Upload motif_classifier/sota_compare_100/run.log with huggingface_hub

motif_classifier/sota_compare_100/run.log

@@ -0,0 +1,103 @@
+[load] 35000 rows from /workspace/dnathinker/runs/_motif_targets/from_tools/extracted.jsonl
+[select] top-100 motifs by between-cell variance (vocab size = 756)
+  top-3 by variance: ZNF384(var=0.3), RREB1(var=0.3), CTCF(var=0.2)
+[data] train=15000 val=4000
+
+=== Variant B-deepstarr: shared deepstarr backbone + 100 per-motif heads ===
+  params: 4.10M
+  [B-deepstarr] ep0 mse=0.2468 val_R²=0.2252 r2>0.85: 0/100 dt=4.1s
+  [B-deepstarr] ep1 mse=0.1918 val_R²=0.2712 r2>0.85: 0/100 dt=0.9s
+  [B-deepstarr] ep2 mse=0.1765 val_R²=0.3013 r2>0.85: 0/100 dt=0.8s
+  [B-deepstarr] ep3 mse=0.1679 val_R²=0.2941 r2>0.85: 0/100 dt=0.9s
+  [B-deepstarr] ep4 mse=0.1591 val_R²=0.3161 r2>0.85: 0/100 dt=1.0s
+  [B-deepstarr] ep5 mse=0.1496 val_R²=0.3261 r2>0.85: 0/100 dt=0.9s
+  [B-deepstarr] ep6 mse=0.1414 val_R²=0.3377 r2>0.85: 0/100 dt=1.0s
+  [B-deepstarr] ep7 mse=0.1348 val_R²=0.3271 r2>0.85: 0/100 dt=0.8s
+  [B-deepstarr] ep8 mse=0.1288 val_R²=0.3429 r2>0.85: 0/100 dt=1.0s
+  [B-deepstarr] ep9 mse=0.1240 val_R²=0.3426 r2>0.85: 0/100 dt=0.9s
+  [B-deepstarr] ep10 mse=0.1208 val_R²=0.3428 r2>0.85: 0/100 dt=0.9s
+  [B-deepstarr] ep11 mse=0.1199 val_R²=0.3427 r2>0.85: 0/100 dt=1.0s
+
+=== Variant B-resnet: shared resnet backbone + 100 per-motif heads ===
+  params: 3.55M
+  [B-resnet] ep0 mse=0.2339 val_R²=0.2319 r2>0.85: 0/100 dt=4.1s
+  [B-resnet] ep1 mse=0.1883 val_R²=0.2826 r2>0.85: 0/100 dt=1.4s
+  [B-resnet] ep2 mse=0.1758 val_R²=0.2973 r2>0.85: 0/100 dt=1.5s
+  [B-resnet] ep3 mse=0.1641 val_R²=0.3050 r2>0.85: 0/100 dt=1.4s
+  [B-resnet] ep4 mse=0.1541 val_R²=0.3303 r2>0.85: 0/100 dt=1.4s
+  [B-resnet] ep5 mse=0.1452 val_R²=0.3393 r2>0.85: 0/100 dt=1.4s
+  [B-resnet] ep6 mse=0.1372 val_R²=0.3401 r2>0.85: 0/100 dt=1.5s
+  [B-resnet] ep7 mse=0.1294 val_R²=0.3461 r2>0.85: 0/100 dt=1.5s
+  [B-resnet] ep8 mse=0.1228 val_R²=0.3535 r2>0.85: 0/100 dt=1.5s
+  [B-resnet] ep9 mse=0.1173 val_R²=0.3532 r2>0.85: 0/100 dt=1.5s
+  [B-resnet] ep10 mse=0.1145 val_R²=0.3543 r2>0.85: 0/100 dt=1.4s
+  [B-resnet] ep11 mse=0.1131 val_R²=0.3548 r2>0.85: 0/100 dt=1.4s
+
+=== Variant B-basset: shared basset backbone + 100 per-motif heads ===
+  params: 4.71M
+  [B-basset] ep0 mse=0.2509 val_R²=0.2204 r2>0.85: 0/100 dt=2.9s
+  [B-basset] ep1 mse=0.1986 val_R²=0.2600 r2>0.85: 0/100 dt=0.9s
+  [B-basset] ep2 mse=0.1835 val_R²=0.2471 r2>0.85: 0/100 dt=1.2s
+  [B-basset] ep3 mse=0.1739 val_R²=0.2499 r2>0.85: 0/100 dt=0.9s
+  [B-basset] ep4 mse=0.1672 val_R²=0.2442 r2>0.85: 0/100 dt=1.2s
+  [B-basset] ep5 mse=0.1588 val_R²=0.2805 r2>0.85: 0/100 dt=1.0s
+  [B-basset] ep6 mse=0.1507 val_R²=0.2779 r2>0.85: 0/100 dt=1.0s
+  [B-basset] ep7 mse=0.1443 val_R²=0.2916 r2>0.85: 0/100 dt=1.0s
+  [B-basset] ep8 mse=0.1389 val_R²=0.2909 r2>0.85: 0/100 dt=1.0s
+  [B-basset] ep9 mse=0.1354 val_R²=0.2944 r2>0.85: 0/100 dt=1.1s
+  [B-basset] ep10 mse=0.1325 val_R²=0.2979 r2>0.85: 0/100 dt=0.9s
+  [B-basset] ep11 mse=0.1316 val_R²=0.2985 r2>0.85: 0/100 dt=1.0s
+
+=== Variant A-medium: 100 separate medium CNNs ===
+  params: 4.19M (≈ 41.9K each)
+  [A-medium] ep0 mse=0.2633 val_R²=0.2114 r2>0.85: 0/100 dt=21.0s
+  [A-medium] ep1 mse=0.1962 val_R²=0.2942 r2>0.85: 0/100 dt=19.2s
+  [A-medium] ep2 mse=0.1663 val_R²=0.3829 r2>0.85: 1/100 dt=19.5s
+  [A-medium] ep3 mse=0.1335 val_R²=0.4846 r2>0.85: 3/100 dt=19.4s
+  [A-medium] ep4 mse=0.1051 val_R²=0.5358 r2>0.85: 4/100 dt=19.4s
+  [A-medium] ep5 mse=0.0860 val_R²=0.5658 r2>0.85: 7/100 dt=19.4s
+  [A-medium] ep6 mse=0.0721 val_R²=0.5860 r2>0.85: 7/100 dt=19.2s
+  [A-medium] ep7 mse=0.0620 val_R²=0.5961 r2>0.85: 9/100 dt=19.6s
+  [A-medium] ep8 mse=0.0549 val_R²=0.6012 r2>0.85: 10/100 dt=19.4s
+  [A-medium] ep9 mse=0.0504 val_R²=0.6023 r2>0.85: 10/100 dt=19.4s
+  [A-medium] ep10 mse=0.0479 val_R²=0.6026 r2>0.85: 10/100 dt=19.6s
+  [A-medium] ep11 mse=0.0470 val_R²=0.6027 r2>0.85: 10/100 dt=19.3s
+
+=== Variant A-resnet: 100 separate resnet CNNs ===
+  params: 3.28M (≈ 32.8K each)
+  [A-resnet] ep0 mse=0.2776 val_R²=0.3291 r2>0.85: 1/100 dt=57.2s
+  [A-resnet] ep1 mse=0.1221 val_R²=0.4019 r2>0.85: 1/100 dt=55.5s
+  [A-resnet] ep2 mse=0.0871 val_R²=0.4402 r2>0.85: 1/100 dt=56.3s
+  [A-resnet] ep3 mse=0.0677 val_R²=0.4623 r2>0.85: 3/100 dt=55.5s
+  [A-resnet] ep4 mse=0.0551 val_R²=0.4726 r2>0.85: 4/100 dt=55.8s
+  [A-resnet] ep5 mse=0.0463 val_R²=0.4833 r2>0.85: 4/100 dt=55.8s
+  [A-resnet] ep6 mse=0.0400 val_R²=0.4798 r2>0.85: 4/100 dt=55.0s
+  [A-resnet] ep7 mse=0.0348 val_R²=0.4916 r2>0.85: 4/100 dt=55.7s
+  [A-resnet] ep8 mse=0.0307 val_R²=0.4940 r2>0.85: 4/100 dt=55.9s
+  [A-resnet] ep9 mse=0.0278 val_R²=0.4956 r2>0.85: 4/100 dt=55.6s
+  [A-resnet] ep10 mse=0.0266 val_R²=0.4968 r2>0.85: 4/100 dt=55.2s
+  [A-resnet] ep11 mse=0.0255 val_R²=0.4963 r2>0.85: 4/100 dt=55.6s
+
+=== Variant A-basset: 100 separate basset CNNs ===
+  params: 6.48M (≈ 64.8K each)
+  [A-basset] ep0 mse=0.2280 val_R²=0.2362 r2>0.85: 0/100 dt=34.2s
+  [A-basset] ep1 mse=0.0899 val_R²=0.2507 r2>0.85: 0/100 dt=32.3s
+  [A-basset] ep2 mse=0.0458 val_R²=0.2452 r2>0.85: 0/100 dt=31.8s
+  [A-basset] ep3 mse=0.0277 val_R²=0.2699 r2>0.85: 0/100 dt=32.4s
+  [A-basset] ep4 mse=0.0174 val_R²=0.2764 r2>0.85: 0/100 dt=32.2s
+  [A-basset] ep5 mse=0.0112 val_R²=0.2831 r2>0.85: 0/100 dt=32.0s
+  [A-basset] ep6 mse=0.0077 val_R²=0.2885 r2>0.85: 0/100 dt=32.2s
+  [A-basset] ep7 mse=0.0061 val_R²=0.2845 r2>0.85: 0/100 dt=32.1s
+  [A-basset] ep8 mse=0.0051 val_R²=0.2919 r2>0.85: 0/100 dt=32.2s
+  [A-basset] ep9 mse=0.0041 val_R²=0.2921 r2>0.85: 0/100 dt=32.2s
+  [A-basset] ep10 mse=0.0036 val_R²=0.2915 r2>0.85: 0/100 dt=32.3s
+  [A-basset] ep11 mse=0.0035 val_R²=0.2914 r2>0.85: 0/100 dt=32.2s
+
+=== COMPARISON ===
+  B_deepstarr              : R²=0.3429  params=4.10M  wallclock=15s
+  B_resnet                 : R²=0.3548  params=3.55M  wallclock=20s
+  B_basset                 : R²=0.2985  params=4.71M  wallclock=14s
+  A_separate_medium        : R²=0.6027  params=4.19M  wallclock=235s
+  A_separate_resnet        : R²=0.4968  params=3.28M  wallclock=669s
+  A_separate_basset        : R²=0.2921  params=6.48M  wallclock=388s
+  WINNER: A_separate_medium (R² = 0.6027)