Add forward hook paths explicitly for recent torchdistill vers.

yoshitomo-matsubara · Aug 30, 2024 · c701fe7 · c701fe7
1 parent 9124113
commit c701fe7
Show file tree

Hide file tree

Showing 16 changed files with 122 additions and 32 deletions.
diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/cola/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/cola/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -108,9 +108,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 7.0
               alpha: 0.9

diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/mnli/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/mnli/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -88,13 +88,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -118,9 +124,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9

diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/mrpc/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/mrpc/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 11.0
               alpha: 0.9

diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/qnli/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/qnli/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 1.0
               alpha: 0.5

diff --git a/...lue/yoshitomo-matsubara/nlp-oss2023/qqp/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...lue/yoshitomo-matsubara/nlp-oss2023/qqp/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9

diff --git a/...lue/yoshitomo-matsubara/nlp-oss2023/rte/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...lue/yoshitomo-matsubara/nlp-oss2023/rte/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9

diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/sst2/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/sst2/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 3.0
               alpha: 0.9

diff --git a/...ue/yoshitomo-matsubara/nlp-oss2023/wnli/kd/bert_base_uncased_from_bert_large_uncased.yaml b/...ue/yoshitomo-matsubara/nlp-oss2023/wnli/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9

diff --git a/configs/sample/glue/cola/kd/bert_base_uncased_from_bert_large_uncased.yaml b/configs/sample/glue/cola/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 7.0
               alpha: 0.9

diff --git a/configs/sample/glue/mnli/kd/bert_base_uncased_from_bert_large_uncased.yaml b/configs/sample/glue/mnli/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -88,13 +88,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -118,9 +124,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9

diff --git a/configs/sample/glue/mrpc/kd/bert_base_uncased_from_bert_large_uncased.yaml b/configs/sample/glue/mrpc/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 11.0
               alpha: 0.9

diff --git a/configs/sample/glue/qnli/kd/bert_base_uncased_from_bert_large_uncased.yaml b/configs/sample/glue/qnli/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 1.0
               alpha: 0.5

diff --git a/configs/sample/glue/qqp/kd/bert_base_uncased_from_bert_large_uncased.yaml b/configs/sample/glue/qqp/kd/bert_base_uncased_from_bert_large_uncased.yaml
@@ -78,13 +78,19 @@ train:
     wrapper:
     requires_grad: False
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   student:
     forward_proc: 'forward_batch_as_kwargs'
     adaptations:
     sequential: []
     wrapper:
     requires_grad: True
     frozen_modules: []
+    forward_hook:
+      input: []
+      output: ['classifier']
   optimizer:
     key: 'optimizer_no_decay'
     kwargs:
@@ -108,9 +114,9 @@ train:
           criterion:
             key: 'KDLoss'
             kwargs:
-              student_module_path: '.classifier'
+              student_module_path: 'classifier'
               student_module_io: 'output'
-              teacher_module_path: '.classifier'
+              teacher_module_path: 'classifier'
               teacher_module_io: 'output'
               temperature: 5.0
               alpha: 0.9