Merge branch 'master' into auto_models

.github/ISSUE_TEMPLATE/bug-report.md

+---
+name: "\U0001F41B Bug Report"
+about: Submit a bug report to help us improve PyTorch Transformers
+---
+
+## 🐛 Bug
+
+<!-- Important information -->
+
+Model I am using (Bert, XLNet....):
+
+Language I am using the model on (English, Chinese....):
+
+The problem arise when using:
+* [ ] the official example scripts: (give details)
+* [ ] my own modified scripts: (give details)
+
+The tasks I am working on is:
+* [ ] an official GLUE/SQUaD task: (give the name)
+* [ ] my own task or dataset: (give details)
+
+## To Reproduce
+
+Steps to reproduce the behavior:
+
+1.
+2.
+3.
+
+<!-- If you have a code sample, error messages, stack traces, please provide it here as well. -->
+
+## Expected behavior
+
+<!-- A clear and concise description of what you expected to happen. -->
+
+## Environment
+
+* OS:
+* Python version:
+* PyTorch version:
+* PyTorch Transformers version (or branch):
+* Using GPU ?
+* Distributed of parallel setup ?
+* Any other relevant information:
+
+## Additional context
+
+<!-- Add any other context about the problem here. -->
\ No newline at end of file

.github/ISSUE_TEMPLATE/feature-request.md

+---
+name: "\U0001F680 Feature Request"
+about: Submit a proposal/request for a new PyTorch Transformers feature
+---
+
+## 🚀 Feature
+
+<!-- A clear and concise description of the feature proposal. Please provide a link to the paper and code in case they exist. -->
+
+## Motivation
+
+<!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too. -->
+
+## Additional context
+
+<!-- Add any other context or screenshots about the feature request here. -->
\ No newline at end of file

.github/ISSUE_TEMPLATE/migration.md

+---
+name: "\U0001F4DA Migration from PyTorch-pretrained-Bert"
+about: Report a problem when migrating from PyTorch-pretrained-Bert to PyTorch-Transformers
+---
+
+## 📚 Migration
+
+<!-- Important information -->
+
+Model I am using (Bert, XLNet....):
+
+Language I am using the model on (English, Chinese....):
+
+The problem arise when using:
+* [ ] the official example scripts: (give details)
+* [ ] my own modified scripts: (give details)
+
+The tasks I am working on is:
+* [ ] an official GLUE/SQUaD task: (give the name)
+* [ ] my own task or dataset: (give details)
+
+Details of the issue:
+
+<!-- A clear and concise description of the migration issue. If you have code snippets, please provide it here as well. -->
+
+## Environment
+
+* OS:
+* Python version:
+* PyTorch version:
+* PyTorch Transformers version (or branch):
+* Using GPU ?
+* Distributed of parallel setup ?
+* Any other relevant information:
+
+## Checklist
+
+- [ ] I have read the migration guide in the readme.
+- [ ] I checked if a related official extension example runs on my machine.
+
+## Additional context
+
+<!-- Add any other context about the problem here. -->
\ No newline at end of file

.github/ISSUE_TEMPLATE/question-help.md

+---
+name: "❓Questions & Help"
+about: Start a general discussion related to PyTorch Transformers
+---
+
+## ❓ Questions & Help
+
+<!-- A clear and concise description of the question. -->
\ No newline at end of file

README.md

@@ -18,7 +18,7 @@ These implementations have been tested on several datasets (see the example scri
 | Section | Description |
 |-|-|
 | [Installation](#installation) | How to install the package |
-| [Quick tour: Usage](#quick-tour-usage) | Tokenizers & models usage: Bert and GPT-2 |
+| [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
 | [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-pytorch-transformers) | Migrating your code from pytorch-pretrained-bert to pytorch-transformers |
 | [Documentation](https://huggingface.co/pytorch-transformers/) | Full API documentation and more |
@@ -56,6 +56,16 @@ python -m pytest -sv ./pytorch_transformers/tests/
 python -m pytest -sv ./examples/
 ```
 
+### Do you want to run a Transformer model on a mobile device?
+
+You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
+
+It contains an example of a conversion script from a Pytorch trained Transformer model (here, `GPT-2`) to a CoreML model that runs on iOS devices.
+
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
+or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
+
+
 ## Quick tour
 
 Let's do a very quick overview of PyTorch-Transformers. Detailed examples for each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [full documentation](https://huggingface.co/pytorch-transformers/).
@@ -195,7 +205,7 @@ python ./examples/run_glue.py \
     --warmup_steps=120
 ```
 
-On this machine we thus have a batch size of 32, please increase `gradient_accumulation_steps` to reach the same batch size if you have a smaller machine. These hyper-parameters should results in a Pearson correlation coefficient of `+0.917` on the development set.
+On this machine we thus have a batch size of 32, please increase `gradient_accumulation_steps` to reach the same batch size if you have a smaller machine. These hyper-parameters should result in a Pearson correlation coefficient of `+0.917` on the development set.
 
 #### Fine-tuning Bert model on the MRPC classification task
 
@@ -265,7 +275,7 @@ This is the model provided as `bert-large-uncased-whole-word-masking-finetuned-s
 ### `run_generation.py`: Text generation with GPT, GPT-2, Transformer-XL and XLNet
 
 A conditional generation script is also included to generate text from a prompt.
-The generation script include the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
+The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
 
 Here is how to run the script with the small version of OpenAI GPT-2 model:
 
@@ -284,7 +294,7 @@ Here is a quick summary of what you should take care of when migrating from `pyt
 
 The main breaking change when migrating from `pytorch-pretrained-bert` to `pytorch-transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
 
-The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
+The exact content of the tuples for each model are detailed in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
 
 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
 
@@ -383,6 +393,7 @@ for batch in train_data:
     torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
     scheduler.step()
     optimizer.step()
+    optimizer.zero_grad()
 ```
 
 ## Citation

docs/source/installation.rst

@@ -49,4 +49,17 @@ If you want to reproduce the original tokenization process of the ``OpenAI GPT``
    pip install spacy ftfy==4.4.3
    python -m spacy download en
 
-If you don't install ``ftfy`` and ``SpaCy``\ , the ``OpenAI GPT`` tokenizer defaults to tokenize using BERT's ``BasicTokenizer`` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
+If you don't install ``ftfy`` and ``SpaCy``\ , the ``OpenAI GPT`` tokenizer will default to tokenize using BERT's ``BasicTokenizer`` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
+
+
+Do you want to run a Transformer model on a mobile device?
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+You should check out our `swift-coreml-transformers <https://github.com/huggingface/swift-coreml-transformers>`_ repo.
+
+It contains an example of a conversion script from a Pytorch trained Transformer model (here, ``GPT-2``) to a CoreML model that runs on iOS devices.
+
+It also contains an implementation of BERT for Question answering.
+
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
+or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
\ No newline at end of file

examples/run_glue.py

@@ -92,6 +92,10 @@ def train(args, train_dataset, model, tokenizer):
             raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
         model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
 
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
     # Distributed training (should be after apex fp16 initialization)
     if args.local_rank != -1:
         model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
@@ -243,6 +247,9 @@ def evaluate(args, model, tokenizer, prefix=""):
 
 
 def load_and_cache_examples(args, task, tokenizer, evaluate=False):
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
     processor = processors[task]()
     output_mode = output_modes[task]
     # Load data features from cache or dataset file
@@ -269,6 +276,9 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
             logger.info("Saving features into cached file %s", cached_features_file)
             torch.save(features, cached_features_file)
 
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
     # Convert to Tensors and build dataset
     all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
     all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
@@ -418,8 +428,6 @@ def main():
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
     model.to(args.device)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
 
     logger.info("Training/evaluation parameters %s", args)
 

examples/run_squad.py

@@ -101,6 +101,10 @@ def train(args, train_dataset, model, tokenizer):
             raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
         model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
 
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
     # Distributed training (should be after apex fp16 initialization)
     if args.local_rank != -1:
         model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
@@ -241,7 +245,10 @@ def evaluate(args, model, tokenizer, prefix=""):
     # Compute predictions
     output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
     output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+    if args.version_2_with_negative:
         output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+    else:
+        output_null_log_odds_file = None
 
     if args.model_type in ['xlnet', 'xlm']:
         # XLNet uses a more complex post-processing procedure
@@ -265,6 +272,9 @@ def evaluate(args, model, tokenizer, prefix=""):
 
 
 def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
     # Load data features from cache or dataset file
     input_file = args.predict_file if evaluate else args.train_file
     cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
@@ -289,6 +299,9 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
             logger.info("Saving features into cached file %s", cached_features_file)
             torch.save(features, cached_features_file)
 
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
     # Convert to Tensors and build dataset
     all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
     all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
@@ -457,8 +470,6 @@ def main():
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
     model.to(args.device)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
 
     logger.info("Training/evaluation parameters %s", args)
 

pytorch_transformers/__init__.py

@@ -10,20 +10,20 @@ from .tokenization_utils import (PreTrainedTokenizer)
 
 from .modeling_auto import (AutoConfig, AutoModel)
 
-from .modeling_bert import (BertConfig, BertModel, BertForPreTraining,
+from .modeling_bert import (BertConfig, BertPreTrainedModel, BertModel, BertForPreTraining,
                             BertForMaskedLM, BertForNextSentencePrediction,
                             BertForSequenceClassification, BertForMultipleChoice,
                             BertForTokenClassification, BertForQuestionAnswering,
                             load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                             BERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
-from .modeling_openai import (OpenAIGPTConfig, OpenAIGPTModel,
+from .modeling_openai import (OpenAIGPTConfig, OpenAIGPTPreTrainedModel, OpenAIGPTModel,
                               OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
                               load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
                               OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
-from .modeling_transfo_xl import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel,
+from .modeling_transfo_xl import (TransfoXLConfig, TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
                                   load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
-from .modeling_gpt2 import (GPT2Config, GPT2Model,
+from .modeling_gpt2 import (GPT2Config, GPT2PreTrainedModel, GPT2Model,
                             GPT2LMHeadModel, GPT2DoubleHeadsModel,
                             load_tf_weights_in_gpt2, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
                             GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
@@ -32,7 +32,7 @@ from .modeling_xlnet import (XLNetConfig,
                              XLNetForSequenceClassification, XLNetForQuestionAnswering,
                              load_tf_weights_in_xlnet, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
                              XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
-from .modeling_xlm import (XLMConfig, XLMModel,
+from .modeling_xlm import (XLMConfig, XLMPreTrainedModel , XLMModel,
                            XLMWithLMHeadModel, XLMForSequenceClassification,
                            XLMForQuestionAnswering, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)

pytorch_transformers/convert_pytorch_checkpoint_to_tf.py

@@ -41,7 +41,7 @@ def convert_pytorch_checkpoint_to_tf(model:BertModel, ckpt_dir:str, model_name:s
         N BertForQuestionAnswering
     """
 
-    tensors_to_transopse = (
+    tensors_to_transpose = (
         "dense.weight",
         "attention.self.query",
         "attention.self.key",
@@ -62,33 +62,33 @@ def convert_pytorch_checkpoint_to_tf(model:BertModel, ckpt_dir:str, model_name:s
     if not os.path.isdir(ckpt_dir):
         os.makedirs(ckpt_dir)
 
-    session = tf.Session()
     state_dict = model.state_dict()
-    tf_vars = []
 
     def to_tf_var_name(name:str):
         for patt, repl in iter(var_map):
             name = name.replace(patt, repl)
         return 'bert/{}'.format(name)
 
-    def assign_tf_var(tensor:np.ndarray, name:str):
-        tmp_var = tf.Variable(initial_value=tensor)
-        tf_var = tf.get_variable(dtype=tmp_var.dtype, shape=tmp_var.shape, name=name)
-        op = tf.assign(ref=tf_var, value=tmp_var)
-        session.run(tf.variables_initializer([tmp_var, tf_var]))
-        session.run(fetches=[op, tf_var])
+    def create_tf_var(tensor:np.ndarray, name:str, session:tf.Session):
+        tf_dtype = tf.dtypes.as_dtype(tensor.dtype)
+        tf_var = tf.get_variable(dtype=tf_dtype, shape=tensor.shape, name=name, initializer=tf.zeros_initializer())
+        session.run(tf.variables_initializer([tf_var]))
+        session.run(tf_var)
         return tf_var
 
+    tf.reset_default_graph()
+    with tf.Session() as session:
         for var_name in state_dict:
             tf_name = to_tf_var_name(var_name)
             torch_tensor = state_dict[var_name].numpy()
-        if any([x in var_name for x in tensors_to_transopse]):
+            if any([x in var_name for x in tensors_to_transpose]):
                 torch_tensor = torch_tensor.T
-        tf_tensor = assign_tf_var(tensor=torch_tensor, name=tf_name)
-        tf_vars.append(tf_tensor)
-        print("{0}{1}initialized".format(tf_name, " " * (60 - len(tf_name))))
+            tf_var = create_tf_var(tensor=torch_tensor, name=tf_name, session=session)
+            tf.keras.backend.set_value(tf_var, torch_tensor)
+            tf_weight = session.run(tf_var)
+            print("Successfully created {}: {}".format(tf_name, np.allclose(tf_weight, torch_tensor)))
 
-    saver = tf.train.Saver(tf_vars)
+        saver = tf.train.Saver(tf.trainable_variables())
         saver.save(session, os.path.join(ckpt_dir, model_name.replace("-", "_") + ".ckpt"))
 
 

pytorch_transformers/convert_transfo_xl_checkpoint_to_pytorch.py

@@ -24,10 +24,9 @@ from io import open
 import torch
 
 import pytorch_transformers.tokenization_transfo_xl as data_utils
-from pytorch_transformers.modeling_transfo_xl import (CONFIG_NAME,
-                                                         WEIGHTS_NAME,
-                                                         TransfoXLConfig,
-                                                         TransfoXLLMHeadModel,
+
+from pytorch_transformers import CONFIG_NAME, WEIGHTS_NAME
+from pytorch_transformers.modeling_transfo_xl import (TransfoXLConfig, TransfoXLLMHeadModel,
                                                       load_tf_weights_in_transfo_xl)
 from pytorch_transformers.tokenization_transfo_xl import (CORPUS_NAME, VOCAB_FILES_NAMES)
 

pytorch_transformers/modeling_openai.py

@@ -538,7 +538,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
             All labels set to ``-1`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``

pytorch_transformers/modeling_utils.py

@@ -39,6 +39,20 @@ WEIGHTS_NAME = "pytorch_model.bin"
 TF_WEIGHTS_NAME = 'model.ckpt'
 
 
+try:
+    from torch.nn import Identity
+except ImportError:
+    # Older PyTorch compatibility
+    class Identity(nn.Module):
+        r"""A placeholder identity operator that is argument-insensitive.
+        """
+        def __init__(self, *args, **kwargs):
+            super(Identity, self).__init__()
+
+        def forward(self, input):
+            return input
+
+
 if not six.PY2:
     def add_start_docstrings(*docstr):
         def docstring_decorator(fn):
@@ -783,7 +797,7 @@ class SequenceSummary(nn.Module):
             # We can probably just use the multi-head attention module of PyTorch >=1.1.0
             raise NotImplementedError
 
-        self.summary = nn.Identity()
+        self.summary = Identity()
         if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
             if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
                 num_classes = config.num_labels
@@ -791,15 +805,15 @@ class SequenceSummary(nn.Module):
                 num_classes = config.hidden_size
             self.summary = nn.Linear(config.hidden_size, num_classes)
 
-        self.activation = nn.Identity()
+        self.activation = Identity()
         if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
             self.activation = nn.Tanh()
 
-        self.first_dropout = nn.Identity()
+        self.first_dropout = Identity()
         if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
             self.first_dropout = nn.Dropout(config.summary_first_dropout)
 
-        self.last_dropout = nn.Identity()
+        self.last_dropout = Identity()
         if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
             self.last_dropout = nn.Dropout(config.summary_last_dropout)
 

pytorch_transformers/tokenization_utils.py

@@ -226,26 +226,46 @@ class PreTrainedTokenizer(object):
         s3_models = list(cls.max_model_input_sizes.keys())
         vocab_files = {}
         if pretrained_model_name_or_path in s3_models:
+            # Get the vocabulary from AWS S3 bucket
             for file_id, map_list in cls.pretrained_vocab_files_map.items():
                 vocab_files[file_id] = map_list[pretrained_model_name_or_path]
         else:
+            # Get the vocabulary from local files
             logger.info(
                 "Model name '{}' not found in model shortcut name list ({}). "
                 "Assuming '{}' is a path or url to a directory containing tokenizer files.".format(
                     pretrained_model_name_or_path, ', '.join(s3_models),
                     pretrained_model_name_or_path))
-            all_vocab_files_names = {'added_tokens_file': ADDED_TOKENS_FILE,
-                                     'special_tokens_map_file': SPECIAL_TOKENS_MAP_FILE}
-            all_vocab_files_names.update(cls.vocab_files_names)
-            for file_id, file_name in all_vocab_files_names.items():
+
+            # Look for the tokenizer main vocabulary files
+            for file_id, file_name in cls.vocab_files_names.items():
                 if os.path.isdir(pretrained_model_name_or_path):
+                    # If a directory is provided we look for the standard filenames
                     full_file_name = os.path.join(pretrained_model_name_or_path, file_name)
                 else:
+                    # If a path to a file is provided we use it (will only work for non-BPE tokenizer using a single vocabulary file)
                     full_file_name = pretrained_model_name_or_path
                 if not os.path.exists(full_file_name):
                     logger.info("Didn't find file {}. We won't load it.".format(full_file_name))
                     full_file_name = None
                 vocab_files[file_id] = full_file_name
+
+            # Look for the additional tokens files
+            all_vocab_files_names = {'added_tokens_file': ADDED_TOKENS_FILE,
+                                     'special_tokens_map_file': SPECIAL_TOKENS_MAP_FILE}
+
+            # If a path to a file was provided, get the parent directory
+            saved_directory = pretrained_model_name_or_path
+            if os.path.exists(saved_directory) and not os.path.isdir(saved_directory):
+                saved_directory = os.path.dirname(saved_directory)
+
+            for file_id, file_name in all_vocab_files_names.items():
+                full_file_name = os.path.join(saved_directory, file_name)
+                if not os.path.exists(full_file_name):
+                    logger.info("Didn't find file {}. We won't load it.".format(full_file_name))
+                    full_file_name = None
+                vocab_files[file_id] = full_file_name
+
             if all(full_file_name is None for full_file_name in vocab_files.values()):
                 logger.error(
                     "Model name '{}' was not found in model name list ({}). "
@@ -333,7 +353,7 @@ class PreTrainedTokenizer(object):
 
         with open(added_tokens_file, 'w', encoding='utf-8') as f:
             if self.added_tokens_encoder:
-                out_str = json.dumps(self.added_tokens_decoder, ensure_ascii=False)
+                out_str = json.dumps(self.added_tokens_encoder, ensure_ascii=False)
             else:
                 out_str = u"{}"
             f.write(out_str)