Merge pull request #604 from samuelbroscheit/master

Fixing issue "Training beyond specified 't_total' steps with schedule 'warmup_linear'" reported in #556

examples/run_classifier.py

@@ -25,6 +25,7 @@ import random
 import sys
 
 import numpy as np
+import math
 import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
@@ -735,15 +736,6 @@ def main():
 
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
-    train_examples = None
-    num_train_optimization_steps = None
-    if args.do_train:
-        train_examples = processor.get_train_examples(args.data_dir)
-        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
     cache_dir = args.cache_dir if args.cache_dir else os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank))
     model = BertForSequenceClassification.from_pretrained(args.bert_model,
@@ -762,8 +754,35 @@ def main():
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)
 
-    # Prepare optimizer
     if args.do_train:
+
+        # Prepare data loader
+
+        train_examples = processor.get_train_examples(args.data_dir)
+        train_features = convert_examples_to_features(
+            train_examples, label_list, args.max_seq_length, tokenizer, output_mode)
+        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)
+
+        if output_mode == "classification":
+            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
+        elif output_mode == "regression":
+            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.float)
+
+        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+        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()
+
+        # Prepare optimizer
+
         param_optimizer = list(model.named_parameters())
         no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
         optimizer_grouped_parameters = [
@@ -797,28 +816,11 @@ def main():
         global_step = 0
         nb_tr_steps = 0
         tr_loss = 0
-    if args.do_train:
-        train_features = convert_examples_to_features(
-            train_examples, label_list, args.max_seq_length, tokenizer, output_mode)
+
         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([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)
-
-        if output_mode == "classification":
-            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
-        elif output_mode == "regression":
-            all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.float)
-
-        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
-        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"):

examples/run_openai_gpt.py

@@ -190,7 +190,7 @@ 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}
             ]
-        num_train_optimization_steps = len(train_data) * args.num_train_epochs // args.train_batch_size
+        num_train_optimization_steps = len(train_dataloader) * args.num_train_epochs
         optimizer = OpenAIAdam(optimizer_grouped_parameters,
                                lr=args.learning_rate,
                                warmup=args.warmup_proportion,

examples/run_squad.py

@@ -894,16 +894,6 @@ def main():
 
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
-    train_examples = None
-    num_train_optimization_steps = None
-    if args.do_train:
-        train_examples = read_squad_examples(
-            input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
-        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,
                 cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)))
@@ -921,8 +911,47 @@ def main():
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)
 
-    # Prepare optimizer
     if args.do_train:
+
+        # Prepare data loader
+
+        train_examples = read_squad_examples(
+            input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
+        cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format(
+            list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
+        try:
+            with open(cached_train_features_file, "rb") as reader:
+                train_features = pickle.load(reader)
+        except:
+            train_features = convert_examples_to_features(
+                examples=train_examples,
+                tokenizer=tokenizer,
+                max_seq_length=args.max_seq_length,
+                doc_stride=args.doc_stride,
+                max_query_length=args.max_query_length,
+                is_training=True)
+            if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+                logger.info("  Saving train features into cached file %s", cached_train_features_file)
+                with open(cached_train_features_file, "wb") as writer:
+                    pickle.dump(train_features, writer)
+        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)
+        all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
+        all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
+        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
+                                   all_start_positions, all_end_positions)
+        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()
+
+        # Prepare optimizer
+
         param_optimizer = list(model.named_parameters())
 
         # hack to remove pooler, which is not used
@@ -959,42 +988,12 @@ def main():
                                  t_total=num_train_optimization_steps)
 
         global_step = 0
-    if args.do_train:
-        cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format(
-            list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
-        train_features = None
-        try:
-            with open(cached_train_features_file, "rb") as reader:
-                train_features = pickle.load(reader)
-        except:
-            train_features = convert_examples_to_features(
-                examples=train_examples,
-                tokenizer=tokenizer,
-                max_seq_length=args.max_seq_length,
-                doc_stride=args.doc_stride,
-                max_query_length=args.max_query_length,
-                is_training=True)
-            if args.local_rank == -1 or torch.distributed.get_rank() == 0:
-                logger.info("  Saving train features into cached file %s", cached_train_features_file)
-                with open(cached_train_features_file, "wb") as writer:
-                    pickle.dump(train_features, writer)
+
         logger.info("***** Running training *****")
         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_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)
-        all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
-        all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
-        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
-                                   all_start_positions, all_end_positions)
-        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"):

examples/run_swag.py

@@ -358,15 +358,6 @@ def main():
 
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
-    train_examples = 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_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,
         cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)),
@@ -384,8 +375,30 @@ def main():
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)
 
-    # Prepare optimizer
     if args.do_train:
+
+        # Prepare data loader
+
+        train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
+        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()
+
+        # Prepare optimizer
+
         param_optimizer = list(model.named_parameters())
 
         # hack to remove pooler, which is not used
@@ -421,23 +434,11 @@ def main():
                                  t_total=num_train_optimization_steps)
 
         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"):