Fixing the issues reported in https://github.com/huggingface/pytorch-pretrained-BERT/issues/556

Reason for issue was that optimzation steps where computed from example size, which is different from actual size of dataloader when an example is chunked into multiple instances.

Solution in this pull request is to compute num_optimization_steps directly from len(data_loader).

examples/run_swag.py

@@ -362,8 +362,20 @@ def main():
     num_train_optimization_steps = None
     if args.do_train:
         train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
-        num_train_optimization_steps = int(
-            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
+        train_features = convert_examples_to_features(
+            train_examples, tokenizer, args.max_seq_length, True)
+        all_input_ids = torch.tensor(select_field(train_features, 'input_ids'), dtype=torch.long)
+        all_input_mask = torch.tensor(select_field(train_features, 'input_mask'), dtype=torch.long)
+        all_segment_ids = torch.tensor(select_field(train_features, 'segment_ids'), dtype=torch.long)
+        all_label = torch.tensor([f.label for f in train_features], dtype=torch.long)
+        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
+        if args.local_rank == -1:
+            train_sampler = RandomSampler(train_data)
+        else:
+            train_sampler = DistributedSampler(train_data)
+        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
+
+        num_train_optimization_steps = len(train_dataloader) / args.gradient_accumulation_steps * args.num_train_epochs
         if args.local_rank != -1:
             num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
 
@@ -422,22 +434,10 @@ def main():
 
     global_step = 0
     if args.do_train:
-        train_features = convert_examples_to_features(
-            train_examples, tokenizer, args.max_seq_length, True)
         logger.info("***** Running training *****")
         logger.info("  Num examples = %d", len(train_examples))
         logger.info("  Batch size = %d", args.train_batch_size)
         logger.info("  Num steps = %d", num_train_optimization_steps)
-        all_input_ids = torch.tensor(select_field(train_features, 'input_ids'), dtype=torch.long)
-        all_input_mask = torch.tensor(select_field(train_features, 'input_mask'), dtype=torch.long)
-        all_segment_ids = torch.tensor(select_field(train_features, 'segment_ids'), dtype=torch.long)
-        all_label = torch.tensor([f.label for f in train_features], dtype=torch.long)
-        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
-        if args.local_rank == -1:
-            train_sampler = RandomSampler(train_data)
-        else:
-            train_sampler = DistributedSampler(train_data)
-        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
 
         model.train()
         for _ in trange(int(args.num_train_epochs), desc="Epoch"):