more explicit notation: num_train_step => num_train_optimization_steps

.gitignore

@@ -119,4 +119,7 @@ dmypy.json
 .vscode
 
 # TF code
-tensorflow_code
+tensorflow_code
+
+# Models
+models

examples/run_classifier.py

@@ -438,11 +438,13 @@ def main():
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
     train_examples = None
-    num_train_steps = None
+    num_train_optimization_steps = None
     if args.do_train:
         train_examples = processor.get_train_examples(args.data_dir)
-        num_train_steps =
-            len(train_examples) // args.train_batch_size // args.gradient_accumulation_steps * args.num_train_epochs
+        num_train_optimization_steps = int(
+            len(train_examples) / args.train_batch_size / 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()
 
     # Prepare model
     model = BertForSequenceClassification.from_pretrained(args.bert_model,
@@ -468,9 +470,6 @@ def main():
         {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
         {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
     if args.fp16:
         try:
             from apex.optimizers import FP16_Optimizer
@@ -491,7 +490,7 @@ def main():
         optimizer = BertAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              warmup=args.warmup_proportion,
-                             t_total=t_total)
+                             t_total=num_train_optimization_steps)
 
     global_step = 0
     nb_tr_steps = 0
@@ -502,7 +501,7 @@ def main():
         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_steps)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
         all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
         all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
         all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
@@ -539,7 +538,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_lm_finetuning.py

@@ -515,13 +515,15 @@ def main():
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
     #train_examples = None
-    num_train_steps = None
+    num_train_optimization_steps = None
     if args.do_train:
         print("Loading Train Dataset", args.train_file)
         train_dataset = BERTDataset(args.train_file, tokenizer, seq_len=args.max_seq_length,
                                     corpus_lines=None, on_memory=args.on_memory)
-        num_train_steps =
-            len(train_dataset) // args.train_batch_size // args.gradient_accumulation_steps * args.num_train_epochs
+        num_train_optimization_steps = int(
+            len(train_dataset) / args.train_batch_size / 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()
 
     # Prepare model
     model = BertForPreTraining.from_pretrained(args.bert_model)
@@ -545,9 +547,6 @@ def main():
         {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
 
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
     if args.fp16:
         try:
             from apex.optimizers import FP16_Optimizer
@@ -568,14 +567,14 @@ def main():
         optimizer = BertAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              warmup=args.warmup_proportion,
-                             t_total=t_total)
+                             t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
         logger.info("***** Running training *****")
         logger.info("  Num examples = %d", len(train_dataset))
         logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_steps)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
 
         if args.local_rank == -1:
             train_sampler = RandomSampler(train_dataset)
@@ -608,7 +607,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_squad.py

@@ -784,12 +784,14 @@ def main():
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
     train_examples = None
-    num_train_steps = None
+    num_train_optimization_steps = None
     if args.do_train:
         train_examples = read_squad_examples(
             input_file=args.train_file, is_training=True)
-        num_train_steps =
-            len(train_examples) // args.train_batch_size // args.gradient_accumulation_steps * args.num_train_epochs
+        num_train_optimization_steps = int(
+            len(train_examples) / args.train_batch_size / 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()
 
     # Prepare model
     model = BertForQuestionAnswering.from_pretrained(args.bert_model,
@@ -821,9 +823,6 @@ def main():
         {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
 
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
     if args.fp16:
         try:
             from apex.optimizers import FP16_Optimizer
@@ -843,7 +842,7 @@ def main():
         optimizer = BertAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              warmup=args.warmup_proportion,
-                             t_total=t_total)
+                             t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -869,7 +868,7 @@ def main():
         logger.info("  Num orig examples = %d", len(train_examples))
         logger.info("  Num split examples = %d", len(train_features))
         logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_steps)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
         all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
         all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
         all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
@@ -903,7 +902,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_squad2.py

@@ -877,12 +877,14 @@ def main():
     tokenizer = BertTokenizer.from_pretrained(args.bert_model)
 
     train_examples = None
-    num_train_steps = None
+    num_train_optimization_steps = None
     if args.do_train:
         train_examples = read_squad_examples(
             input_file=args.train_file, is_training=True)
-        num_train_steps =
-            len(train_examples) // args.train_batch_size // args.gradient_accumulation_steps * args.num_train_epochs
+        num_train_optimization_steps = int(
+            len(train_examples) / args.train_batch_size / 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()
 
     # Prepare model
     model = BertForQuestionAnswering.from_pretrained(args.bert_model,
@@ -914,9 +916,6 @@ def main():
         {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
 
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
     if args.fp16:
         try:
             from apex.optimizers import FP16_Optimizer
@@ -936,7 +935,7 @@ def main():
         optimizer = BertAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              warmup=args.warmup_proportion,
-                             t_total=t_total)
+                             t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -962,7 +961,7 @@ def main():
         logger.info("  Num orig examples = %d", len(train_examples))
         logger.info("  Num split examples = %d", len(train_features))
         logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_steps)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
         all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
         all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
         all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
@@ -997,7 +996,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_swag.py

@@ -349,11 +349,13 @@ def main():
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
     train_examples = None
-    num_train_steps = None
+    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_steps =
-            len(train_examples) // args.train_batch_size // args.gradient_accumulation_steps * args.num_train_epochs
+        num_train_optimization_steps = int(
+            len(train_examples) / args.train_batch_size / 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()
 
     # Prepare model
     model = BertForMultipleChoice.from_pretrained(args.bert_model,
@@ -384,9 +386,6 @@ def main():
         {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
         {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
         ]
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
     if args.fp16:
         try:
             from apex.optimizers import FP16_Optimizer
@@ -406,7 +405,7 @@ def main():
         optimizer = BertAdam(optimizer_grouped_parameters,
                              lr=args.learning_rate,
                              warmup=args.warmup_proportion,
-                             t_total=t_total)
+                             t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -415,7 +414,7 @@ def main():
         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_steps)
+        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)
@@ -455,7 +454,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used that handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()