train generate model updates for transformers 4.0.0

Author: bjascob

amrlib/models/generate_t5/trainer.py

@@ -31,6 +31,8 @@ def __getitem__(self, idx):
 # prepares lm_labels from target_ids, returns examples with keys as expected by the forward method
 # this is necessacry because the trainer directly passes this dict as arguments to the model
 # so make sure the keys match the parameter names of the forward method
+# Note*1: The original code (with transformers v3.4.0) returned dict with "lm_labels".
+# Support for this was removed in transformers v4.0.0 and replaced it with "labels"
 class T2TDataCollator:
     def __call__(self, batch):
         input_ids = torch.stack([example['input_ids']  for example in batch])
@@ -39,7 +41,7 @@ def __call__(self, batch):
         attention_mask = torch.stack([example['attention_mask'] for example in batch])
         decoder_attention_mask = torch.stack([example['target_attention_mask'] for example in batch])
         return {'input_ids': input_ids, 'attention_mask': attention_mask,
-                'lm_labels': lm_labels, 'decoder_attention_mask': decoder_attention_mask }
+                'labels': lm_labels, 'decoder_attention_mask': decoder_attention_mask }     # Note*1
 
 
 # Note that for save_steps, steps means gradient updates (not batch) so if
@@ -80,8 +82,11 @@ def train(self):
             len(valid_dataset), len(valid_dataset.bad_indexes)))
         # Train the model
         print('Training')
+        # trainer = T5Trainer(model=self.model, args=self.training_args, train_dataset=train_dataset,
+        #         eval_dataset=valid_dataset, data_collator=T2TDataCollator(), prediction_loss_only=True)
+        # prediction_loss_only=True moved to training_args for compatibility with transformers v4.0.0
         trainer = T5Trainer(model=self.model, args=self.training_args, train_dataset=train_dataset,
-                eval_dataset=valid_dataset, data_collator=T2TDataCollator(), prediction_loss_only=True)
+                eval_dataset=valid_dataset, data_collator=T2TDataCollator())
         trainer.train()
         # Save the results
         print('Saving model')

amrlib/models/parse_gsii/modules/transformer.py

@@ -200,6 +200,10 @@ def in_proj_qkv(self, query):
         # See release notes for v1.7 (torch.chunk) for an explanation.  A temporary fix is to use unsafe_chunk instead.
         # See https://discuss.pytorch.org/t/runtimeerror-for-chunk-inplace-operation-new-with-torch-1-7/105334
         return self._in_proj(query).unsafe_chunk(3, dim=-1)
+        # Possible solution...
+        # proj = self._in_proj(query)
+        # sz   = proj.size()[2] // 3
+        # return proj[:,:,:sz], proj[:,:,sz:2*sz], proj[:,:,2*sz:]        
 
     def in_proj_kv(self, key):
         return self._in_proj(key, start=self.embed_dim).chunk(2, dim=-1)

configs/model_generate_t5.json

@@ -13,8 +13,8 @@
         "output_dir"                    : "amrlib/data/model_generate_t5",
         "do_train"                      : true,
         "do_eval"                       : false,
-        "evaluate_during_training"      : false,
         "overwrite_output_dir"          : false,
+        "prediction_loss_only"          : true,
         "num_train_epochs"              : 8,
         "save_steps"                    : 1000,
         "save_total_limit"              : 2,