789 0.803       0.622    47      59  52/29/13/93/32
+     Adam          16             128 relu/relu/tanh 0.781     0.789 0.790       0.676    42      54  48/30/12/90/39
+     Adam          32             128 tanh/relu/relu 0.799     0.789 0.804       0.568    44      56  52/30/11/97/29
+     Adam          32              64 tanh/relu/tanh 0.776     0.778 0.785       0.649    56      68  48/31/10/93/37
+     Adam          32             128 tanh/relu/tanh 0.781     0.773 0.793       0.730    26      38  49/28/14/88/40
+     Adam          32              64 tanh/tanh/tanh 0.776     0.757 0.789       0.703    28      40  51/27/17/88/36
+     Adam          32             128 relu/relu/tanh 0.781     0.756 0.786       0.568    24      36 43/29/10/104/33
+     Adam          16              64 tanh/tanh/tanh 0.763     0.752 0.768       0.486     8      20 51/29/11/101/27
+     Adam          16              64 tanh/relu/tanh 0.758     0.741 0.768       0.595     8      20  50/27/13/96/33
+     Adam          32              64 relu/relu/tanh 0.763     0.737 0.773       0.622    24      36   48/29/9/96/37
+     Adam          16              64 tanh/tanh/relu 0.749     0.733 0.753       0.459    10      22 50/29/15/100/25
+     Adam          16             128 tanh/tanh/tanh 0.753     0.733 0.759       0.541    24      36 44/31/13/100/31
+     Adam          16             128 tanh/relu/relu 0.763     0.728 0.760       0.378    13      25 49/29/16/108/17
+     Adam          16             128 tanh/relu/tanh 0.767     0.727 0.770       0.486    24      36 52/32/10/104/21
+     Adam          32             128 tanh/tanh/relu 0.753     0.725 0.763       0.545    19      31   55/33/9/79/43
+     Adam          16              64 tanh/relu/relu 0.749     0.723 0.757       0.568    22      34  44/27/18/98/32
+     Adam          32              64 relu/tanh/tanh 0.744     0.713 0.754       0.621    24      36 40/26/12/101/40
+     Adam          16             128 tanh/tanh/relu 0.740     0.712 0.737       0.351    11      23 50/30/16/103/20
+     Adam          16             128 relu/tanh/tanh 0.731     0.701 0.734       0.486    38      50 38/29/11/110/31
+     Adam          32             128 tanh/tanh/tanh 0.726     0.696 0.725       0.351    10      22 53/29/17/100/20
+
+BY optimizer
+             n  f1_mean  f1_max  collapsed  working
+optimizer                                          
+Adam       108    0.270   0.789         79       19
+SGD        108    0.112   0.133        106        0
+
+BY batch_size
+              n  f1_mean  f1_max  collapsed  working
+batch_size                                          
+16          108    0.211   0.789         91       10
+32          108    0.171   0.789         94        9
+
+BY projection_dim
+                  n  f1_mean  f1_max  collapsed  working
+projection_dim                                          
+64              108    0.189   0.789         93        9
+128             108    0.192   0.789         92       10
+
+BY act1
+          n  f1_mean  f1_max  collapsed  working
+act1                                            
+relu     72    0.190   0.789         62        5
+sigmoid  72    0.121   0.133         72        0
+tanh     72    0.262   0.789         51       14
+
+BY act2
+          n  f1_mean  f1_max  collapsed  working
+act2                                            
+relu     72    0.216   0.789         58       10
+sigmoid  72    0.118   0.133         72        0
+tanh     72    0.238   0.789         55        9
+
+BY act3
+          n  f1_mean  f1_max  collapsed  working
+act3                                            
+relu     72    0.186   0.789         61        7
+sigmoid  72    0.144   0.545         68        0
+tanh     72    0.242   0.789         56       12

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_check_singlecell.jsonl

@@ -0,0 +1,48 @@
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 3, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.684931506849315, "f1_macro": 0.6444813586786613, "f1_weighted": 0.6749256063279981, "epochs_completed": 19, "best_epoch": 7, "pred_counts": [44, 28, 6, 122, 19], "collapsed": false, "recall": [0.6896551724137931, 0.8181818181818182, 0.5454545454545454, 0.8073394495412844, 0.24324324324324326]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 1, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.726027397260274, "f1_macro": 0.7144281742354031, "f1_weighted": 0.7332409934786565, "epochs_completed": 25, "best_epoch": 13, "pred_counts": [54, 32, 9, 86, 38], "collapsed": false, "recall": [0.9310344827586207, 0.8787878787878788, 0.6363636363636364, 0.6880733944954128, 0.5675675675675675]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 2, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7397260273972602, "f1_macro": 0.7138206206911931, "f1_weighted": 0.7466811668840856, "epochs_completed": 28, "best_epoch": 16, "pred_counts": [54, 26, 20, 93, 26], "collapsed": false, "recall": [0.896551724137931, 0.7878787878787878, 1.0, 0.7522935779816514, 0.4594594594594595]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 0, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7488584474885844, "f1_macro": 0.7236842897667166, "f1_weighted": 0.7535285657158763, "epochs_completed": 40, "best_epoch": 28, "pred_counts": [48, 28, 9, 105, 29], "collapsed": false, "recall": [0.8275862068965517, 0.8181818181818182, 0.6363636363636364, 0.7889908256880734, 0.5405405405405406]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 2024, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7541439764590951, "f1_weighted": 0.7898422552519037, "epochs_completed": 43, "best_epoch": 31, "pred_counts": [49, 29, 12, 97, 32], "collapsed": false, "recall": [0.8275862068965517, 0.8181818181818182, 0.8181818181818182, 0.8165137614678899, 0.5945945945945946]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 123, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.7359501479955786, "f1_weighted": 0.7626088247020677, "epochs_completed": 49, "best_epoch": 37, "pred_counts": [51, 29, 12, 90, 37], "collapsed": false, "recall": [0.896551724137931, 0.8787878787878788, 0.7272727272727273, 0.7339449541284404, 0.5945945945945946]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 7, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7671232876712328, "f1_macro": 0.7406338069110205, "f1_weighted": 0.7784375791314946, "epochs_completed": 45, "best_epoch": 33, "pred_counts": [50, 24, 14, 96, 35], "collapsed": false, "recall": [0.8275862068965517, 0.7272727272727273, 0.8181818181818182, 0.7889908256880734, 0.6756756756756757]}
+{"config": "A b32 p64 tanh/tanh/relu", "seed": 42, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "tanh", "projection_activation": "tanh", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7945205479452054, "f1_macro": 0.78874693921005, "f1_weighted": 0.8026041446389447, "epochs_completed": 59, "best_epoch": 47, "pred_counts": [52, 29, 13, 93, 32], "collapsed": false, "recall": [0.9310344827586207, 0.8787878787878788, 0.9090909090909091, 0.7798165137614679, 0.6216216216216216]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 3, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7168949771689498, "f1_macro": 0.6634952144629563, "f1_weighted": 0.7175294877460144, "epochs_completed": 48, "best_epoch": 36, "pred_counts": [49, 30, 4, 108, 28], "collapsed": false, "recall": [0.7931034482758621, 0.8181818181818182, 0.36363636363636365, 0.7798165137614679, 0.4864864864864865]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 2, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.730593607305936, "f1_macro": 0.6682730017572305, "f1_weighted": 0.7194362526214491, "epochs_completed": 38, "best_epoch": 26, "pred_counts": [54, 27, 14, 110, 14], "collapsed": false, "recall": [0.896551724137931, 0.8181818181818182, 0.7272727272727273, 0.8256880733944955, 0.24324324324324326]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 0, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7488584474885844, "f1_macro": 0.7283944490396104, "f1_weighted": 0.7570745621268528, "epochs_completed": 39, "best_epoch": 27, "pred_counts": [52, 29, 9, 101, 28], "collapsed": false, "recall": [0.8275862068965517, 0.8484848484848485, 0.6363636363636364, 0.7706422018348624, 0.5675675675675675]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 1, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7625570776255708, "f1_macro": 0.7017018823470437, "f1_weighted": 0.7589129909059207, "epochs_completed": 49, "best_epoch": 37, "pred_counts": [41, 29, 15, 116, 18], "collapsed": false, "recall": [0.7241379310344828, 0.8484848484848485, 0.7272727272727273, 0.8715596330275229, 0.40540540540540543]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 2024, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.7013460186286767, "f1_weighted": 0.7603570398639725, "epochs_completed": 35, "best_epoch": 23, "pred_counts": [45, 29, 12, 99, 34], "collapsed": false, "recall": [0.7931034482758621, 0.7878787878787878, 0.6363636363636364, 0.8073394495412844, 0.5675675675675675]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 7, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.751011131155479, "f1_weighted": 0.7887772320889953, "epochs_completed": 29, "best_epoch": 17, "pred_counts": [50, 28, 8, 105, 28], "collapsed": false, "recall": [0.8275862068965517, 0.7878787878787878, 0.6363636363636364, 0.8348623853211009, 0.6216216216216216]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 123, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.7071455263854697, "f1_weighted": 0.7550850295543412, "epochs_completed": 34, "best_epoch": 22, "pred_counts": [52, 31, 12, 100, 24], "collapsed": false, "recall": [0.896551724137931, 0.9090909090909091, 0.6363636363636364, 0.7889908256880734, 0.43243243243243246]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 2, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.4977168949771689, "f1_macro": 0.1329268292682927, "f1_weighted": 0.3307996436128745, "epochs_completed": 23, "best_epoch": 11, "pred_counts": [0, 0, 0, 219, 0], "collapsed": true, "recall": [0.0, 0.0, 0.0, 1.0, 0.0]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 3, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7397260273972602, "f1_macro": 0.7148184407361271, "f1_weighted": 0.7493292901128092, "epochs_completed": 42, "best_epoch": 30, "pred_counts": [48, 29, 7, 94, 41], "collapsed": false, "recall": [0.7931034482758621, 0.8181818181818182, 0.5454545454545454, 0.7431192660550459, 0.6756756756756757]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 0, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.776255707762557, "f1_macro": 0.7578915715890517, "f1_weighted": 0.7873287183808084, "epochs_completed": 58, "best_epoch": 46, "pred_counts": [54, 29, 9, 91, 36], "collapsed": false, "recall": [0.896551724137931, 0.8484848484848485, 0.6363636363636364, 0.7522935779816514, 0.7297297297297297]}
+{"config": "B b16 p128 relu/relu/tanh", "seed": 42, "hyperparameters": {"epochs": 200, "batch_size": 16, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "relu", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.788652381544761, "f1_weighted": 0.7895010559630641, "epochs_completed": 54, "best_epoch": 42, "pred_counts": [48, 30, 12, 90, 39], "collapsed": false, "recall": [0.8620689655172413, 0.9090909090909091, 0.9090909090909091, 0.7431192660550459, 0.6756756756756757]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 123, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7853881278538812, "f1_macro": 0.7597321666231736, "f1_weighted": 0.7932886529725107, "epochs_completed": 51, "best_epoch": 39, "pred_counts": [48, 29, 15, 97, 30], "collapsed": false, "recall": [0.8620689655172413, 0.8787878787878788, 0.8181818181818182, 0.7889908256880734, 0.6216216216216216]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 2024, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7945205479452054, "f1_macro": 0.7775926070493133, "f1_weighted": 0.8037755930922895, "epochs_completed": 46, "best_epoch": 34, "pred_counts": [45, 29, 10, 100, 35], "collapsed": false, "recall": [0.7586206896551724, 0.8181818181818182, 0.8181818181818182, 0.8348623853211009, 0.6756756756756757]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 7, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7168949771689498, "f1_macro": 0.6666606642657062, "f1_weighted": 0.7075349317809315, "epochs_completed": 65, "best_epoch": 53, "pred_counts": [56, 27, 9, 115, 12], "collapsed": false, "recall": [0.8620689655172413, 0.7575757575757576, 0.6363636363636364, 0.8256880733944955, 0.2702702702702703]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 1, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.8127853881278538, "f1_macro": 0.8004969316657136, "f1_weighted": 0.8177554227696595, "epochs_completed": 65, "best_epoch": 53, "pred_counts": [50, 31, 9, 98, 31], "collapsed": false, "recall": [0.9655172413793104, 0.9090909090909091, 0.7272727272727273, 0.8073394495412844, 0.6486486486486487]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 2, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7899543378995434, "f1_macro": 0.7647661601172038, "f1_weighted": 0.7982709673173486, "epochs_completed": 58, "best_epoch": 46, "pred_counts": [51, 29, 16, 95, 28], "collapsed": false, "recall": [0.896551724137931, 0.8484848484848485, 0.9090909090909091, 0.7981651376146789, 0.5945945945945946]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 3, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7579908675799086, "f1_macro": 0.7450983463384329, "f1_weighted": 0.7676536929640603, "epochs_completed": 59, "best_epoch": 47, "pred_counts": [44, 27, 7, 100, 41], "collapsed": false, "recall": [0.7241379310344828, 0.7878787878787878, 0.6363636363636364, 0.7798165137614679, 0.7297297297297297]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 0, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7579908675799086, "f1_macro": 0.7305752281147861, "f1_weighted": 0.7672308198023153, "epochs_completed": 40, "best_epoch": 28, "pred_counts": [46, 28, 10, 98, 37], "collapsed": false, "recall": [0.7586206896551724, 0.8181818181818182, 0.6363636363636364, 0.7706422018348624, 0.7027027027027027]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 2024, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.738332484131178, "f1_weighted": 0.7656099462973935, "epochs_completed": 42, "best_epoch": 30, "pred_counts": [52, 28, 10, 93, 36], "collapsed": false, "recall": [0.8275862068965517, 0.8181818181818182, 0.7272727272727273, 0.7614678899082569, 0.6216216216216216]}
+{"config": "C b32 p128 tanh/relu/relu", "seed": 42, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "relu", "sgd_momentum": 0.0}, "accuracy": 0.7990867579908676, "f1_macro": 0.7885889251585044, "f1_weighted": 0.8040701819868671, "epochs_completed": 56, "best_epoch": 44, "pred_counts": [52, 30, 11, 97, 29], "collapsed": false, "recall": [0.9655172413793104, 0.9090909090909091, 0.8181818181818182, 0.7981651376146789, 0.5675675675675675]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 123, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7359225592888959, "f1_weighted": 0.791296097847027, "epochs_completed": 55, "best_epoch": 43, "pred_counts": [55, 33, 10, 93, 28], "collapsed": false, "recall": [0.896551724137931, 0.9090909090909091, 0.5454545454545454, 0.7798165137614679, 0.6486486486486487]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 7, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.7357654892283638, "f1_weighted": 0.765273263266187, "epochs_completed": 44, "best_epoch": 32, "pred_counts": [50, 24, 13, 96, 36], "collapsed": false, "recall": [0.7931034482758621, 0.7272727272727273, 0.8181818181818182, 0.7706422018348624, 0.6756756756756757]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 2, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7595667299177735, "f1_weighted": 0.7904695167211089, "epochs_completed": 32, "best_epoch": 20, "pred_counts": [51, 29, 13, 95, 31], "collapsed": false, "recall": [0.8620689655172413, 0.8484848484848485, 0.8181818181818182, 0.7889908256880734, 0.6216216216216216]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 3, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7123287671232876, "f1_macro": 0.7019184281569603, "f1_weighted": 0.7172640652092707, "epochs_completed": 47, "best_epoch": 35, "pred_counts": [48, 27, 7, 109, 28], "collapsed": false, "recall": [0.7586206896551724, 0.7878787878787878, 0.6363636363636364, 0.7706422018348624, 0.4594594594594595]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 1, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7666271513372672, "f1_weighted": 0.7841213035584601, "epochs_completed": 46, "best_epoch": 34, "pred_counts": [53, 32, 9, 101, 24], "collapsed": false, "recall": [0.9310344827586207, 0.9090909090909091, 0.7272727272727273, 0.7981651376146789, 0.5135135135135135]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 123, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7671232876712328, "f1_macro": 0.7396749831626857, "f1_weighted": 0.7699385457307235, "epochs_completed": 34, "best_epoch": 22, "pred_counts": [52, 29, 14, 102, 22], "collapsed": false, "recall": [0.896551724137931, 0.8787878787878788, 0.8181818181818182, 0.7981651376146789, 0.4594594594594595]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 0, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7579908675799086, "f1_macro": 0.743470919668342, "f1_weighted": 0.7665191757209968, "epochs_completed": 57, "best_epoch": 45, "pred_counts": [48, 29, 8, 100, 34], "collapsed": false, "recall": [0.7931034482758621, 0.8484848484848485, 0.6363636363636364, 0.7706422018348624, 0.6486486486486487]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 2024, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7990867579908676, "f1_macro": 0.7616315451274192, "f1_weighted": 0.8058428826086493, "epochs_completed": 35, "best_epoch": 23, "pred_counts": [45, 29, 10, 106, 29], "collapsed": false, "recall": [0.7586206896551724, 0.8181818181818182, 0.7272727272727273, 0.8715596330275229, 0.6216216216216216]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 7, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7534246575342466, "f1_macro": 0.7446451031909788, "f1_weighted": 0.7653821355518723, "epochs_completed": 33, "best_epoch": 21, "pred_counts": [51, 25, 12, 93, 38], "collapsed": false, "recall": [0.8275862068965517, 0.7575757575757576, 0.8181818181818182, 0.7522935779816514, 0.6756756756756757]}
+{"config": "D b32 p64 tanh/relu/tanh", "seed": 42, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.776255707762557, "f1_macro": 0.7776900614771901, "f1_weighted": 0.7854714654440004, "epochs_completed": 68, "best_epoch": 56, "pred_counts": [48, 31, 10, 93, 37], "collapsed": false, "recall": [0.8275862068965517, 0.9090909090909091, 0.8181818181818182, 0.7614678899082569, 0.6486486486486487]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 1, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7671232876712328, "f1_macro": 0.7387577790490057, "f1_weighted": 0.7727522475446735, "epochs_completed": 34, "best_epoch": 22, "pred_counts": [53, 34, 7, 92, 33], "collapsed": false, "recall": [0.9310344827586207, 0.9090909090909091, 0.5454545454545454, 0.7522935779816514, 0.6216216216216216]}
+{"config": "E b32 p128 tanh/relu/tanh", "seed": 42, "hyperparameters": {"epochs": 200, "batch_size": 32, "learning_rate": 0.01, "weight_decay": 0.01, "dropout": 0.3, "projection_dim": 128, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "tanh", "projection_activation": "relu", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7734406826426466, "f1_weighted": 0.7928328973182098, "epochs_completed": 38, "best_epoch": 26, "pred_counts": [49, 28, 14, 88, 40], "collapsed": false, "recall": [0.8275862068965517, 0.8484848484848485, 0.9090909090909091, 0.7522935779816514, 0.7297297297297297]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 0, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 0, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7625570776255708, "f1_macro": 0.7241977003359705, "f1_weighted": 0.7688921750314536, "epochs_completed": 61, "best_epoch": 49, "pred_counts": [47, 28, 8, 105, 31], "collapsed": false, "recall": [0.7931034482758621, 0.8181818181818182, 0.5454545454545454, 0.8073394495412844, 0.6216216216216216]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 2, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.8082191780821918, "f1_macro": 0.792199067264165, "f1_weighted": 0.8205970478829074, "epochs_completed": 81, "best_epoch": 69, "pred_counts": [53, 29, 14, 91, 32], "collapsed": false, "recall": [0.896551724137931, 0.8484848484848485, 0.9090909090909091, 0.7981651376146789, 0.7027027027027027]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 3, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 3, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7397260273972602, "f1_macro": 0.7190529426628272, "f1_weighted": 0.7474937309171131, "epochs_completed": 83, "best_epoch": 71, "pred_counts": [45, 28, 6, 102, 38], "collapsed": false, "recall": [0.7586206896551724, 0.8181818181818182, 0.5454545454545454, 0.7706422018348624, 0.6216216216216216]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 7, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 7, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.771689497716895, "f1_macro": 0.7524499679210638, "f1_weighted": 0.7826279144428353, "epochs_completed": 44, "best_epoch": 32, "pred_counts": [51, 28, 8, 91, 41], "collapsed": false, "recall": [0.8620689655172413, 0.7878787878787878, 0.6363636363636364, 0.7614678899082569, 0.7567567567567568]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 1, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 1, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7808219178082192, "f1_macro": 0.7513846597558472, "f1_weighted": 0.7897822474196114, "epochs_completed": 62, "best_epoch": 50, "pred_counts": [49, 30, 12, 98, 30], "collapsed": false, "recall": [0.8275862068965517, 0.8484848484848485, 0.7272727272727273, 0.7981651376146789, 0.6486486486486487]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 123, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 123, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7625570776255708, "f1_macro": 0.7396509109600558, "f1_weighted": 0.773059563038045, "epochs_completed": 49, "best_epoch": 37, "pred_counts": [53, 29, 13, 90, 34], "collapsed": false, "recall": [0.896551724137931, 0.8787878787878788, 0.7272727272727273, 0.7431192660550459, 0.6216216216216216]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 2024, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 2024, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7579908675799086, "f1_macro": 0.7372566114501599, "f1_weighted": 0.7695326473497044, "epochs_completed": 53, "best_epoch": 41, "pred_counts": [48, 29, 10, 95, 37], "collapsed": false, "recall": [0.7586206896551724, 0.8181818181818182, 0.7272727272727273, 0.7798165137614679, 0.6486486486486487]}
+{"config": "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh", "seed": 42, "hyperparameters": {"epochs": 100, "batch_size": 16, "learning_rate": 0.001, "weight_decay": 0.0001, "dropout": 0.3, "projection_dim": 64, "fusion_dim": 12, "validation_fraction": 0.25, "optimizer": "Adam", "class_weighting": true, "early_stopping_patience": 12, "seed": 42, "biological_activation": "relu", "projection_activation": "sigmoid", "fusion_activation": "tanh", "sgd_momentum": 0.0}, "accuracy": 0.7671232876712328, "f1_macro": 0.7474258785478297, "f1_weighted": 0.7740214080920172, "epochs_completed": 50, "best_epoch": 38, "pred_counts": [46, 32, 16, 96, 29], "collapsed": false, "recall": [0.7586206896551724, 0.9393939393939394, 0.9090909090909091, 0.7706422018348624, 0.5675675675675675]}

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_check_summary.txt

@@ -0,0 +1,25 @@
+                                                acc       f1_macro           f1_w    f1m min/max  collapsed  min recall (mean)  best epoch  max epochs run  recall c0  recall c1  recall c2  recall c3  recall c4
+config                                                                                                                                                                                                           
+D b32 p64 tanh/relu/tanh              0.769 ± 0.015  0.749 ± 0.018  0.778 ± 0.013  0.731 / 0.778          0              0.609        39.5              68       0.83       0.84       0.73       0.78       0.64
+BASE lr1e-3 wd1e-4 relu/sigmoid/tanh  0.769 ± 0.020  0.745 ± 0.022  0.778 ± 0.021  0.719 / 0.792          0              0.615        48.4              83       0.82       0.84       0.72       0.78       0.65
+E b32 p128 tanh/relu/tanh             0.765 ± 0.026  0.745 ± 0.021  0.773 ± 0.027  0.702 / 0.773          0              0.594        26.8              57       0.83       0.84       0.74       0.78       0.60
+A b32 p64 tanh/tanh/relu              0.749 ± 0.034  0.727 ± 0.041  0.755 ± 0.040  0.644 / 0.789          0              0.537        26.5              59       0.85       0.83       0.76       0.77       0.54
+B b16 p128 relu/relu/tanh             0.753 ± 0.022  0.714 ± 0.042  0.756 ± 0.027  0.663 / 0.789          0              0.485        28.8              54       0.83       0.84       0.66       0.80       0.50
+C b32 p128 tanh/relu/relu             0.740 ± 0.103  0.675 ± 0.223  0.724 ± 0.163  0.133 / 0.800          1              0.496        38.8              65       0.76       0.74       0.62       0.82       0.52
+
+A b32 p64 tanh/tanh/relu     vs BASE (same seed): mean Δ macro F1 -0.018, better on 2/8 seeds, per-seed [-0.001, -0.037, -0.078, -0.075, -0.012, 0.041, -0.004, 0.017]
+B b16 p128 relu/relu/tanh    vs BASE (same seed): mean Δ macro F1 -0.032, better on 2/8 seeds, per-seed [0.004, -0.05, -0.124, -0.056, -0.001, 0.041, -0.033, -0.036]
+C b32 p128 tanh/relu/relu    vs BASE (same seed): mean Δ macro F1 -0.071, better on 5/8 seeds, per-seed [0.034, 0.049, -0.659, -0.004, -0.086, 0.041, 0.02, 0.04]
+D b32 p64 tanh/relu/tanh     vs BASE (same seed): mean Δ macro F1 +0.004, better on 5/8 seeds, per-seed [0.006, 0.015, -0.027, 0.026, -0.017, 0.03, -0.004, 0.001]
+E b32 p128 tanh/relu/tanh    vs BASE (same seed): mean Δ macro F1 -0.000, better on 4/8 seeds, per-seed [0.019, -0.013, -0.033, -0.017, -0.008, 0.026, 0.0, 0.024]
+
+config  A b32 p64 tanh/tanh/relu  B b16 p128 relu/relu/tanh  BASE lr1e-3 wd1e-4 relu/sigmoid/tanh  C b32 p128 tanh/relu/relu  D b32 p64 tanh/relu/tanh  E b32 p128 tanh/relu/tanh
+seed                                                                                                                                                                             
+0                          0.724                      0.728                                 0.724                      0.758                     0.731                      0.743
+1                          0.714                      0.702                                 0.751                      0.800                     0.767                      0.739
+2                          0.714                      0.668                                 0.792                      0.133                     0.765                      0.760
+3                          0.644                      0.663                                 0.719                      0.715                     0.745                      0.702
+7                          0.741                      0.751                                 0.752                      0.667                     0.736                      0.745
+42                         0.789                      0.789                                 0.747                      0.789                     0.778                      0.773
+123                        0.736                      0.707                                 0.740                      0.760                     0.736                      0.740
+2024                       0.754                      0.701                                 0.737                      0.778                     0.738                      0.762

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_shard0.log

@@ -0,0 +1,13 @@
+[0] A b32 p64 tanh/tanh/relu seed=0 f1m=0.724 collapsed=False
+[0] A b32 p64 tanh/tanh/relu seed=7 f1m=0.741 collapsed=False
+[0] B b16 p128 relu/relu/tanh seed=0 f1m=0.728 collapsed=False
+[0] B b16 p128 relu/relu/tanh seed=7 f1m=0.751 collapsed=False
+[0] C b32 p128 tanh/relu/relu seed=0 f1m=0.758 collapsed=False
+[0] C b32 p128 tanh/relu/relu seed=7 f1m=0.667 collapsed=False
+[0] D b32 p64 tanh/relu/tanh seed=0 f1m=0.731 collapsed=False
+[0] D b32 p64 tanh/relu/tanh seed=7 f1m=0.736 collapsed=False
+[0] E b32 p128 tanh/relu/tanh seed=0 f1m=0.743 collapsed=False
+[0] E b32 p128 tanh/relu/tanh seed=7 f1m=0.745 collapsed=False
+[0] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=0 f1m=0.724 collapsed=False
+[0] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=7 f1m=0.752 collapsed=False
+SHARD 0 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_shard1.log

@@ -0,0 +1,13 @@
+[1] A b32 p64 tanh/tanh/relu seed=1 f1m=0.714 collapsed=False
+[1] A b32 p64 tanh/tanh/relu seed=42 f1m=0.789 collapsed=False
+[1] B b16 p128 relu/relu/tanh seed=1 f1m=0.702 collapsed=False
+[1] B b16 p128 relu/relu/tanh seed=42 f1m=0.789 collapsed=False
+[1] C b32 p128 tanh/relu/relu seed=1 f1m=0.800 collapsed=False
+[1] C b32 p128 tanh/relu/relu seed=42 f1m=0.789 collapsed=False
+[1] D b32 p64 tanh/relu/tanh seed=1 f1m=0.767 collapsed=False
+[1] D b32 p64 tanh/relu/tanh seed=42 f1m=0.778 collapsed=False
+[1] E b32 p128 tanh/relu/tanh seed=1 f1m=0.739 collapsed=False
+[1] E b32 p128 tanh/relu/tanh seed=42 f1m=0.773 collapsed=False
+[1] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=1 f1m=0.751 collapsed=False
+[1] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=42 f1m=0.747 collapsed=False
+SHARD 1 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_shard2.log

@@ -0,0 +1,13 @@
+[2] A b32 p64 tanh/tanh/relu seed=2 f1m=0.714 collapsed=False
+[2] A b32 p64 tanh/tanh/relu seed=123 f1m=0.736 collapsed=False
+[2] B b16 p128 relu/relu/tanh seed=2 f1m=0.668 collapsed=False
+[2] B b16 p128 relu/relu/tanh seed=123 f1m=0.707 collapsed=False
+[2] C b32 p128 tanh/relu/relu seed=2 f1m=0.133 collapsed=True
+[2] C b32 p128 tanh/relu/relu seed=123 f1m=0.760 collapsed=False
+[2] D b32 p64 tanh/relu/tanh seed=2 f1m=0.765 collapsed=False
+[2] D b32 p64 tanh/relu/tanh seed=123 f1m=0.736 collapsed=False
+[2] E b32 p128 tanh/relu/tanh seed=2 f1m=0.760 collapsed=False
+[2] E b32 p128 tanh/relu/tanh seed=123 f1m=0.740 collapsed=False
+[2] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=2 f1m=0.792 collapsed=False
+[2] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=123 f1m=0.740 collapsed=False
+SHARD 2 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/results/seed_shard3.log

@@ -0,0 +1,13 @@
+[3] A b32 p64 tanh/tanh/relu seed=3 f1m=0.644 collapsed=False
+[3] A b32 p64 tanh/tanh/relu seed=2024 f1m=0.754 collapsed=False
+[3] B b16 p128 relu/relu/tanh seed=3 f1m=0.663 collapsed=False
+[3] B b16 p128 relu/relu/tanh seed=2024 f1m=0.701 collapsed=False
+[3] C b32 p128 tanh/relu/relu seed=3 f1m=0.715 collapsed=False
+[3] C b32 p128 tanh/relu/relu seed=2024 f1m=0.778 collapsed=False
+[3] D b32 p64 tanh/relu/tanh seed=3 f1m=0.745 collapsed=False
+[3] D b32 p64 tanh/relu/tanh seed=2024 f1m=0.738 collapsed=False
+[3] E b32 p128 tanh/relu/tanh seed=3 f1m=0.702 collapsed=False
+[3] E b32 p128 tanh/relu/tanh seed=2024 f1m=0.762 collapsed=False
+[3] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=3 f1m=0.719 collapsed=False
+[3] BASE lr1e-3 wd1e-4 relu/sigmoid/tanh seed=2024 f1m=0.737 collapsed=False
+SHARD 3 DONE

diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/build_workspaces.py

@@ -0,0 +1,30 @@
+"""Build and cache the BRCA workspace with both GenePT files, using the app's own prepare step.
+Usage: build_workspaces.py <cache_dir>"""
+import pickle
+import sys
+from pathlib import Path
+
+import numpy as np
+
+import app
+
+CACHE = Path(sys.argv[1])
+CACHE.mkdir(parents=True, exist_ok=True)
+OPTIONS = {
+    "bulk": "Bulk · large-3 context",
+    "singlecell": "Single-cell · cell type, tissue, drug & pathway",
+}
+for key, option in OPTIONS.items():
+    result = app.prepare_workspace(
+        "BioLM-NET examples", "BRCA", "", None, None, None, None, None, None, None,
+        True, option, progress=lambda *a, **k: None,
+    )
+    workspace = result[0]
+    if workspace is None:
+        sys.exit(f"{key}: {result[1]}")
+    (CACHE / f"workspace_BRCA_{key}.pkl").write_bytes(pickle.dumps(workspace))
+    for name, branch in (("gene", workspace.gene_branch), ("dna", workspace.dna_branch)):
+        m = branch.biological_mask
+        print(f"{key:10s} {name}: inputs={m.shape[0]} hidden genes={m.shape[1]} edges={(m > 0).sum()} "
+              f"median in-degree={np.median((m > 0).sum(0)):.0f} pathways={len(branch.pathways)} "
+              f"GenePT dim={branch.embeddings.shape[1]}", flush=True)

diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/grid.py

@@ -0,0 +1,70 @@
+"""Jubair's hyperparameter grid (email of 2026-10-17) on BRCA, using the Workbench's train() unchanged.
+
+Usage: grid.py <workspace.pkl> <out.jsonl> <shard> <n_shards> [seed]
+Resumable: configs already in <out.jsonl> are skipped.
+
+Grid as sent, mapped onto Workbench Hyperparameters:
+  layer_one -> projection_dim; activation_function_one/two/three -> biological/projection/fusion activation;
+  l2_reg -> weight_decay; epoch -> epochs (early stopping, patience 12, still applies);
+  SGD gets momentum 0.03 as instructed. num_ex_neuron is not supported by the Workbench (fixed at 0).
+"""
+import itertools
+import json
+import sys
+from dataclasses import asdict, replace
+from pathlib import Path
+
+import numpy as np
+import pickle
+import torch
+from sklearn.metrics import recall_score
+
+import biolmnet.training as training
+
+WORKSPACE, OUT, SHARD, SHARDS = Path(sys.argv[1]), Path(sys.argv[2]), int(sys.argv[3]), int(sys.argv[4])
+SEED = int(sys.argv[5]) if len(sys.argv) > 5 else 42
+torch.set_num_threads(2)
+
+ACT = ["relu", "sigmoid", "tanh"]
+GRID = list(itertools.product([64, 128], ACT, ACT, ACT, [16, 32], ["SGD", "Adam"]))
+
+workspace = pickle.loads(WORKSPACE.read_bytes())
+names = workspace.label_names
+done = set()
+if OUT.exists():
+    done = {json.dumps(json.loads(line)["key"]) for line in OUT.read_text().splitlines() if line.strip()}
+
+for index, (proj, a1, a2, a3, batch, opt) in enumerate(GRID):
+    if index % SHARDS != SHARD:
+        continue
+    key = {"projection_dim": proj, "act1": a1, "act2": a2, "act3": a3, "batch_size": batch, "optimizer": opt, "seed": SEED}
+    if json.dumps(key) in done:
+        continue
+    hp = replace(
+        training.Hyperparameters(),
+        epochs=200, batch_size=batch, learning_rate=0.01, weight_decay=0.01, dropout=0.3,
+        projection_dim=proj, optimizer=opt, sgd_momentum=0.03 if opt == "SGD" else 0.0,
+        biological_activation=a1, projection_activation=a2, fusion_activation=a3, seed=SEED,
+    )
+    try:
+        result = training.train(workspace, hp)
+    except Exception as exc:  # record and keep going
+        row = {"key": key, "error": repr(exc)}
+    else:
+        m = result.bundle.metrics
+        vp = result.validation_predictions
+        counts = vp["predicted"].value_counts().reindex(names, fill_value=0).astype(int).tolist()
+        history = result.bundle.history
+        best_epoch = int(min(history, key=lambda h: h["validation_loss"])["epoch"])
+        row = {
+            "key": key, "hyperparameters": asdict(hp),
+            "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"],
+            "epochs_completed": m["epochs_completed"], "best_epoch": best_epoch,
+            "pred_counts": counts, "classes_predicted": int((np.array(counts) > 0).sum()),
+            "recall": recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0).tolist(),
+            "final_train_loss": history[-1]["training_loss"], "best_val_loss": m["validation_loss"],
+        }
+    with OUT.open("a") as handle:
+        handle.write(json.dumps(row) + "\n")
+    print(f"[{SHARD}] {index} {key} -> " + (row.get("error") or f"f1m={row['f1_macro']:.3f} classes={row['classes_predicted']}"), flush=True)
+print(f"SHARD {SHARD} DONE", flush=True)

diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/grid_summary.py

@@ -0,0 +1,31 @@
+"""Summarise grid.py output. Usage: grid_summary.py <grid.jsonl>"""
+import json
+import sys
+
+import pandas as pd
+
+rows = []
+for line in open(sys.argv[1]):
+    r = json.loads(line)
+    if "error" in r:
+        print("ERROR", r["key"], r["error"])
+        continue
+    rows.append({**r["key"], "acc": r["accuracy"], "f1_macro": r["f1_macro"], "f1_w": r["f1_weighted"],
+                 "epochs": r["epochs_completed"], "best": r["best_epoch"], "classes": r["classes_predicted"],
+                 "collapsed": r["classes_predicted"] == 1, "pred": "/".join(map(str, r["pred_counts"])),
+                 "min_recall": min(r["recall"])})
+df = pd.DataFrame(rows)
+df["acts"] = df.act1 + "/" + df.act2 + "/" + df.act3
+pd.set_option("display.width", 220)
+pd.set_option("display.max_rows", 300)
+
+print(f"configs: {len(df)}  collapsed (1 class): {df.collapsed.sum()}  "
+      f"all 5 classes predicted: {(df.classes == 5).sum()}  macro F1 >= 0.70: {(df.f1_macro >= 0.70).sum()}")
+print("\nTOP 20 by macro F1")
+print(df.sort_values("f1_macro", ascending=False).head(20)[
+    ["optimizer", "batch_size", "projection_dim", "acts", "acc", "f1_macro", "f1_w", "min_recall", "best", "epochs", "pred"]
+].round(3).to_string(index=False))
+for factor in ["optimizer", "batch_size", "projection_dim", "act1", "act2", "act3"]:
+    g = df.groupby(factor).agg(n=("f1_macro", "size"), f1_mean=("f1_macro", "mean"), f1_max=("f1_macro", "max"),
+                               collapsed=("collapsed", "sum"), working=("f1_macro", lambda s: int((s >= 0.70).sum())))
+    print(f"\nBY {factor}\n" + g.round(3).to_string())

diagnostics/2026-10-30-lab-hyperparameter-grid/scripts/seed_check.py

@@ -0,0 +1,62 @@
+"""Re-run the best grid configurations (plus the 2026-10-16 stable baseline) across 8 seeds.
+Usage: seed_check.py <workspace.pkl> <out.jsonl> <shard> <n_shards>
+Resumable: (config, seed) pairs already in <out.jsonl> are skipped."""
+import json
+import pickle
+import sys
+from dataclasses import asdict, replace
+from pathlib import Path
+
+import numpy as np
+import torch
+from sklearn.metrics import recall_score
+
+import biolmnet.training as training
+
+WORKSPACE, OUT, SHARD, SHARDS = Path(sys.argv[1]), Path(sys.argv[2]), int(sys.argv[3]), int(sys.argv[4])
+torch.set_num_threads(2)
+SEEDS = [0, 1, 2, 3, 7, 42, 123, 2024]
+
+GRID_BASE = dict(epochs=200, learning_rate=0.01, weight_decay=0.01, dropout=0.3, optimizer="Adam")
+CONFIGS = {
+    "A b32 p64 tanh/tanh/relu": dict(GRID_BASE, batch_size=32, projection_dim=64,
+                                     biological_activation="tanh", projection_activation="tanh", fusion_activation="relu"),
+    "B b16 p128 relu/relu/tanh": dict(GRID_BASE, batch_size=16, projection_dim=128,
+                                      biological_activation="relu", projection_activation="relu", fusion_activation="tanh"),
+    "C b32 p128 tanh/relu/relu": dict(GRID_BASE, batch_size=32, projection_dim=128,
+                                      biological_activation="tanh", projection_activation="relu", fusion_activation="relu"),
+    "D b32 p64 tanh/relu/tanh": dict(GRID_BASE, batch_size=32, projection_dim=64,
+                                     biological_activation="tanh", projection_activation="relu", fusion_activation="tanh"),
+    "E b32 p128 tanh/relu/tanh": dict(GRID_BASE, batch_size=32, projection_dim=128,
+                                      biological_activation="tanh", projection_activation="relu", fusion_activation="tanh"),
+    # 2026-10-16 stable setting: Adam lr 1e-3, weight_decay 1e-4, published activations, 100-epoch limit.
+    "BASE lr1e-3 wd1e-4 relu/sigmoid/tanh": dict(epochs=100, weight_decay=1e-4),
+}
+
+workspace = pickle.loads(WORKSPACE.read_bytes())
+names = workspace.label_names
+done = set()
+if OUT.exists():
+    done = {(r["config"], r["seed"]) for r in map(json.loads, OUT.read_text().splitlines()) if r}
+
+jobs = [(c, s) for c in CONFIGS for s in SEEDS]
+for index, (config, seed) in enumerate(jobs):
+    if index % SHARDS != SHARD or (config, seed) in done:
+        continue
+    hp = replace(training.Hyperparameters(), seed=seed, **CONFIGS[config])
+    result = training.train(workspace, hp)
+    m = result.bundle.metrics
+    vp = result.validation_predictions
+    counts = vp["predicted"].value_counts().reindex(names, fill_value=0).astype(int).tolist()
+    row = {
+        "config": config, "seed": seed, "hyperparameters": asdict(hp),
+        "accuracy": m["accuracy"], "f1_macro": m["f1_macro"], "f1_weighted": m["f1_weighted"],
+        "epochs_completed": m["epochs_completed"],
+        "best_epoch": int(min(result.bundle.history, key=lambda h: h["validation_loss"])["epoch"]),
+        "pred_counts": counts, "collapsed": int((np.array(counts) > 0).sum()) == 1,
+        "recall": recall_score(vp["observed"], vp["predicted"], labels=names, average=None, zero_division=0).tolist(),
+    }
+    with OUT.open("a") as handle:
+        handle.write(json.dumps(row) + "\n")
+    print(f"[{SHARD}] {config} seed={seed} f1m={row['f1_macro']:.3f} collapsed={row['collapsed']}", flush=True)
+print(f"SHARD {SHARD} DONE", flush=True)

tests/test_core.py

@@ -1,9 +1,11 @@
 from __future__ import annotations
 
 import pickle
+from dataclasses import replace
 
 import numpy as np
 import pandas as pd
+import pytest
 import torch
 
 from biolmnet.artifacts import load_bundle, save_bundle
@@ -15,7 +17,13 @@ from biolmnet.data import (
     deterministic_gene_embeddings,
 )
 from biolmnet.model import BioMaskedLinear, BioLMNet
-from biolmnet.training import Hyperparameters, predict, train
+from biolmnet.training import (
+    DEFAULT_VALIDATION_FRACTION,
+    Hyperparameters,
+    predict,
+    split_train_validation,
+    train,
+)
 
 
 def test_biological_mask_uses_pdi_and_undirected_ppi() -> None:
@@ -232,3 +240,217 @@ def test_zerogpu_duration_estimator_is_bounded_and_scales() -> None:
     long = estimate_training_duration(common[0], 200, *common[1:])
     assert 30 <= short <= 300
     assert short <= long <= 300
+
+
+_FAST = Hyperparameters(
+    epochs=3,
+    batch_size=8,
+    projection_dim=4,
+    fusion_dim=3,
+    dropout=0.0,
+    early_stopping_patience=3,
+)
+
+
+def test_default_split_is_75_training_25_validation() -> None:
+    assert DEFAULT_VALIDATION_FRACTION == 0.25
+    assert Hyperparameters().validation_fraction == 0.25
+    labels = np.repeat(np.array([0, 1]), 50)
+    train_index, validation_index = split_train_validation(
+        labels, Hyperparameters().validation_fraction, Hyperparameters().seed
+    )
+    assert len(train_index) == 75
+    assert len(validation_index) == 25
+
+
+def test_split_partitions_every_sample_exactly_once() -> None:
+    labels = np.repeat(np.array([0, 1, 2]), [40, 25, 18])
+    train_index, validation_index = split_train_validation(labels, 0.25, 42)
+    assert len(train_index) + len(validation_index) == len(labels)
+    assert set(train_index).isdisjoint(validation_index)
+    assert sorted(np.concatenate([train_index, validation_index])) == list(range(len(labels)))
+
+
+def test_split_preserves_class_proportions() -> None:
+    labels = np.repeat(np.array([0, 1, 2]), [120, 60, 20])  # 60 / 30 / 10
+    train_index, validation_index = split_train_validation(labels, 0.25, 42)
+    expected = np.array([0.6, 0.3, 0.1])
+    for partition in (train_index, validation_index):
+        observed = np.bincount(labels[partition], minlength=3) / len(partition)
+        np.testing.assert_allclose(observed, expected, atol=0.02)
+
+
+def test_split_is_reproducible_for_a_fixed_seed() -> None:
+    labels = np.repeat(np.array([0, 1]), [30, 34])
+    first = split_train_validation(labels, 0.25, 42)
+    second = split_train_validation(labels, 0.25, 42)
+    other_seed = split_train_validation(labels, 0.25, 7)
+    np.testing.assert_array_equal(first[0], second[0])
+    np.testing.assert_array_equal(first[1], second[1])
+    assert set(first[1]) != set(other_seed[1])
+
+
+def test_split_rejects_partitions_too_small_to_stratify() -> None:
+    with pytest.raises(ValueError, match="at least two samples"):
+        split_train_validation(np.array([0, 0, 0, 1]), 0.25, 42)
+    # 4 classes x 2 samples: 25% leaves 2 validation samples for 4 classes.
+    with pytest.raises(ValueError, match="one sample per class"):
+        split_train_validation(np.repeat(np.arange(4), 2), 0.25, 42)
+
+
+def test_training_uses_paired_rows_and_train_only_scaler() -> None:
+    workspace = _tiny_workspace()
+    hyperparameters = replace(_FAST, validation_fraction=0.25)
+    train_index, validation_index = split_train_validation(
+        workspace.labels, hyperparameters.validation_fraction, hyperparameters.seed
+    )
+    # Shift only the validation rows: a scaler fitted on all samples would move.
+    workspace.gene_expression[validation_index] += 100.0
+    workspace.dna_methylation[validation_index] -= 100.0
+
+    result = train(workspace, hyperparameters)
+    bundle = result.bundle
+
+    assert bundle.metrics["training_samples"] == 36
+    assert bundle.metrics["validation_samples"] == 12
+    np.testing.assert_array_equal(
+        result.validation_predictions["sample_row"].to_numpy(), validation_index
+    )
+    expected_labels = [workspace.label_names[value] for value in workspace.labels[validation_index]]
+    assert result.validation_predictions["observed"].tolist() == expected_labels
+
+    # Both omics scalers come from the same training rows only.
+    np.testing.assert_allclose(
+        bundle.gene_mean, workspace.gene_expression[train_index].mean(axis=0), rtol=1e-5
+    )
+    np.testing.assert_allclose(
+        bundle.gene_scale, workspace.gene_expression[train_index].std(axis=0), rtol=1e-5
+    )
+    np.testing.assert_allclose(
+        bundle.dna_mean, workspace.dna_methylation[train_index].mean(axis=0), rtol=1e-5
+    )
+    np.testing.assert_allclose(
+        bundle.dna_scale, workspace.dna_methylation[train_index].std(axis=0), rtol=1e-5
+    )
+
+    # Validation rows are transformed with that training-fitted scaler: the
+    # stored validation probabilities match predict(), which uses the bundle's scaler.
+    rescored = predict(
+        pd.DataFrame(workspace.gene_expression[validation_index], columns=bundle.gene_features),
+        pd.DataFrame(workspace.dna_methylation[validation_index], columns=bundle.dna_features),
+        bundle,
+    )
+    probability_columns = [f"P({label})" for label in workspace.label_names]
+    np.testing.assert_allclose(
+        result.validation_predictions[probability_columns].to_numpy(),
+        rescored[probability_columns].to_numpy(),
+        atol=1e-5,
+    )
+
+
+def test_class_weights_use_training_labels_only(monkeypatch) -> None:
+    import biolmnet.training as training
+
+    captured = {}
+    original = training.nn.CrossEntropyLoss
+
+    def spy(*args, weight=None, **kwargs):
+        captured["weight"] = None if weight is None else weight.cpu().numpy()
+        return original(*args, weight=weight, **kwargs)
+
+    monkeypatch.setattr(training.nn, "CrossEntropyLoss", spy)
+    workspace = _tiny_workspace()  # 24 control / 24 case: all-sample weights are [1, 1].
+    # Force a lopsided split so training-only weights clearly differ from all-sample ones.
+    train_index = np.concatenate([np.arange(0, 22), np.arange(24, 30)])
+    validation_index = np.setdiff1d(np.arange(48), train_index)
+    monkeypatch.setattr(
+        training, "split_train_validation", lambda *args: (train_index, validation_index)
+    )
+    train(workspace, _FAST)
+
+    counts = np.bincount(workspace.labels[train_index], minlength=2).astype(np.float32)
+    assert counts.tolist() == [22.0, 6.0]
+    np.testing.assert_allclose(captured["weight"], len(train_index) / (2 * counts), rtol=1e-6)
+
+
+def test_results_page_receives_training_history_plot() -> None:
+    from app import refresh_results
+
+    assert refresh_results(None, {})[-1] is None
+
+    result = train(_tiny_workspace(), _FAST)
+    figure = refresh_results(result.bundle, {})[-1]
+    traces = {trace.name: trace for trace in figure.data}
+    assert set(traces) == {"Training", "Validation"}
+    history = result.bundle.history
+    assert list(traces["Training"].y) == [row["training_loss"] for row in history]
+    assert list(traces["Validation"].y) == [row["validation_loss"] for row in history]
+
+    # Artifacts saved without history still render the rest of the page.
+    result.bundle.history = []
+    assert refresh_results(result.bundle, {})[-1] is None
+
+
+def test_activation_choices_reach_model_config_and_artifact(tmp_path) -> None:
+    hyperparameters = replace(
+        _FAST,
+        biological_activation="tanh",
+        projection_activation="relu",
+        fusion_activation="sigmoid",
+    )
+    result = train(_tiny_workspace(), hyperparameters)
+    branch = result.bundle.model.gene_branch
+    assert isinstance(branch.biological_activation, torch.nn.Tanh)
+    assert isinstance(branch.projection_activation, torch.nn.ReLU)
+    assert isinstance(result.bundle.model.fusion_activation, torch.nn.Sigmoid)
+    assert result.bundle.config["architecture"]["biological_activation"] == "tanh"
+
+    restored = load_bundle(save_bundle(result.bundle, tmp_path / "model.zip"))
+    assert isinstance(restored.model.gene_branch.biological_activation, torch.nn.Tanh)
+    assert isinstance(restored.model.fusion_activation, torch.nn.Sigmoid)
+
+
+def test_defaults_follow_lab_grid_configuration() -> None:
+    defaults = Hyperparameters()
+    assert (defaults.optimizer, defaults.learning_rate, defaults.weight_decay) == ("Adam", 0.01, 0.01)
+    assert (defaults.batch_size, defaults.projection_dim, defaults.dropout, defaults.epochs) == (32, 64, 0.3, 200)
+    assert (defaults.biological_activation, defaults.projection_activation, defaults.fusion_activation) == (
+        "tanh",
+        "relu",
+        "tanh",
+    )
+    assert defaults.sgd_momentum == 0.03
+
+
+def test_auto_genept_uses_lab_recommended_single_cell_file() -> None:
+    from app import GENEPT_OPTIONS, _resolve_embedding_file
+
+    auto = next(label for label, value in GENEPT_OPTIONS.items() if value == "auto")
+    for dataset in ("BRCA", "COAD", "GBM", "scTrioseq2"):
+        assert _resolve_embedding_file(auto, "BioLM-NET examples", dataset) == (
+            "embedding_associations_cell_type_tissue_drug_pathway_openai_large.parquet"
+        )
+
+
+def test_sgd_momentum_reaches_optimizer(monkeypatch) -> None:
+    import biolmnet.training as training
+
+    captured = {}
+    original = training.torch.optim.SGD
+
+    def spy(params, **kwargs):
+        captured.update(kwargs)
+        return original(params, **kwargs)
+
+    monkeypatch.setattr(training.torch.optim, "SGD", spy)
+    train(_tiny_workspace(), replace(_FAST, optimizer="SGD", learning_rate=0.01, sgd_momentum=0.03))
+    assert captured["momentum"] == 0.03
+
+
+@pytest.mark.parametrize(
+    "field, value",
+    [("biological_activation", "softmax"), ("fusion_activation", "gelu"), ("sgd_momentum", 1.0)],
+)
+def test_invalid_activation_or_momentum_is_rejected(field, value) -> None:
+    with pytest.raises(ValueError):
+        replace(Hyperparameters(), **{field: value}).validate()