Add halt_heads/halthead_k562_meta.json
Browse files
halt_heads/halthead_k562_meta.json
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cell_line": "k562",
|
| 3 |
+
"split": "fewshot",
|
| 4 |
+
"backbone_ckpt": "/workspace/halt/hf_cache_stse_all/fewshot/k562/checkpoints/best.ckpt",
|
| 5 |
+
"backbone": "ST-SE-Replogle llama (frozen, no grad)",
|
| 6 |
+
"n_perturbations": 968,
|
| 7 |
+
"num_rounds": 8,
|
| 8 |
+
"architecture": {
|
| 9 |
+
"head_module": "AdaptiveStateRefine",
|
| 10 |
+
"H": 328,
|
| 11 |
+
"n_layers": 8,
|
| 12 |
+
"exit_layers": [
|
| 13 |
+
1,
|
| 14 |
+
2,
|
| 15 |
+
3,
|
| 16 |
+
4,
|
| 17 |
+
5,
|
| 18 |
+
6,
|
| 19 |
+
7,
|
| 20 |
+
8
|
| 21 |
+
],
|
| 22 |
+
"num_rounds": 8,
|
| 23 |
+
"error_head": true,
|
| 24 |
+
"halt_head": "Sequential(LayerNorm(H), Linear(H,64), SiLU, Linear(64,1)); last-bias init -2.0",
|
| 25 |
+
"error_head_arch": "Sequential(LayerNorm(H), Linear(H,64), SiLU, Linear(64,1)) + softplus",
|
| 26 |
+
"refinement_token": "Parameter(H) init randn*0.02, appended at seq index S",
|
| 27 |
+
"pert_dim": 2024,
|
| 28 |
+
"input_dim": 2058,
|
| 29 |
+
"n_genes": 2000
|
| 30 |
+
},
|
| 31 |
+
"hparams": {
|
| 32 |
+
"epochs": 50,
|
| 33 |
+
"warmup": 15,
|
| 34 |
+
"n_seeds": 4,
|
| 35 |
+
"cell_set_len_S": 64,
|
| 36 |
+
"lr": 0.003,
|
| 37 |
+
"tau": 0.05,
|
| 38 |
+
"alpha": 0.5,
|
| 39 |
+
"beta": 1.0,
|
| 40 |
+
"gamma": 0.1,
|
| 41 |
+
"delta": 0.1,
|
| 42 |
+
"min_cells": 20,
|
| 43 |
+
"optimizer": "AdamW",
|
| 44 |
+
"grad_clip": 1.0,
|
| 45 |
+
"ponder_gated_after_warmup": true,
|
| 46 |
+
"KL": 0
|
| 47 |
+
},
|
| 48 |
+
"seeds": [
|
| 49 |
+
0,
|
| 50 |
+
1,
|
| 51 |
+
2,
|
| 52 |
+
3
|
| 53 |
+
],
|
| 54 |
+
"input_feature_spec": {
|
| 55 |
+
"basal_state": "obsm[X_state] SE embedding (input_dim=2058), control cells sampled S=64/pert",
|
| 56 |
+
"pert": "one-hot over pert_dim=2024 via pert_onehot_map.pt",
|
| 57 |
+
"batch": "gem_group via batch_onehot_map.pkl",
|
| 58 |
+
"target": "magnitude-free: L2-normalized control-relative HVG response direction (n_genes=2000)",
|
| 59 |
+
"Dr": "per-round cosine distance between round-r decoded prediction direction and target direction (computed from frozen backbone)"
|
| 60 |
+
},
|
| 61 |
+
"dr_source": "recomputed on-the-fly from frozen-backbone forward passes (no disk D_r cache; resp_{line}.npz caches response TARGETS only)",
|
| 62 |
+
"device": "cuda",
|
| 63 |
+
"weight_contents": "per-seed head state_dicts under keys seed_0..seed_3 (refinement_token + halt_head.* + error_head.*); frozen backbone NOT included. E[N] signature = mean over seeds.",
|
| 64 |
+
"reproduction": {
|
| 65 |
+
"rho_spearman_refit_vs_saved": 1.0,
|
| 66 |
+
"pearson_refit_vs_saved": 1.0,
|
| 67 |
+
"per_seed_rho_vs_saved": [
|
| 68 |
+
0.8492980401982247,
|
| 69 |
+
0.9007505332762896,
|
| 70 |
+
0.9428977430658868,
|
| 71 |
+
0.917032986383472
|
| 72 |
+
],
|
| 73 |
+
"n_common_perts": 968,
|
| 74 |
+
"refit_EN_mean": 4.581273315890023,
|
| 75 |
+
"refit_EN_std": 1.18076755960427,
|
| 76 |
+
"refit_EN_range": [
|
| 77 |
+
1.3633559942245483,
|
| 78 |
+
6.960995435714722
|
| 79 |
+
],
|
| 80 |
+
"saved_EN_mean": 4.581273315890023,
|
| 81 |
+
"saved_EN_std": 1.18076755960427,
|
| 82 |
+
"refit_seed_reproducibility_spearman": 0.7606562251931633
|
| 83 |
+
},
|
| 84 |
+
"elapsed_sec": 458.7
|
| 85 |
+
}
|