Merge remote-tracking branch 'upstream/master' into convert-back-to-tf

merging

examples/lm_finetuning/finetune_on_pregenerated.py

@@ -315,8 +315,7 @@ 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.get_lr(global_step/num_train_optimization_steps,
-                                                                                 args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/lm_finetuning/simple_lm_finetuning.py

@@ -534,36 +534,37 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'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}
-        ]
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'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}
+            ]
 
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
 
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
         else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-        warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
-                                             t_total=num_train_optimization_steps)
-
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=num_train_optimization_steps)
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -603,8 +604,7 @@ 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.get_lr(global_step/num_train_optimization_steps,
-                                                                                 args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_classifier.py

@@ -271,7 +271,7 @@ class StsbProcessor(DataProcessor):
 
 
 class QqpProcessor(DataProcessor):
-    """Processor for the STS-B data set (GLUE version)."""
+    """Processor for the QQP data set (GLUE version)."""
 
     def get_train_examples(self, data_dir):
         """See base class."""
@@ -306,7 +306,7 @@ class QqpProcessor(DataProcessor):
 
 
 class QnliProcessor(DataProcessor):
-    """Processor for the STS-B data set (GLUE version)."""
+    """Processor for the QNLI data set (GLUE version)."""
 
     def get_train_examples(self, data_dir):
         """See base class."""
@@ -763,35 +763,36 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'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}
-        ]
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'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}
+            ]
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
 
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
         else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-        warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
-                                             t_total=num_train_optimization_steps)
-
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=num_train_optimization_steps)
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
 
     global_step = 0
     nb_tr_steps = 0
@@ -854,8 +855,7 @@ 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.get_lr(global_step/num_train_optimization_steps,
-                                                                                 args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_openai_gpt.py

@@ -183,19 +183,20 @@ def main():
     eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'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
-    optimizer = OpenAIAdam(optimizer_grouped_parameters,
-                           lr=args.learning_rate,
-                           warmup=args.warmup_proportion,
-                           max_grad_norm=args.max_grad_norm,
-                           weight_decay=args.weight_decay,
-                           t_total=num_train_optimization_steps)
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'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
+        optimizer = OpenAIAdam(optimizer_grouped_parameters,
+                               lr=args.learning_rate,
+                               warmup=args.warmup_proportion,
+                               max_grad_norm=args.max_grad_norm,
+                               weight_decay=args.weight_decay,
+                               t_total=num_train_optimization_steps)
 
     if args.do_train:
         nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None

examples/run_squad.py

@@ -922,40 +922,41 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
 
-    # hack to remove pooler, which is not used
-    # thus it produce None grad that break apex
-    param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
+        # hack to remove pooler, which is not used
+        # thus it produce None grad that break apex
+        param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
 
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'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}
-        ]
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'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}
+            ]
 
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
 
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
         else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-        warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
-                                             t_total=num_train_optimization_steps)
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=num_train_optimization_steps)
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -1015,8 +1016,7 @@ def main():
                     if args.fp16:
                         # modify learning rate with special warm up BERT uses
                         # if args.fp16 is False, BertAdam is used and handles this automatically
-                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step/num_train_optimization_steps,
-                                                                                 args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

examples/run_swag.py

@@ -385,39 +385,40 @@ def main():
         model = torch.nn.DataParallel(model)
 
     # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
 
-    # hack to remove pooler, which is not used
-    # thus it produce None grad that break apex
-    param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
+        # hack to remove pooler, which is not used
+        # thus it produce None grad that break apex
+        param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
 
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'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}
-        ]
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'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}
+            ]
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
 
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
         else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-        warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
-                                             t_total=num_train_optimization_steps)
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=num_train_optimization_steps)
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
 
     global_step = 0
     if args.do_train:
@@ -466,8 +467,7 @@ 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.get_lr(global_step/num_train_optimization_steps,
-                                                                                 args.warmup_proportion)
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
                         for param_group in optimizer.param_groups:
                             param_group['lr'] = lr_this_step
                     optimizer.step()

hubconf.py

@@ -84,7 +84,7 @@ def bertTokenizer(*args, **kwargs):
 
     Example:
         >>> sentence = 'Hello, World!'
-        >>> tokenizer = torch.hub.load('ailzhang/pytorch-pretrained-BERT:hubconf', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False, force_reload=False)
+        >>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT:hubconf', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False, force_reload=False)
         >>> toks = tokenizer.tokenize(sentence)
         ['Hello', '##,', 'World', '##!']
         >>> ids = tokenizer.convert_tokens_to_ids(toks)

pytorch_pretrained_bert/file_utils.py

@@ -22,6 +22,15 @@ import requests
 from botocore.exceptions import ClientError
 from tqdm import tqdm
 
+try:
+    from torch.hub import _get_torch_home
+    torch_cache_home = _get_torch_home()
+except ImportError:
+    torch_cache_home = os.path.expanduser(
+        os.getenv('TORCH_HOME', os.path.join(
+            os.getenv('XDG_CACHE_HOME', '~/.cache'), 'torch')))
+default_cache_path = os.path.join(torch_cache_home, 'pytorch_pretrained_bert')
+
 try:
     from urllib.parse import urlparse
 except ImportError:
@@ -29,11 +38,11 @@ except ImportError:
 
 try:
     from pathlib import Path
-    PYTORCH_PRETRAINED_BERT_CACHE = Path(os.getenv('PYTORCH_PRETRAINED_BERT_CACHE',
-                                                   Path.home() / '.pytorch_pretrained_bert'))
+    PYTORCH_PRETRAINED_BERT_CACHE = Path(
+        os.getenv('PYTORCH_PRETRAINED_BERT_CACHE', default_cache_path))
 except (AttributeError, ImportError):
     PYTORCH_PRETRAINED_BERT_CACHE = os.getenv('PYTORCH_PRETRAINED_BERT_CACHE',
-                                              os.path.join(os.path.expanduser("~"), '.pytorch_pretrained_bert'))
+                                              default_cache_path)
 
 CONFIG_NAME = "config.json"
 WEIGHTS_NAME = "pytorch_model.bin"

pytorch_pretrained_bert/modeling.py

@@ -145,7 +145,8 @@ class BertConfig(object):
                  attention_probs_dropout_prob=0.1,
                  max_position_embeddings=512,
                  type_vocab_size=2,
-                 initializer_range=0.02):
+                 initializer_range=0.02,
+                 layer_norm_eps=1e-12):
         """Constructs BertConfig.
 
         Args:
@@ -169,6 +170,7 @@ class BertConfig(object):
                 `BertModel`.
             initializer_range: The sttdev of the truncated_normal_initializer for
                 initializing all weight matrices.
+            layer_norm_eps: The epsilon used by LayerNorm.
         """
         if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                         and isinstance(vocab_size_or_config_json_file, unicode)):
@@ -188,6 +190,7 @@ class BertConfig(object):
             self.max_position_embeddings = max_position_embeddings
             self.type_vocab_size = type_vocab_size
             self.initializer_range = initializer_range
+            self.layer_norm_eps = layer_norm_eps
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
                              "or the path to a pretrained model config file (str)")
@@ -254,7 +257,7 @@ class BertEmbeddings(nn.Module):
 
         # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
         # any TensorFlow checkpoint file
-        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, input_ids, token_type_ids=None):
@@ -329,7 +332,7 @@ class BertSelfOutput(nn.Module):
     def __init__(self, config):
         super(BertSelfOutput, self).__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
-        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
@@ -370,7 +373,7 @@ class BertOutput(nn.Module):
     def __init__(self, config):
         super(BertOutput, self).__init__()
         self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
-        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
@@ -434,7 +437,7 @@ class BertPredictionHeadTransform(nn.Module):
             self.transform_act_fn = ACT2FN[config.hidden_act]
         else:
             self.transform_act_fn = config.hidden_act
-        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=1e-12)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
 
     def forward(self, hidden_states):
         hidden_states = self.dense(hidden_states)