Archive 4 checkpoints (1-4)

.gitattributes

@@ -63,3 +63,5 @@ paper-releases/pbisc-diffusion/updates/20260930T065633Z-summary-deck-scsemi-fina
 paper-releases/pbisc-diffusion/updates/20260930T065633Z-summary-deck-scsemi-final/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20260930T065633Z-summary-deck-scsemi-final/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20260930T065633Z-summary-deck-scsemi-final/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/sec01_slide_a.pdf filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx filter=lfs diff=lfs merge=lfs -text

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea779ae9af4771ce0a5b1d96d59ecf06c04cb59892eb8c404af821317c20850d
+size 4547408

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47a36cef0470ced87a47e41b57674d8b2d02f8fe7761f7759246054794d7a817
+size 5100576

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/README.md

@@ -0,0 +1,59 @@
+# PBISC-Diffusion paper summary deck (2026-09-30, r00)
+
+28-slide English summary of the PBISC-Diffusion (Model 8) paper experiments: title,
+outline, 10 sections (23 content slides), then a Backup divider and 3 backup slides.
+Short keyword lines only; the figures carry the results. Section 10 is planned work.
+
+- `PBISC-Diffusion_paper_summary_20260930.pptx`: the deck
+- `PBISC-Diffusion_paper_summary_20260930.pdf`: LibreOffice PDF of the deck
+- `data/slides_spec.json`: slide order, titles, keywords, footnotes, speaker-note text
+- `generate_deck.py`: builds the pptx from `data/slides_spec.json` only (python-pptx)
+- `figures/`: all slide figures (PNG + PDF), made by `tools/sec*.py`
+- `data/sec*.csv`, `data/sec09_slide_text.json`: every plotted or quoted number with its source path
+- `render/`: PDF and 70-dpi slide JPGs for checking
+
+## Rebuild
+
+Everything is CPU only; no GPU or checkpoint is touched.
+
+```bash
+cd /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00
+# figures + data tables (sec03_common.py and style.py are shared helpers, not run directly)
+for f in sec01_overview sec02_backbone sec03_width_loss sec03_width_bench sec03_panel \
+         sec04_sle_benchmark sec05_condition_space sec05_cells_offmanifold \
+         sec05_behcet_direction sec05_shift_cohorts sec06_finetune_method sec07_ft_modes \
+         sec07_k_aug sec08_pairs_needed sec09_paired_validation sec10_paired_subset_cohorts; do
+  python3 "tools/$f.py" || break
+done
+python3 tools/sync_sec09_text.py      # copy regenerated section 9 text into data/slides_spec.json
+python3 generate_deck.py              # -> PBISC-Diffusion_paper_summary_20260930.pptx
+soffice --headless --convert-to pdf --outdir render PBISC-Diffusion_paper_summary_20260930.pptx
+pdftoppm -jpeg -r 70 render/PBISC-Diffusion_paper_summary_20260930.pdf render/slide
+cp render/PBISC-Diffusion_paper_summary_20260930.pdf .
+```
+
+`generate_deck.py` stops with an error if a figure is missing, a title does not fit on one
+line, a footnote needs more than 2 lines, or any text contains Korean or placeholder words.
+
+## Notes
+
+- Section 9 (few-pair scSemiProfiler comparison) uses all 25 scSemiProfiler draws
+  (5 cohorts x 5 draws). If scores change, rerun `tools/sec09_paired_validation.py`, then
+  `tools/sync_sec09_text.py`, then check the short footnotes it lists and rebuild.
+- On-slide footnotes are one short line; the full source and method text is in the speaker
+  notes (`footnote_full` in the spec).
+- `DRAFT_1150_*`: frozen copy shown at the 2026-09-30 12:00 meeting (23/25 scSemiProfiler
+  draws, long footnotes). Superseded by the files above.
+- Story rule (user decision 2026-09-30): the deck and paper do not mention PBISC-VAE, reference
+  calibration (refcal), the TPM input variant or any calibration arm. The no-fine-tune baseline is
+  zero-shot on real bulk as measured; the sc-only control is the sc-only fine-tune fed raw bulk.
+  `tools/sec04_immicro.py` (refcal vs scSemiProfiler on ImMicro) is kept but no longer in the deck.
+  The pre-revision deck, scripts and figures are in `_archive/2026-09-30_before_no_vae_refcal/`.
+
+## Sources
+
+Final model: `runs/f01_6k_1x_seed20260627` (epoch 16, EMA). Scores and receipts come from
+`plans/utility_studies_final6k_20260925/`, `plans/paired_finetune_intradataset_20260927/`,
+`plans/immicro_scsemi_comparison_20260927/`, `plans/floor01_campaign_20260923/`,
+`analysis/scsemi_benchmark_diffusion_20260913/` and the run folders named in each
+`data/sec*.csv`. Manuscript: `/home/sj_server_1/PBISC/paper-release-20260928/manuscript/`.

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec03_panel.py

@@ -0,0 +1,197 @@
+#!/usr/bin/env python3
+"""Section 3c: gene-context (panel) ablation on the 0.1% detection-floor panels, 1x architecture.
+
+Benchmark: part 16 of each panel, Bernoulli (drawn) read-out, scANVI metrics on the 1,006 benchmark genes
+present in every panel (the 2,980-gene benchmark panel intersected with the 2k tier;
+evaluation_manifest_common.json), with scSemiProfiler on the same axis.
+Scores come from the utility-study re-run of the same 1,006-gene pipeline (metrics_1006/summary_tidy.csv),
+which adds 4 extra sampling seeds of the final 6k checkpoint and reproduces every floor01 row exactly
+(asserted here against analysis/floor01_benchmark_20260923/metrics/summary_tidy.csv).
+6k is drawn as the 5-seed mean +- SD with the seeds as dots; the base seed used by the floor01 ranking is
+the best of the five. 2k, 3k and 15k have one seed each.
+Loss: validation loss divided by the panel's own part-16 value. Absolute losses are not compared
+across panels because each panel predicts a different gene set.
+
+Writes figures/sec03_panel.{png,pdf}, data/sec03_panel_bench.csv, data/sec03_panel_seeds.csv,
+data/sec03_panel_loss.csv, data/sec03_panel_headline.csv.
+"""
+from __future__ import annotations
+
+import json
+import re
+import sys
+
+sys.path.insert(0, "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+                   "2026-09-30--paper-summary--r00/tools")
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+from matplotlib.lines import Line2D
+from matplotlib.patches import Patch
+
+import style
+from sec03_common import (DATA, FIG, M8, PANEL_RUNS, PANEL_SCORES, PANEL_SCORES_MD, PANEL_SEED_MANIFEST,
+                          PANEL_SEED_SCORES, SAT_TOL, fmt_params, n_donors, read_run, saturation_epoch)
+
+COMMON = M8 / "plans/floor01_campaign_20260923/evaluation_manifest_common.json"
+FINAL_CKPT = M8 / "runs/f01_6k_1x_seed20260627/checkpoint_epoch_016.pt"
+METHOD = {"2k": "diffusion_f012k_e16", "3k": "diffusion_f013k_e16", "6k": "diffusion_f016k_e16",
+          "15k": "diffusion_f01p15k_e16"}
+SEED_METHODS = ["diffusion_f016k_e16"] + [f"diffusion_f016ks{i}_e16" for i in range(1, 5)]   # base seed first
+METRICS = [("logfc_r", "mean_logFC_pearson", "logFC r (higher better)"),
+           ("top50", "total_top50_overlap", "Top-50 DEG overlap (higher better)"),
+           ("comp_l1", "composition_l1", "Composition L1 (lower better)")]
+REFS = ("scSemiProfiler",)
+
+
+def load_scores(path) -> pd.DataFrame:
+    t = pd.read_csv(path)
+    t = t[t["celltype_gate"] == "common_celltype_gate"]
+    # each comparison block carries the reference rows once; they are identical across blocks
+    for b in REFS:
+        rb = t[t["method"] == b]
+        assert rb[[c for _, c, _ in METRICS]].nunique().max() == 1, b
+    return t.drop_duplicates("method").set_index("method")
+
+
+def main() -> int:
+    n_common = len(json.loads(COMMON.read_text())["evaluation_gene_ids"])
+    donors = n_donors(PANEL_SCORES_MD)
+    t = load_scores(PANEL_SEED_SCORES)
+    assert set(t["panel_genes"]) == {n_common}, (set(t["panel_genes"]), n_common)
+    # the seed re-run must reproduce the floor01 scoring call row for row
+    f01 = load_scores(PANEL_SCORES)
+    for m in list(METHOD.values()) + list(REFS):
+        for _, c, _ in METRICS:
+            assert abs(float(t.loc[m, c]) - float(f01.loc[m, c])) < 1e-9, (m, c)
+    # all five 6k seeds come from the one final checkpoint
+    ck = json.loads(PANEL_SEED_MANIFEST.read_text())["files"]["checkpoint"]
+    assert len(ck) == 1 and ck[0]["path"] == str(FINAL_CKPT), ck
+    ckpt_sha = ck[0]["sha256"]
+
+    seeds = pd.DataFrame([{"method": m, "seed_rank_in_list": i, "is_base_seed": i == 0,
+                           **{k: float(t.loc[m, c]) for k, c, _ in METRICS}} for i, m in enumerate(SEED_METHODS)])
+    seeds["checkpoint"], seeds["checkpoint_sha256"] = str(FINAL_CKPT), ckpt_sha
+    seeds["source"] = str(PANEL_SEED_SCORES)
+    seeds.to_csv(DATA / "sec03_panel_seeds.csv", index=False)
+    s_mean = {k: seeds[k].mean() for k, _, _ in METRICS}
+    s_sd = {k: seeds[k].std(ddof=1) for k, _, _ in METRICS}
+
+    runs, loss_rows, bench_rows = {}, [], []
+    for key, name, host in PANEL_RUNS:
+        r = read_run(name, host)
+        runs[key] = r
+        h = r["hist"]
+        assert int(h["epoch"].max()) == 16
+        v16 = float(h["val_loss"].iloc[-1])
+        sat = saturation_epoch(h["val_loss"].tolist())
+        r["sat"], r["v16"] = sat, v16
+        for _, e in h.iterrows():
+            loss_rows.append({"panel": key, "run": name, "source": r["source"] + "/history.json",
+                              "n_genes": r["cfg"]["data"]["n_genes"], "params": r["status"]["parameters"],
+                              "epoch": int(e["epoch"]), "train_loss": e["train_loss"], "val_loss": e["val_loss"],
+                              "val_rel_e16": e["val_loss"] / v16, "sat_epoch": sat,
+                              "sat_rule": f"val within {SAT_TOL:.0%} of part-16 value"})
+        row = {"arm": key, "method": METHOD[key], "n_genes": r["cfg"]["data"]["n_genes"],
+               "params": r["status"]["parameters"], "val_loss_e16": v16}
+        if key == "6k":
+            row.update({"n_seeds": len(seeds), **{k: s_mean[k] for k in s_mean},
+                        **{f"{k}_sd": s_sd[k] for k in s_sd},
+                        **{f"{k}_base_seed": float(t.loc[METHOD[key], c]) for k, c, _ in METRICS}})
+        else:
+            row.update({"n_seeds": 1, **{k: float(t.loc[METHOD[key], c]) for k, c, _ in METRICS}})
+        bench_rows.append(row)
+    for b in REFS:
+        bench_rows.append({"arm": b, "method": b, **{k: float(t.loc[b, c]) for k, c, _ in METRICS}})
+    bench = pd.DataFrame(bench_rows)
+    bench["eval_genes"], bench["n_donors"], bench["part"] = n_common, donors, 16
+    bench["readout"] = "Bernoulli (drawn)"
+    bench["source"] = str(PANEL_SEED_SCORES)
+    bench.to_csv(DATA / "sec03_panel_bench.csv", index=False)
+    loss = pd.DataFrame(loss_rows)
+    assert loss["params"].notna().all()
+    loss.to_csv(DATA / "sec03_panel_loss.csv", index=False)
+    B = bench.set_index("arm")
+
+    # headline checks for the keyword lines
+    lo = {k: seeds[k].min() for k, _, _ in METRICS}
+    hi = {k: seeds[k].max() for k, _, _ in METRICS}
+    in_range_3k = {k: bool(lo[k] - 1e-12 <= B.loc["3k", k] <= hi[k] + 1e-12) for k, _, _ in METRICS}
+    base_rank = {k: int((seeds[k] > seeds.loc[0, k]).sum()) + 1 if k != "comp_l1"
+                 else int((seeds[k] < seeds.loc[0, k]).sum()) + 1 for k, _, _ in METRICS}
+    m = re.search(r"\|\s*CLEC9A\s*\|\s*([^|]+?)\s*\|\s*([\d,]+)\s*\|\s*(\w+)\s*\|", PANEL_SCORES_MD.read_text())
+    assert m, "CLEC9A row missing"
+    head = pd.DataFrame([
+        {"quantity": "6k_seed_mean_r_top50_l1", "value": "/".join(f"{s_mean[k]:.3f}" for k in s_mean)},
+        {"quantity": "6k_seed_sd_r_top50_l1", "value": "/".join(f"{s_sd[k]:.3f}" for k in s_sd)},
+        {"quantity": "6k_seed_min_max_r", "value": f"{lo['logfc_r']:.3f}-{hi['logfc_r']:.3f}"},
+        {"quantity": "6k_seed_min_max_top50", "value": f"{lo['top50']:.0f}-{hi['top50']:.0f}"},
+        {"quantity": "6k_base_seed_rank_of_5_r_top50_l1", "value": "/".join(str(v) for v in base_rank.values())},
+        {"quantity": "3k_inside_6k_seed_range_r_top50_l1", "value": "/".join(str(v) for v in in_range_3k.values())},
+        {"quantity": "6k_mean_minus_3k_r", "value": s_mean["logfc_r"] - B.loc["3k", "logfc_r"]},
+        {"quantity": "6k_mean_minus_3k_top50", "value": s_mean["top50"] - B.loc["3k", "top50"]},
+        {"quantity": "best_comp_l1_panel", "value": B.loc[["2k", "3k", "6k", "15k"], "comp_l1"].idxmin()},
+        {"quantity": "worst_comp_l1_panel", "value": B.loc[["2k", "3k", "6k", "15k"], "comp_l1"].idxmax()},
+        {"quantity": "CLEC9A_population_rank_smallest_panel", "value": f"{m.group(1)}/{m.group(2)}/{m.group(3)}"},
+        {"quantity": "final_checkpoint_sha256", "value": ckpt_sha},
+    ])
+    head["source"] = [str(PANEL_SEED_SCORES)] * 10 + [str(PANEL_SCORES_MD), str(PANEL_SEED_MANIFEST)]
+    head.to_csv(DATA / "sec03_panel_headline.csv", index=False)
+
+    # --- one full-width figure: three benchmark panels + relative validation loss ----------------
+    keys = [k for k, _, _ in PANEL_RUNS]
+    i6 = keys.index("6k")
+    fig, axes = plt.subplots(1, 4, figsize=style.FULL, gridspec_kw={"width_ratios": [1, 1, 1, 1.55]})
+    rng = np.random.default_rng(0)
+    for ax, (k, _, ylab) in zip(axes[:3], METRICS):
+        for i, key in enumerate(keys):
+            ax.bar(i, B.loc[key, k], width=0.72, color=style.PANEL[key], zorder=2)
+        # SD whisker left of centre, the five seeds right of centre, so neither hides the other
+        ax.errorbar(i6 - 0.14, s_mean[k], yerr=s_sd[k], color=style.INK, lw=1.4, capsize=4, capthick=1.4, zorder=4)
+        ax.scatter(i6 + 0.16 + rng.uniform(-0.06, 0.06, len(seeds)), seeds[k], s=22, color=style.INK, zorder=5,
+                   edgecolors="white", linewidths=0.6)
+        ax.axhline(B.loc["scSemiProfiler", k], color=style.ARM["scsemi"], lw=1.8, ls="--", zorder=3)
+        ax.set_xticks(range(len(keys)))
+        ax.set_xticklabels(keys)
+        ax.get_xticklabels()[i6].set_fontweight("bold")
+        ax.set_xlim(-0.6, len(keys) - 0.4)
+        ax.set_xlabel("Gene panel")
+        ax.set_ylabel(ylab)
+        ax.set_ylim(0, max(B[k].max(), seeds[k].max()) * 1.12)
+
+    ax = axes[3]
+    for key in keys:
+        r = runs[key]
+        h = r["hist"]
+        rel = h["val_loss"] / r["v16"]
+        ax.plot(h["epoch"], rel, "-o", color=style.PANEL[key], ms=3.5, lw=1.8,
+                label=f"{key}, {fmt_params(int(r['status']['parameters']))}, sat. E{r['sat']}")
+        ax.plot([r["sat"]], [rel.iloc[r["sat"] - 1]], marker="o", ms=11, mfc="none", mec=style.INK, mew=1.3)
+    ax.axhline(1.0 + SAT_TOL, color=style.GREY, lw=0.9, ls=":")
+    ax.text(16.3, 1.0 + SAT_TOL + 0.002, f"+{SAT_TOL:.0%}", va="bottom", ha="right", fontsize=12,
+            color=style.GREY)
+    ax.set_xlabel("Schedule part (epoch)")
+    ax.set_ylabel("Val. loss / own part-16 value")
+    ax.set_xticks([1, 4, 8, 12, 16])
+    ax.set_xlim(0.5, 16.5)
+    ax.legend(loc="upper right", handlelength=1.4, borderaxespad=0.1)
+
+    handles = [Patch(facecolor=style.GREY, label="bars 2k, 3k, 15k: 1 seed"),
+               Line2D([], [], color=style.INK, lw=1.4, marker="o", ms=4, mec="white",
+                      label=f"6k: {len(seeds)} seeds, mean ± SD"),
+               Line2D([], [], color=style.ARM["scsemi"], lw=1.8, ls="--", label="scSemiProfiler")]
+    fig.legend(handles=handles, loc="upper left", ncol=len(handles), bbox_to_anchor=(0.06, 1.07),
+               handlelength=2.0, columnspacing=2.4)
+    fig.tight_layout(w_pad=1.6)
+    style.save(fig, "sec03_panel", FIG)
+
+    print(f"common genes {n_common}, donors {donors}, checkpoint sha {ckpt_sha[:8]}")
+    print(bench.drop(columns=["source"]).round(4).to_string())
+    print(seeds[["method", "logfc_r", "top50", "comp_l1"]].round(4).to_string())
+    print(head[["quantity", "value"]].to_string())
+    print({k: runs[k]["sat"] for k in runs})
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())