Create nli_head_alignbanked_conv5d_76pct_output.txt

training_metrics/nli_head_alignbanked_conv5d_76pct_output.txt

@@ -0,0 +1,279 @@
+=================================================================
+NLI HEAD: Compositional Convolution (conv5d)
+=================================================================
+  Device: cuda
+
+=================================================================
+LOADING MODEL
+=================================================================
+Loading weights: 100%
+ 112/112 [00:00<00:00, 3881.72it/s, Materializing param=token_emb.weight]
+  Model: 32,424,960 params (frozen)
+  Bank: present
+
+=================================================================
+LOADING SNLI
+=================================================================
+  Train: 549,367  Val: 9,842
+
+=================================================================
+PRE-ENCODING
+=================================================================
+  Encoding: 100%|██████████| 537/537 [02:21<00:00,  3.80it/s]
+  Encoding: 100%|██████████| 10/10 [00:02<00:00,  4.16it/s]
+  Enriched: 896 (raw=768 + bank=128)
+  Train: 549,000  Val: 9,800
+    entailment: 183,293 (33.4%)
+    neutral: 182,642 (33.3%)
+    contradiction: 183,065 (33.3%)
+
+=================================================================
+COMPOSITIONAL CONV NLI HEAD
+=================================================================
+    Compositions of 5: 16 paths
+      (1, 1, 1, 1, 1)
+      (1, 1, 1, 2)
+      (1, 1, 2, 1)
+      (1, 1, 3)
+      (1, 2, 1, 1)
+      (1, 2, 2)
+      (1, 3, 1)
+      (1, 4)
+      (2, 1, 1, 1)
+      (2, 1, 2)
+      (2, 2, 1)
+      (2, 3)
+      (3, 1, 1)
+      (3, 2)
+      (4, 1)
+      (5,)
+  Head params: 1,808,657
+
+=================================================================
+TRAINING (20 epochs, 4289 batches/epoch)
+=================================================================
+
+  E 1: 29s
+    Task:      loss=0.7502  t_acc=0.6717  v_acc=0.7094  v_loss=0.6855
+    Per-class: ent=0.833  neu=0.687  con=0.605
+    Paths:     (5,)=0.066  (4, 1)=0.065  (2, 1, 2)=0.064  spread=0.0066
+    Protos:    sim=-0.0341  temp=9.96
+    ★ New best: 0.7094
+
+  E 2: 29s
+    Task:      loss=0.6658  t_acc=0.7171  v_acc=0.7282  v_loss=0.6359
+    Per-class: ent=0.773  neu=0.663  con=0.747
+    Paths:     (5,)=0.074  (2, 1, 2)=0.068  (3, 2)=0.067  spread=0.0176
+    Protos:    sim=-0.0761  temp=9.95
+    ★ New best: 0.7282
+
+  E 3: 29s
+    Task:      loss=0.6256  t_acc=0.7370  v_acc=0.7299  v_loss=0.6288
+    Per-class: ent=0.740  neu=0.738  con=0.711
+    Paths:     (5,)=0.083  (2, 1, 2)=0.071  (3, 2)=0.070  spread=0.0314
+    Protos:    sim=-0.1379  temp=9.98
+    ★ New best: 0.7299
+
+  E 4: 29s
+    Task:      loss=0.5921  t_acc=0.7536  v_acc=0.7401  v_loss=0.6132
+    Per-class: ent=0.758  neu=0.676  con=0.785
+    Paths:     (5,)=0.092  (2, 1, 2)=0.073  (1, 4)=0.073  spread=0.0441
+    Protos:    sim=-0.2124  temp=10.04
+    ★ New best: 0.7401
+
+  E 5: 29s
+    Task:      loss=0.5618  t_acc=0.7688  v_acc=0.7457  v_loss=0.6052
+    Per-class: ent=0.797  neu=0.692  con=0.747
+    Paths:     (5,)=0.100  (1, 4)=0.083  (2, 1, 2)=0.074  spread=0.0552
+    Protos:    sim=-0.2886  temp=10.14
+    ★ New best: 0.7457
+
+  E 6: 29s
+    Task:      loss=0.5313  t_acc=0.7834  v_acc=0.7540  v_loss=0.5993
+    Per-class: ent=0.786  neu=0.721  con=0.754
+    Paths:     (5,)=0.106  (1, 4)=0.095  (2, 3)=0.075  spread=0.0640
+    Protos:    sim=-0.3571  temp=10.27
+    ★ New best: 0.7540
+
+  E 7: 29s
+    Task:      loss=0.5010  t_acc=0.7977  v_acc=0.7603  v_loss=0.5865
+    Per-class: ent=0.803  neu=0.721  con=0.755
+    Paths:     (5,)=0.109  (1, 4)=0.109  (2, 3)=0.079  spread=0.0708
+    Protos:    sim=-0.4113  temp=10.44
+    ★ New best: 0.7603
+
+  E 8: 29s
+    Task:      loss=0.4705  t_acc=0.8131  v_acc=0.7563  v_loss=0.5946
+    Per-class: ent=0.792  neu=0.695  con=0.781
+    Paths:     (1, 4)=0.122  (5,)=0.112  (2, 3)=0.082  spread=0.0864
+    Protos:    sim=-0.4490  temp=10.62
+
+  E 9: 29s
+    Task:      loss=0.4413  t_acc=0.8273  v_acc=0.7600  v_loss=0.5955
+    Per-class: ent=0.795  neu=0.719  con=0.765
+    Paths:     (1, 4)=0.135  (5,)=0.113  (2, 3)=0.085  spread=0.1014
+    Protos:    sim=-0.4716  temp=10.80
+
+  E10: 29s
+    Task:      loss=0.4135  t_acc=0.8419  v_acc=0.7609  v_loss=0.5967
+    Per-class: ent=0.780  neu=0.718  con=0.784
+    Paths:     (1, 4)=0.146  (5,)=0.113  (2, 3)=0.087  spread=0.1141
+    Protos:    sim=-0.4840  temp=10.98
+    ★ New best: 0.7609
+
+  E11: 29s
+    Task:      loss=0.3878  t_acc=0.8552  v_acc=0.7602  v_loss=0.5929
+    Per-class: ent=0.791  neu=0.730  con=0.759
+    Paths:     (1, 4)=0.155  (5,)=0.113  (2, 3)=0.089  spread=0.1241
+    Protos:    sim=-0.4904  temp=11.14
+
+  E12: 29s
+    Task:      loss=0.3643  t_acc=0.8680  v_acc=0.7634  v_loss=0.5950
+    Per-class: ent=0.795  neu=0.736  con=0.758
+    Paths:     (1, 4)=0.161  (5,)=0.113  (2, 3)=0.091  spread=0.1317
+    Protos:    sim=-0.4938  temp=11.28
+    ★ New best: 0.7634
+
+  E13: 29s
+    Task:      loss=0.3438  t_acc=0.8795  v_acc=0.7597  v_loss=0.6002
+    Per-class: ent=0.804  neu=0.722  con=0.752
+    Paths:     (1, 4)=0.166  (5,)=0.113  (2, 3)=0.092  spread=0.1373
+    Protos:    sim=-0.4958  temp=11.39
+
+  E14: 29s
+    Task:      loss=0.3263  t_acc=0.8899  v_acc=0.7604  v_loss=0.6013
+    Per-class: ent=0.797  neu=0.718  con=0.765
+    Paths:     (1, 4)=0.169  (5,)=0.113  (2, 3)=0.093  spread=0.1412
+    Protos:    sim=-0.4969  temp=11.48
+
+  E15: 29s
+    Task:      loss=0.3110  t_acc=0.8987  v_acc=0.7587  v_loss=0.6004
+    Per-class: ent=0.797  neu=0.720  con=0.758
+    Paths:     (1, 4)=0.171  (5,)=0.113  (2, 3)=0.093  spread=0.1439
+    Protos:    sim=-0.4976  temp=11.55
+
+  E16: 29s
+    Task:      loss=0.2988  t_acc=0.9056  v_acc=0.7621  v_loss=0.6004
+    Per-class: ent=0.794  neu=0.732  con=0.760
+    Paths:     (1, 4)=0.173  (5,)=0.113  (2, 3)=0.094  spread=0.1458
+    Protos:    sim=-0.4980  temp=11.60
+
+  E17: 29s
+    Task:      loss=0.2892  t_acc=0.9115  v_acc=0.7615  v_loss=0.6010
+    Per-class: ent=0.792  neu=0.732  con=0.759
+    Paths:     (1, 4)=0.174  (5,)=0.113  (2, 3)=0.094  spread=0.1469
+    Protos:    sim=-0.4982  temp=11.63
+
+  E18: 29s
+    Task:      loss=0.2824  t_acc=0.9159  v_acc=0.7601  v_loss=0.6022
+    Per-class: ent=0.792  neu=0.724  con=0.764
+    Paths:     (1, 4)=0.174  (5,)=0.113  (2, 3)=0.094  spread=0.1475
+    Protos:    sim=-0.4984  temp=11.64
+
+  E19: 29s
+    Task:      loss=0.2778  t_acc=0.9186  v_acc=0.7611  v_loss=0.6022
+    Per-class: ent=0.789  neu=0.732  con=0.762
+    Paths:     (1, 4)=0.174  (5,)=0.113  (2, 3)=0.094  spread=0.1478
+    Protos:    sim=-0.4984  temp=11.65
+
+  E20: 29s
+    Task:      loss=0.2754  t_acc=0.9199  v_acc=0.7606  v_loss=0.6020
+    Per-class: ent=0.790  neu=0.728  con=0.763
+    Paths:     (1, 4)=0.174  (5,)=0.113  (2, 3)=0.094  spread=0.1480
+    Protos:    sim=-0.4985  temp=11.66
+
+=================================================================
+PATH WEIGHT ANALYSIS
+=================================================================
+
+  Path                        Weight Type           
+  --------------------------------------------------
+  (1, 4)                      0.1609 geo→rest        ████████████████
+  (5,)                        0.1133 holistic        ███████████
+  (2, 3)                      0.0905 geo+struct→...  █████████
+  (3, 2)                      0.0717 geo-first       ███████
+  (2, 1, 2)                   0.0677 geo+struct→...  ██████
+  (1, 1, 3)                   0.0669 geo→rest        ██████
+  (1, 1, 1, 2)                0.0552 geo→rest        █████
+  (4, 1)                      0.0549 geo-first       █████
+  (2, 2, 1)                   0.0471 geo+struct→...  ████
+  (3, 1, 1)                   0.0468 geo-first       ████
+  (1, 2, 2)                   0.0434 geo→rest        ████
+  (2, 1, 1, 1)                0.0422 geo+struct→...  ████
+  (1, 3, 1)                   0.0380 geo→rest        ███
+  (1, 1, 2, 1)                0.0370 geo→rest        ███
+  (1, 1, 1, 1, 1)             0.0351 independent     ███
+  (1, 2, 1, 1)                0.0292 geo→rest        ██
+
+=================================================================
+COMPOSITIONAL ORDER TEST
+=================================================================
+A new version of the following files was downloaded from https://huggingface.co/AbstractPhil/geolip-captionbert-8192:
+- modeling_caption_bert.py
+. Make sure to double-check they do not contain any added malicious code. To avoid downloading new versions of the code file, you can pin a revision.
+Loading weights: 100%
+ 112/112 [00:00<00:00, 3890.69it/s, Materializing param=token_emb.weight]
+
+  P: a potato on top of a table
+  H: a table on top of a potato
+  Pooled cos: 0.987
+  NLI: entailment  [E=0.507 N=0.056 C=0.437]
+
+  P: a potato on top of a table
+  H: there is a potato
+  Pooled cos: 0.502
+  NLI: entailment  [E=0.842 N=0.093 C=0.065]
+
+  P: a cat is sitting on a mat
+  H: a mat is sitting on a cat
+  Pooled cos: 0.993
+  NLI: entailment  [E=0.917 N=0.014 C=0.070]
+
+  P: a dog chased the cat
+  H: the cat chased the dog
+  Pooled cos: 0.977
+  NLI: entailment  [E=0.427 N=0.201 C=0.372]
+
+  P: a woman is holding a baby
+  H: a baby is holding a woman
+  Pooled cos: 0.996
+  NLI: entailment  [E=0.991 N=0.005 C=0.004]
+
+  P: the boy kicked the ball
+  H: the ball kicked the boy
+  Pooled cos: 0.986
+  NLI: contradiction  [E=0.462 N=0.014 C=0.524]
+
+  P: a man is riding a horse
+  H: a horse is riding a man
+  Pooled cos: 0.995
+  NLI: entailment  [E=0.862 N=0.004 C=0.134]
+
+  P: a girl is painting a picture
+  H: a girl is creating art
+  Pooled cos: 0.796
+  NLI: neutral  [E=0.313 N=0.596 C=0.092]
+
+  P: two dogs are playing in a park
+  H: animals are outdoors
+  Pooled cos: 0.676
+  NLI: entailment  [E=0.985 N=0.013 C=0.001]
+
+  P: a person is swimming in the ocean
+  H: nobody is in the water
+  Pooled cos: 0.778
+  NLI: contradiction  [E=0.059 N=0.004 C=0.937]
+
+=================================================================
+SUMMARY
+=================================================================
+  Best val accuracy: 0.7634
+  Head params:       1,808,657
+  Paths:             16
+  Components:        5 → d_path=256
+  Bank present:      True
+  Saved: nli_conv5d_best.pt
+
+=================================================================
+DONE
+=================================================================