Upload sc_transcriptomics_esmc/config.yaml with huggingface_hub

sc_transcriptomics_esmc/config.yaml

@@ -0,0 +1,59 @@
+# Training configuration for the OmicsFM checkpoint in this folder.
+# Extracted verbatim from best_model.ckpt.
+
+checkpoint: sc_transcriptomics_esmc
+training_run: trans_prot_mapped_emsc
+epoch: 42
+val_loss: 11.131809487802968
+vocabulary_size: 20272
+config:
+  data:
+    train_path: data/sc_trans/protein_mapped/train.h5ad
+    valid_path: data/sc_trans/protein_mapped/valid.h5ad
+    test_path: data/sc_trans/protein_mapped/test.h5ad
+    num_bins: 10
+    detect_groups: false
+    cache_build_workers: 1
+  transfer:
+    checkpoint_path: null
+    freeze_transformer: false
+    freeze_epochs: 0
+  device: cuda
+  model:
+    d_model: 256
+    n_heads: 8
+    n_layers: 6
+    d_ff: 1024
+    dropout: 0.1
+    input_features: 1024
+    feature_context_ratio: 0.85
+    use_esmc: true
+    fasta_path: fasta/human_proteome_canonical_31032022.fasta
+    esmc_model: esmc_600m
+    esmc_cache: data/esmc_human_cache.pt
+    absent_species:
+      enabled: false
+      input_absent: 50
+      absent_context_ratio: 0.5
+    protein_groups:
+      enabled: false
+      max_group_size: 20
+      input_groups: 256
+      group_context_ratio: 0.2
+      encoder: mean
+      attention:
+        n_heads: 4
+        n_encoder_layers: 1
+  training:
+    batch_size: 64
+    learning_rate: 0.0001
+    weight_decay: 1.0e-05
+    max_epochs: 200
+    val_frequency: 0.2
+    early_stopping_patience: 5000
+    early_stopping_delta: 0.001
+    gradient_clip: 1.0
+    num_workers: 0
+  output:
+    save_dir: model
+    experiment_name: trans_prot_mapped_emsc