Release: 2.1.1

GLUE on TPU

.circleci/config.yml

@@ -9,7 +9,7 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install torch
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -38,7 +38,7 @@ jobs:
         parallelism: 1
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: sudo pip install tensorboardX scikit-learn
@@ -65,7 +65,7 @@ jobs:
             - image: circleci/python:2.7
         steps:
             - checkout
-            - run: sudo pip install tensorflow==2.0.0-rc0
+            - run: sudo pip install tensorflow
             - run: sudo pip install --progress-bar off .
             - run: sudo pip install pytest codecov pytest-cov
             - run: python -m pytest -sv ./transformers/tests/ --cov

README.md

@@ -545,4 +545,14 @@ for batch in train_data:
 
 ## Citation
 
-At the moment, there is no paper associated with Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.
+We now have a paper you can cite for the 🤗 Transformers library:
+```
+@misc{wolf2019transformers,
+    title={Transformers: State-of-the-art Natural Language Processing},
+    author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Jamie Brew},
+    year={2019},
+    eprint={1910.03771},
+    archivePrefix={arXiv},
+    primaryClass={cs.CL}
+}
+```

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.1.0'
+release = u'2.1.1'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/pretrained_models.rst

@@ -53,6 +53,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-cased-finetuned-mrpc``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | The ``bert-base-cased`` model fine-tuned on MRPC                                                                                    |
 |                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__)                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-cased``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased German text by DBMDZ                                                                                               |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-dbmdz-uncased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
+|                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |

docs/source/serialization.rst

@@ -33,6 +33,8 @@ where
     * ``bert-large-uncased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-cased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
     * ``bert-large-uncased-whole-word-masking-finetuned-squad``: The ``bert-large-uncased-whole-word-masking`` model finetuned on SQuAD (using the ``run_bert_squad.py`` examples). Results: *exact_match: 86.91579943235573, f1: 93.1532499015869*
+    * ``bert-base-german-dbmdz-cased``: Trained on German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
+    * ``bert-base-german-dbmdz-uncased``: Trained on (uncased) German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
     * ``openai-gpt``: OpenAI GPT English model, 12-layer, 768-hidden, 12-heads, 110M parameters
     * ``gpt2``: OpenAI GPT-2 English model, 12-layer, 768-hidden, 12-heads, 117M parameters
     * ``gpt2-medium``: OpenAI GPT-2 English model, 24-layer, 1024-hidden, 16-heads, 345M parameters

examples/contrib/run_swag.py

@@ -31,9 +31,13 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForMultipleChoice, BertTokenizer)

examples/distillation/distiller.py

@@ -19,7 +19,6 @@ import os
 import math
 import psutil
 import time
-from tensorboardX import SummaryWriter
 from tqdm import trange, tqdm
 import numpy as np
 import psutil
@@ -31,6 +30,11 @@ from torch.optim import AdamW
 from torch.utils.data.distributed import DistributedSampler
 from torch.utils.data import RandomSampler, BatchSampler, DataLoader
 
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from transformers import WarmupLinearSchedule
 
 from utils import logger

examples/distillation/run_squad_w_distillation.py

@@ -30,9 +30,13 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
 from torch.utils.data.distributed import DistributedSampler
 import torch.nn.functional as F
 import torch.nn as nn
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,

examples/run_generation.py

@@ -107,7 +107,8 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
     return logits
 
 
-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0, is_xlnet=False, xlm_lang=None, device='cpu'):
+def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
+                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
     context = torch.tensor(context, dtype=torch.long, device=device)
     context = context.unsqueeze(0).repeat(num_samples, 1)
     generated = context
@@ -125,10 +126,16 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
                 target_mapping[0, 0, -1] = 1.0  # predict last token
                 inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
 
+            if is_xlm_mlm and xlm_mask_token:
+                # XLM MLM models are direct models (predict same token, not next token)
+                # => need one additional dummy token in the input (will be masked and guessed)
+                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
+                inputs = {'input_ids': input_ids}
+
             if xlm_lang is not None:
                 inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
 
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
+            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
             next_token_logits = outputs[0][0, -1, :] / (temperature if temperature > 0 else 1.)
 
             # reptition penalty from CTRL (https://arxiv.org/abs/1909.05858)
@@ -167,10 +174,7 @@ def main():
     parser.add_argument('--stop_token', type=str, default=None,
                         help="Token at which text generation is stopped")
     args = parser.parse_args()
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7 : 
-            print('CTRL typically works better with lower temperatures (and lower top_k).')
-        
+
     args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
     args.n_gpu = torch.cuda.device_count()
 
@@ -190,7 +194,11 @@ def main():
     elif args.length < 0:
         args.length = MAX_LENGTH  # avoid infinite loop
 
-    print(args)
+    logger.info(args)
+    if args.model_type in ["ctrl"]:
+        if args.temperature > 0.7 : 
+            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
+
     while True:
         xlm_lang = None
         # XLM Language usage detailed in the issues #1414
@@ -204,6 +212,13 @@ def main():
                     language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
             xlm_lang = tokenizer.lang2id[language]
 
+        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
+        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
+        if is_xlm_mlm:
+            xlm_mask_token = tokenizer.mask_token_id
+        else:
+            xlm_mask_token = None
+
         raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
         if args.model_type in ["transfo-xl", "xlnet"]:
             # Models with memory likes to have a long prompt for short inputs.
@@ -218,6 +233,8 @@ def main():
             top_p=args.top_p,
             repetition_penalty=args.repetition_penalty,
             is_xlnet=bool(args.model_type == "xlnet"),
+            is_xlm_mlm=is_xlm_mlm,
+            xlm_mask_token=xlm_mask_token,
             xlm_lang=xlm_lang,
             device=args.device,
         )

examples/run_glue.py

@@ -28,7 +28,12 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
@@ -155,7 +160,7 @@ def train(args, train_dataset, model, tokenizer):
                 torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
 
             tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
+            if (step + 1) % args.gradient_accumulation_steps == 0 and not args.tpu:
                 optimizer.step()
                 scheduler.step()  # Update learning rate schedule
                 model.zero_grad()
@@ -181,6 +186,11 @@ def train(args, train_dataset, model, tokenizer):
                     torch.save(args, os.path.join(output_dir, 'training_args.bin'))
                     logger.info("Saving model checkpoint to %s", output_dir)
 
+            if args.tpu:
+                args.xla_model.optimizer_step(optimizer, barrier=True)
+                model.zero_grad()
+                global_step += 1
+
             if args.max_steps > 0 and global_step > args.max_steps:
                 epoch_iterator.close()
                 break
@@ -380,6 +390,15 @@ def main():
     parser.add_argument('--seed', type=int, default=42,
                         help="random seed for initialization")
 
+    parser.add_argument('--tpu', action='store_true',
+                        help="Whether to run on the TPU defined in the environment variables")
+    parser.add_argument('--tpu_ip_address', type=str, default='',
+                        help="TPU IP address if none are set in the environment variables")
+    parser.add_argument('--tpu_name', type=str, default='',
+                        help="TPU name if none are set in the environment variables")
+    parser.add_argument('--xrt_tpu_config', type=str, default='',
+                        help="XRT TPU config if none are set in the environment variables")
+
     parser.add_argument('--fp16', action='store_true',
                         help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
     parser.add_argument('--fp16_opt_level', type=str, default='O1',
@@ -413,6 +432,23 @@ def main():
         args.n_gpu = 1
     args.device = device
 
+    if args.tpu:
+        if args.tpu_ip_address:
+            os.environ["TPU_IP_ADDRESS"] = args.tpu_ip_address
+        if args.tpu_name:
+            os.environ["TPU_NAME"] = args.tpu_name
+        if args.xrt_tpu_config:
+            os.environ["XRT_TPU_CONFIG"] = args.xrt_tpu_config
+
+        assert "TPU_IP_ADDRESS" in os.environ
+        assert "TPU_NAME" in os.environ
+        assert "XRT_TPU_CONFIG" in os.environ
+
+        import torch_xla
+        import torch_xla.core.xla_model as xm
+        args.device = xm.xla_device()
+        args.xla_model = xm
+
     # Setup logging
     logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                         datefmt = '%m/%d/%Y %H:%M:%S',
@@ -458,7 +494,7 @@ def main():
 
 
     # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0) and not args.tpu:
         # Create output directory if needed
         if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
             os.makedirs(args.output_dir)

examples/run_lm_finetuning.py

@@ -34,7 +34,12 @@ import numpy as np
 import torch
 from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
@@ -61,7 +66,7 @@ class TextDataset(Dataset):
     def __init__(self, tokenizer, file_path='train', block_size=512):
         assert os.path.isfile(file_path)
         directory, filename = os.path.split(file_path)
-        cached_features_file = os.path.join(directory, 'cached_lm_' + block_size + '_' + filename)
+        cached_features_file = os.path.join(directory, 'cached_lm_' + str(block_size) + '_' + filename)
 
         if os.path.exists(cached_features_file):
             logger.info("Loading features from cached file %s", cached_features_file)

examples/run_multiple_choice.py

@@ -29,7 +29,12 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tensorboardX import SummaryWriter
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
 from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,

examples/run_squad.py

@@ -28,9 +28,13 @@ import torch
 from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                               TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
 
-from tensorboardX import SummaryWriter
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except:
+    from tensorboardX import SummaryWriter
+
+from tqdm import tqdm, trange
 
 from transformers import (WEIGHTS_NAME, BertConfig,
                                   BertForQuestionAnswering, BertTokenizer,

setup.py

@@ -38,7 +38,7 @@ from setuptools import find_packages, setup
 
 setup(
     name="transformers",
-    version="2.1.0",
+    version="2.1.1",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",

transformers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "2.1.0"
+__version__ = "2.1.1"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -110,65 +110,55 @@ if is_tf_available():
                                    TFBertForMaskedLM, TFBertForNextSentencePrediction,
                                    TFBertForSequenceClassification, TFBertForMultipleChoice,
                                    TFBertForTokenClassification, TFBertForQuestionAnswering,
-                                   load_bert_pt_weights_in_tf2,
                                    TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_gpt2 import (TFGPT2PreTrainedModel, TFGPT2MainLayer,
                                    TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel,
-                                   load_gpt2_pt_weights_in_tf2,
                                    TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_openai import (TFOpenAIGPTPreTrainedModel, TFOpenAIGPTMainLayer,
                                      TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TFOpenAIGPTDoubleHeadsModel,
-                                     load_openai_gpt_pt_weights_in_tf2,
                                      TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_transfo_xl import (TFTransfoXLPreTrainedModel, TFTransfoXLMainLayer,
                                          TFTransfoXLModel, TFTransfoXLLMHeadModel,
-                                         load_transfo_xl_pt_weights_in_tf2,
                                          TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_xlnet import (TFXLNetPreTrainedModel, TFXLNetMainLayer,
                                     TFXLNetModel, TFXLNetLMHeadModel,
                                     TFXLNetForSequenceClassification,
                                     TFXLNetForQuestionAnsweringSimple,
-                                    load_xlnet_pt_weights_in_tf2,
                                     TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_xlm import (TFXLMPreTrainedModel, TFXLMMainLayer,
                                   TFXLMModel, TFXLMWithLMHeadModel,
                                   TFXLMForSequenceClassification,
                                   TFXLMForQuestionAnsweringSimple,
-                                  load_xlm_pt_weights_in_tf2,
                                   TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_roberta import (TFRobertaPreTrainedModel, TFRobertaMainLayer,
                                       TFRobertaModel, TFRobertaForMaskedLM,
                                       TFRobertaForSequenceClassification,
-                                      load_roberta_pt_weights_in_tf2,
                                       TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
                                          TFDistilBertModel, TFDistilBertForMaskedLM,
                                          TFDistilBertForSequenceClassification,
                                          TFDistilBertForQuestionAnswering,
-                                         load_distilbert_pt_weights_in_tf2,
                                          TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_ctrl import (TFCTRLPreTrainedModel, TFCTRLModel,
                                     TFCTRLLMHeadModel,
-                                    load_ctrl_pt_weights_in_tf2,
                                     TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
 # TF 2.0 <=> PyTorch conversion utilities
-if is_tf_available() and is_torch_available():
-    from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
-                                            load_pytorch_checkpoint_in_tf2_model,
-                                            load_pytorch_weights_in_tf2_model,
-                                            load_pytorch_model_in_tf2_model,
-                                            load_tf2_checkpoint_in_pytorch_model,
-                                            load_tf2_weights_in_pytorch_model,
-                                            load_tf2_model_in_pytorch_model)
+from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
+                                        load_pytorch_checkpoint_in_tf2_model,
+                                        load_pytorch_weights_in_tf2_model,
+                                        load_pytorch_model_in_tf2_model,
+                                        load_tf2_checkpoint_in_pytorch_model,
+                                        load_tf2_weights_in_pytorch_model,
+                                        load_tf2_model_in_pytorch_model)
 
 if not is_tf_available() and not is_torch_available():
     logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."

transformers/configuration_bert.py

@@ -40,6 +40,8 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
+    'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
+    'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
 }
 
 

transformers/configuration_utils.py

@@ -53,7 +53,8 @@ class PretrainedConfig(object):
         self.num_labels = kwargs.pop('num_labels', 2)
         self.output_attentions = kwargs.pop('output_attentions', False)
         self.output_hidden_states = kwargs.pop('output_hidden_states', False)
-        self.torchscript = kwargs.pop('torchscript', False)
+        self.output_past = kwargs.pop('output_past', True)  # Not used by all models
+        self.torchscript = kwargs.pop('torchscript', False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop('use_bfloat16', False)
         self.pruned_heads = kwargs.pop('pruned_heads', {})
 
@@ -130,20 +131,19 @@ class PretrainedConfig(object):
         # redirect to the cache, if necessary
         try:
             resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
-        except EnvironmentError as e:
+        except EnvironmentError:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
-                logger.error(
-                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
-                        config_file))
+                msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
+                        config_file)
             else:
-                logger.error(
-                    "Model name '{}' was not found in model name list ({}). "
-                    "We assumed '{}' was a path or url but couldn't find any file "
-                    "associated to this path or url.".format(
+                msg = "Model name '{}' was not found in model name list ({}). " \
+                      "We assumed '{}' was a path or url to a configuration file named {} or " \
+                      "a directory containing such a file but couldn't find any such file at this path or url.".format(
                         pretrained_model_name_or_path,
                         ', '.join(cls.pretrained_config_archive_map.keys()),
-                        config_file))
-            raise e
+                        config_file, CONFIG_NAME)
+            raise EnvironmentError(msg)
+
         if resolved_config_file == config_file:
             logger.info("loading configuration file {}".format(config_file))
         else:
@@ -154,7 +154,7 @@ class PretrainedConfig(object):
         config = cls.from_json_file(resolved_config_file)
 
         if hasattr(config, 'pruned_heads'):
-            config.pruned_heads = dict((int(key), set(value)) for key, value in config.pruned_heads.items())
+            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
 
         # Update config with kwargs if needed
         to_remove = []
@@ -165,7 +165,7 @@ class PretrainedConfig(object):
         for key in to_remove:
             kwargs.pop(key, None)
 
-        logger.info("Model config %s", config)
+        logger.info("Model config %s", str(config))
         if return_unused_kwargs:
             return config, kwargs
         else:

transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -24,15 +24,16 @@ import tensorflow as tf
 
 from transformers import is_torch_available, cached_path
 
-from transformers import (BertConfig, TFBertForPreTraining, TFBertForQuestionAnswering, TFBertForSequenceClassification, load_bert_pt_weights_in_tf2, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  GPT2Config, TFGPT2LMHeadModel, load_gpt2_pt_weights_in_tf2, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  XLNetConfig, TFXLNetLMHeadModel, load_xlnet_pt_weights_in_tf2, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  XLMConfig, TFXLMWithLMHeadModel, load_xlm_pt_weights_in_tf2, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  TransfoXLConfig, TFTransfoXLLMHeadModel, load_transfo_xl_pt_weights_in_tf2, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, load_openai_gpt_pt_weights_in_tf2, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, load_roberta_pt_weights_in_tf2, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, load_distilbert_pt_weights_in_tf2, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  CTRLConfig, TFCTRLLMHeadModel, load_ctrl_pt_weights_in_tf2, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+from transformers import (load_pytorch_checkpoint_in_tf2_model,
+    BertConfig, TFBertForPreTraining, TFBertForQuestionAnswering, TFBertForSequenceClassification, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  GPT2Config, TFGPT2LMHeadModel, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  XLNetConfig, TFXLNetLMHeadModel, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  XLMConfig, TFXLMWithLMHeadModel, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  TransfoXLConfig, TFTransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
 
 if is_torch_available():
     import torch
@@ -71,27 +72,27 @@ import logging
 logging.basicConfig(level=logging.INFO)
 
 MODEL_CLASSES = {
-    'bert': (BertConfig, TFBertForPreTraining, load_bert_pt_weights_in_tf2, BertForPreTraining, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-large-uncased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, load_bert_pt_weights_in_tf2, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-large-cased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, load_bert_pt_weights_in_tf2, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'bert-base-cased-finetuned-mrpc': (BertConfig, TFBertForSequenceClassification, load_bert_pt_weights_in_tf2, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'gpt2': (GPT2Config, TFGPT2LMHeadModel, load_gpt2_pt_weights_in_tf2, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'xlnet': (XLNetConfig, TFXLNetLMHeadModel, load_xlnet_pt_weights_in_tf2, XLNetLMHeadModel, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'xlm': (XLMConfig, TFXLMWithLMHeadModel, load_xlm_pt_weights_in_tf2, XLMWithLMHeadModel, XLM_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'transfo-xl': (TransfoXLConfig, TFTransfoXLLMHeadModel, load_transfo_xl_pt_weights_in_tf2, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'openai-gpt': (OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, load_openai_gpt_pt_weights_in_tf2, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'roberta': (RobertaConfig, TFRobertaForMaskedLM, load_roberta_pt_weights_in_tf2, RobertaForMaskedLM, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, load_roberta_pt_weights_in_tf2, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, load_distilbert_pt_weights_in_tf2, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, load_distilbert_pt_weights_in_tf2, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, load_ctrl_pt_weights_in_tf2, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
+    'bert': (BertConfig, TFBertForPreTraining, BertForPreTraining, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-large-uncased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-large-cased-whole-word-masking-finetuned-squad': (BertConfig, TFBertForQuestionAnswering, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'bert-base-cased-finetuned-mrpc': (BertConfig, TFBertForSequenceClassification, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'gpt2': (GPT2Config, TFGPT2LMHeadModel, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'xlnet': (XLNetConfig, TFXLNetLMHeadModel, XLNetLMHeadModel, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'xlm': (XLMConfig, TFXLMWithLMHeadModel, XLMWithLMHeadModel, XLM_PRETRAINED_MODEL_ARCHIVE_MAP, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'transfo-xl': (TransfoXLConfig, TFTransfoXLLMHeadModel, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'openai-gpt': (OpenAIGPTConfig, TFOpenAIGPTLMHeadModel, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'roberta': (RobertaConfig, TFRobertaForMaskedLM, RobertaForMaskedLM, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'roberta-large-mnli': (RobertaConfig, TFRobertaForSequenceClassification, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP)
 }
 
 def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file, tf_dump_path, compare_with_pt_model=False, use_cached_models=True):
     if model_type not in MODEL_CLASSES:
         raise ValueError("Unrecognized model type, should be one of {}.".format(list(MODEL_CLASSES.keys())))
 
-    config_class, model_class, loading_fct, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
+    config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
 
     # Initialise TF model
     if config_file in aws_config_map:
@@ -105,7 +106,8 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
     # Load weights from tf checkpoint
     if pytorch_checkpoint_path in aws_model_maps:
         pytorch_checkpoint_path = cached_path(aws_model_maps[pytorch_checkpoint_path], force_download=not use_cached_models)
-    tf_model = loading_fct(tf_model, pytorch_checkpoint_path)
+    # Load PyTorch checkpoint in tf2 model:
+    tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
 
     if compare_with_pt_model:
         inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
@@ -147,7 +149,7 @@ def convert_all_pt_checkpoints_to_tf(args_model_type, tf_dump_path, model_shortc
         if model_type not in MODEL_CLASSES:
             raise ValueError("Unrecognized model type {}, should be one of {}.".format(model_type, list(MODEL_CLASSES.keys())))
 
-        config_class, model_class, loading_fct, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
+        config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
 
         if model_shortcut_names_or_path is None:
             model_shortcut_names_or_path = list(aws_model_maps.keys())

transformers/modeling_bert.py

@@ -48,6 +48,8 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
+    'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
+    'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
 }
 
 def load_tf_weights_in_bert(model, config, tf_checkpoint_path):

transformers/modeling_ctrl.py

@@ -269,16 +269,16 @@ class CTRLModel(CTRLPreTrainedModel):
     def __init__(self, config):
         super(CTRLModel, self).__init__(config)
         self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
+
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
-        
+
         self.pos_encoding = positional_encoding(config.n_positions, self.d_model_size, torch.float)
-            
-        self.output_attentions = config.output_attentions
 
         self.w = nn.Embedding(config.vocab_size, config.n_embd)
 
-        
         self.dropout = nn.Dropout(config.embd_pdrop)
         self.h = nn.ModuleList([EncoderLayer(config.n_embd,
                                              config.n_head,
@@ -378,7 +378,8 @@ class CTRLModel(CTRLPreTrainedModel):
                         attention_mask=attention_mask,
                         head_mask=head_mask[i])
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -388,7 +389,9 @@ class CTRLModel(CTRLPreTrainedModel):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:

transformers/modeling_gpt2.py

@@ -347,6 +347,7 @@ class GPT2Model(GPT2PreTrainedModel):
         super(GPT2Model, self).__init__(config)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
 
         self.wte = nn.Embedding(config.vocab_size, config.n_embd)
         self.wpe = nn.Embedding(config.n_positions, config.n_embd)
@@ -440,7 +441,8 @@ class GPT2Model(GPT2PreTrainedModel):
                             head_mask=head_mask[i])
 
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -452,7 +454,9 @@ class GPT2Model(GPT2PreTrainedModel):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
@@ -460,7 +464,7 @@ class GPT2Model(GPT2PreTrainedModel):
             attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
             all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
             outputs = outputs + (all_attentions,)
-        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+        return outputs  # last hidden state, (presents), (all hidden_states), (attentions)
 
 
 @add_start_docstrings("""The GPT2 Model transformer with a language modeling head on top

transformers/modeling_tf_bert.py

@@ -30,7 +30,6 @@ import tensorflow as tf
 from .configuration_bert import BertConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -52,14 +51,6 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_bert_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """ Gaussian Error Linear Unit.
     Original Implementation of the gelu activation function in Google Bert repo when initially created.
@@ -545,7 +536,6 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
     """
     config_class = BertConfig
     pretrained_model_archive_map = TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_bert_pt_weights_in_tf2
     base_model_prefix = "bert"
 
 

transformers/modeling_tf_ctrl.py

@@ -27,20 +27,11 @@ import tensorflow as tf
 from .configuration_ctrl import CTRLConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list, TFSharedEmbeddings
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
 TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP = {"ctrl": "https://s3.amazonaws.com/models.huggingface.co/bert/ctrl-tf_model.h5"}
 
-def load_ctrl_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def angle_defn(pos, i, d_model_size):
     angle_rates = 1 / np.power(10000, (2 * (i//2)) / np.float32(d_model_size))
     return pos * angle_rates
@@ -177,12 +168,14 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
     def __init__(self, config, **kwargs):
         super(TFCTRLMainLayer, self).__init__(**kwargs)
         self.output_hidden_states = config.output_hidden_states
+        self.output_attentions = config.output_attentions
+        self.output_past = config.output_past
+
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
 
         self.pos_encoding = positional_encoding(config.n_positions, self.d_model_size)
 
-        self.output_attentions = config.output_attentions
 
         self.w = TFSharedEmbeddings(config.vocab_size,
                                     config.n_embd,
@@ -299,7 +292,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
             outputs = h([hidden_states, mask, layer_past, attention_mask, head_mask[i]], training=training)
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+
+            if self.output_past:
+                presents = presents + (present,)
 
             if self.output_attentions:
                 all_attentions.append(outputs[2])
@@ -309,7 +304,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         if self.output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states, presents)
+        outputs = (hidden_states,)
+        if self.output_past:
+            outputs = outputs + (presents,)
         if self.output_hidden_states:
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
@@ -327,7 +324,6 @@ class TFCTRLPreTrainedModel(TFPreTrainedModel):
     config_class = CTRLConfig
     pretrained_model_archive_map = TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "transformer"
-    load_pt_weights = load_ctrl_pt_weights_in_tf2
 
 
 CTRL_START_DOCSTRING = r"""    CTRL model was proposed in 

transformers/modeling_tf_distilbert.py

@@ -31,7 +31,6 @@ import tensorflow as tf
 from .configuration_distilbert import DistilBertConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -66,14 +65,6 @@ def gelu_new(x):
         (np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
     return x * cdf
 
-def load_distilbert_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-    attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    tf_inputs = [inputs_list, attns_list]
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
 class TFEmbeddings(tf.keras.layers.Layer):
     def __init__(self, config, **kwargs):
         super(TFEmbeddings, self).__init__(**kwargs)
@@ -454,7 +445,6 @@ class TFDistilBertPreTrainedModel(TFPreTrainedModel):
     """
     config_class = DistilBertConfig
     pretrained_model_archive_map = TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_distilbert_pt_weights_in_tf2
     base_model_prefix = "distilbert"
 
 

transformers/modeling_tf_gpt2.py

@@ -32,7 +32,6 @@ from .modeling_tf_utils import (TFPreTrainedModel, TFConv1D, TFSharedEmbeddings,
                                 TFSequenceSummary, shape_list, get_initializer)
 from .configuration_gpt2 import GPT2Config
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -42,14 +41,6 @@ TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models
                                      "distilgpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-tf_model.h5",}
 
 
-def load_gpt2_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """Gaussian Error Linear Unit.
     This is a smoother version of the RELU.
@@ -350,7 +341,6 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
     """
     config_class = GPT2Config
     pretrained_model_archive_map = TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_gpt2_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 

transformers/modeling_tf_openai.py

@@ -32,21 +32,12 @@ from .modeling_tf_utils import (TFPreTrainedModel, TFConv1D, TFSharedEmbeddings,
                                 TFSequenceSummary, shape_list, get_initializer)
 from .configuration_openai import OpenAIGPTConfig
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
 TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP = {"openai-gpt": "https://s3.amazonaws.com/models.huggingface.co/bert/openai-gpt-tf_model.h5"}
 
 
-def load_openai_gpt_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """Gaussian Error Linear Unit.
     This is a smoother version of the RELU.
@@ -335,7 +326,6 @@ class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel):
     """
     config_class = OpenAIGPTConfig
     pretrained_model_archive_map = TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_openai_gpt_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 

transformers/modeling_tf_pytorch_utils.py

@@ -25,8 +25,6 @@ import numpy
 
 logger = logging.getLogger(__name__)
 
-DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-
 def convert_tf_weight_name_to_pt_weight_name(tf_name, start_prefix_to_remove=''):
     """ Convert a TF 2.0 model variable name in a pytorch model weight name.
 
@@ -105,7 +103,7 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
         raise e
 
     if tf_inputs is None:
-        tf_inputs = tf.constant(DUMMY_INPUTS)
+        tf_inputs = tf_model.dummy_inputs
 
     if tf_inputs is not None:
         tfo = tf_model(tf_inputs, training=False)  # Make sure model is built

transformers/modeling_tf_roberta.py

@@ -26,7 +26,6 @@ import tensorflow as tf
 from .configuration_roberta import RobertaConfig
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 from .modeling_tf_bert import TFBertEmbeddings, TFBertMainLayer, gelu, gelu_new
 
@@ -38,14 +37,6 @@ TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-tf_model.h5",
 }
 
-def load_roberta_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 class TFRobertaEmbeddings(TFBertEmbeddings):
     """
     Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
@@ -83,7 +74,7 @@ class TFRobertaMainLayer(TFBertMainLayer):
             input_ids = inputs
 
         if tf.not_equal(tf.reduce_sum(input_ids[:, 0]), 0):
-            logger.warning("A sequence with no special tokens has been passed to the RoBERTa model. "
+            tf.print("A sequence with no special tokens has been passed to the RoBERTa model. "
                            "This model requires special tokens in order to work. "
                            "Please specify add_special_tokens=True in your encoding.")
 
@@ -96,7 +87,6 @@ class TFRobertaPreTrainedModel(TFPreTrainedModel):
     """
     config_class = RobertaConfig
     pretrained_model_archive_map = TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_roberta_pt_weights_in_tf2
     base_model_prefix = "roberta"
 
 

transformers/modeling_tf_transfo_xl.py

@@ -33,7 +33,6 @@ from .configuration_transfo_xl import TransfoXLConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFConv1D, TFSequenceSummary, shape_list, get_initializer
 from .modeling_tf_transfo_xl_utilities import TFAdaptiveSoftmaxMask
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -41,14 +40,6 @@ TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'transfo-xl-wt103': "https://s3.amazonaws.com/models.huggingface.co/bert/transfo-xl-wt103-tf_model.h5",
 }
 
-def load_transfo_xl_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 class TFPositionalEmbedding(tf.keras.layers.Layer):
     def __init__(self, demb, **kwargs):
         super(TFPositionalEmbedding, self).__init__(**kwargs)
@@ -577,7 +568,6 @@ class TFTransfoXLPreTrainedModel(TFPreTrainedModel):
     """
     config_class = TransfoXLConfig
     pretrained_model_archive_map = TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_transfo_xl_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 

transformers/modeling_tf_utils.py

@@ -25,9 +25,11 @@ import tensorflow as tf
 
 from .configuration_utils import PretrainedConfig
 from .file_utils import cached_path, WEIGHTS_NAME, TF_WEIGHTS_NAME, TF2_WEIGHTS_NAME
+from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
 
 class TFPreTrainedModel(tf.keras.Model):
     r""" Base class for all TF models.
@@ -48,8 +50,8 @@ class TFPreTrainedModel(tf.keras.Model):
     """
     config_class = None
     pretrained_model_archive_map = {}
-    load_pt_weights = lambda model, config, path: None
     base_model_prefix = ""
+    dummy_inputs = tf.constant(DUMMY_INPUTS)  # dummy inputs to build the network
 
     def __init__(self, config, *inputs, **kwargs):
         super(TFPreTrainedModel, self).__init__(*inputs, **kwargs)
@@ -262,17 +264,16 @@ class TFPreTrainedModel(tf.keras.Model):
 
         if from_pt:
             # Load from a PyTorch checkpoint
-            return cls.load_pt_weights(model, resolved_archive_file)
+            return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file)
 
-        inputs = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-        ret = model(inputs, training=False)  # build the network with dummy inputs
+        ret = model(model.dummy_inputs, training=False)  # build the network with dummy inputs
 
         assert os.path.isfile(resolved_archive_file), "Error retrieving file {}".format(resolved_archive_file)
         # 'by_name' allow us to do transfer learning by skipping/adding layers
         # see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1339-L1357
         model.load_weights(resolved_archive_file, by_name=True)
 
-        ret = model(inputs, training=False)  # Make sure restore ops are run
+        ret = model(model.dummy_inputs, training=False)  # Make sure restore ops are run
 
         return model
 
@@ -393,26 +394,26 @@ class TFSequenceSummary(tf.keras.layers.Layer):
             # We can probably just use the multi-head attention module of PyTorch >=1.1.0
             raise NotImplementedError
 
-        self.summary = None
-        if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
+        self.has_summary = hasattr(config, 'summary_use_proj') and config.summary_use_proj
+        if self.has_summary:
             if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
                 num_classes = config.num_labels
             else:
                 num_classes = config.hidden_size
             self.summary = tf.keras.layers.Dense(num_classes,
-                                                 kernel_initializer=get_initializer(initializer_range),
-                                                 name='summary')
+                                                    kernel_initializer=get_initializer(initializer_range),
+                                                    name='summary')
 
-        self.activation = None
-        if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
+        self.has_activation = hasattr(config, 'summary_activation') and config.summary_activation == 'tanh'
+        if self.has_activation:
             self.activation = tf.keras.activations.tanh
 
-        self.first_dropout = None
-        if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
+        self.has_first_dropout = hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0
+        if self.has_first_dropout:
             self.first_dropout = tf.keras.layers.Dropout(config.summary_first_dropout)
 
-        self.last_dropout = None
-        if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
+        self.has_last_dropout = hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0
+        if self.has_last_dropout:
             self.last_dropout = tf.keras.layers.Dropout(config.summary_last_dropout)
 
     def call(self, inputs, training=False):
@@ -455,17 +456,17 @@ class TFSequenceSummary(tf.keras.layers.Layer):
         elif self.summary_type == 'attn':
             raise NotImplementedError
 
-        if training and self.first_dropout is not None:
-            output = self.first_dropout(output)
+        if self.has_first_dropout:
+            output = self.first_dropout(output, training=training)
 
-        if self.summary is not None:
+        if self.has_summary:
             output = self.summary(output)
 
-        if self.activation is not None:
+        if self.has_activation:
             output = self.activation(output)
 
-        if training and self.last_dropout is not None:
-            output = self.last_dropout(output)
+        if self.has_last_dropout:
+            output = self.last_dropout(output, training=training)
 
         return output
 

transformers/modeling_tf_xlm.py

@@ -25,9 +25,8 @@ import numpy as np
 import tensorflow as tf
 
 from .configuration_xlm import XLMConfig
-from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer
+from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer, DUMMY_INPUTS
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
@@ -45,19 +44,6 @@ TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_xlm_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    # build the network
-    inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
-    attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    if tf_model.config.use_lang_emb and tf_model.config.n_langs > 1:
-        langs_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
-    else:
-        langs_list = None
-    tf_inputs = [inputs_list, attns_list, langs_list]
-    tfo = tf_model(tf_inputs, training=False)
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def create_sinusoidal_embeddings(n_pos, dim, out):
     position_enc = np.array([
         [pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)]
@@ -441,9 +427,19 @@ class TFXLMPreTrainedModel(TFPreTrainedModel):
     """
     config_class = XLMConfig
     pretrained_model_archive_map = TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_xlm_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
+    @property
+    def dummy_inputs(self):
+        # Sometimes XLM has language embeddings so don't forget to build them as well if needed
+        inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
+        attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        if self.config.use_lang_emb and self.config.n_langs > 1:
+            langs_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        else:
+            langs_list = None
+        return [inputs_list, attns_list, langs_list]
+
 
 XLM_START_DOCSTRING = r"""    The XLM model was proposed in
     `Cross-lingual Language Model Pretraining`_

transformers/modeling_tf_xlnet.py

@@ -30,7 +30,6 @@ import tensorflow as tf
 from .configuration_xlnet import XLNetConfig
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list, get_initializer
 from .file_utils import add_start_docstrings
-from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 
 logger = logging.getLogger(__name__)
@@ -41,13 +40,6 @@ TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP = {
 }
 
 
-def load_xlnet_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
-    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-    tf_inputs = tf.constant(inputs_list)
-    tfo = tf_model(tf_inputs, training=False)  # build the network
-    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)
-
-
 def gelu(x):
     """ Implementation of the gelu activation function.
         XLNet is using OpenAI GPT's gelu
@@ -362,6 +354,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         super(TFXLNetMainLayer, self).__init__(**kwargs)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -421,16 +414,13 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
-            if self.reuse_len is not None and self.reuse_len > 0:
-                curr_out = curr_out[:self.reuse_len]
+        if self.reuse_len is not None and self.reuse_len > 0:
+            curr_out = curr_out[:self.reuse_len]
 
-            if prev_mem is None:
-                new_mem = curr_out[-self.mem_len:]
-            else:
-                new_mem = tf.concat([prev_mem, curr_out], 0)[-self.mem_len:]
+        if prev_mem is None:
+            new_mem = curr_out[-self.mem_len:]
+        else:
+            new_mem = tf.concat([prev_mem, curr_out], 0)[-self.mem_len:]
 
         return tf.stop_gradient(new_mem)
 
@@ -546,8 +536,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             raise ValueError('Unsupported attention type: {}'.format(self.attn_type))
 
         # data mask: input mask & perm mask
-        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) "
-        "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
+        assert input_mask is None or attention_mask is None, "You can only use one of input_mask (uses 1 for padding) " \
+            "or attention_mask (uses 0 for padding, added for compatbility with BERT). Please choose one."
         if input_mask is None and attention_mask is not None:
             input_mask = 1.0 - attention_mask
         if input_mask is not None and perm_mask is not None:
@@ -632,7 +622,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
             # cache new mems
-            new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+                new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
                 hidden_states.append((output_h, output_g) if output_g is not None else output_h)
 
@@ -650,7 +641,11 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         output = self.dropout(output_g if output_g is not None else output_h, training=training)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (tf.transpose(output, perm=(1, 0, 2)), new_mems)
+        outputs = (tf.transpose(output, perm=(1, 0, 2)),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(tf.transpose(h, perm=(1, 0, 2)) for hs in hidden_states for h in hs)
@@ -661,7 +656,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 class TFXLNetPreTrainedModel(TFPreTrainedModel):
@@ -670,7 +665,6 @@ class TFXLNetPreTrainedModel(TFPreTrainedModel):
     """
     config_class = XLNetConfig
     pretrained_model_archive_map = TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
-    load_pt_weights = load_xlnet_pt_weights_in_tf2
     base_model_prefix = "transformer"
 
 
@@ -777,7 +771,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -819,7 +813,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **prediction_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -863,7 +857,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -874,7 +868,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **logits**: ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``tf.Tensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -918,7 +912,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel):
 
         outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # return logits, mems, (hidden states), (attentions)
+        return outputs  # return logits, (mems), (hidden states), (attentions)
 
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -932,6 +926,11 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``tf.Tensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -971,7 +970,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel):
 
         outputs = (start_logits, end_logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
 
-        return outputs  # start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
 #     the hidden-states output to compute `span start logits` and `span end logits`). """,

transformers/modeling_utils.py

@@ -316,20 +316,20 @@ class PreTrainedModel(nn.Module):
             # redirect to the cache, if necessary
             try:
                 resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
-            except EnvironmentError as e:
+            except EnvironmentError:
                 if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
-                    logger.error(
-                        "Couldn't reach server at '{}' to download pretrained weights.".format(
-                            archive_file))
+                    msg = "Couldn't reach server at '{}' to download pretrained weights.".format(
+                            archive_file)
                 else:
-                    logger.error(
-                        "Model name '{}' was not found in model name list ({}). "
-                        "We assumed '{}' was a path or url but couldn't find any file "
-                        "associated to this path or url.".format(
+                    msg = "Model name '{}' was not found in model name list ({}). " \
+                        "We assumed '{}' was a path or url to model weight files named one of {} but " \
+                        "couldn't find any such file at this path or url.".format(
                             pretrained_model_name_or_path,
                             ', '.join(cls.pretrained_model_archive_map.keys()),
-                            archive_file))
-                raise e
+                            archive_file,
+                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME])
+                raise EnvironmentError(msg)
+
             if resolved_archive_file == archive_file:
                 logger.info("loading weights file {}".format(archive_file))
             else:

transformers/modeling_xlnet.py

@@ -239,45 +239,60 @@ class XLNetRelativeAttention(nn.Module):
 
         return x
 
+    @staticmethod
+    def rel_shift_bnij(x, klen=-1):
+        x_size = x.shape
+
+        x = x.reshape(x_size[0], x_size[1], x_size[3], x_size[2])
+        x = x[:, :, 1:, :]
+        x = x.reshape(x_size[0], x_size[1], x_size[2], x_size[3]-1)
+        # Note: the tensor-slice form was faster in my testing than torch.index_select
+        #       However, tracing doesn't like the nature of the slice, and if klen changes
+        #       during the run then it'll fail, whereas index_select will be fine.
+        x = torch.index_select(x, 3, torch.arange(klen, device=x.device, dtype=torch.long))
+        # x = x[:, :, :, :klen]
+
+        return x
+
     def rel_attn_core(self, q_head, k_head_h, v_head_h, k_head_r, seg_mat=None, attn_mask=None, head_mask=None):
         """Core relative positional attention operations."""
 
         # content based attention score
-        ac = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_w_bias, k_head_h)
+        ac = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_w_bias, k_head_h)
 
         # position based attention score
-        bd = torch.einsum('ibnd,jbnd->ijbn', q_head + self.r_r_bias, k_head_r)
-        bd = self.rel_shift(bd, klen=ac.shape[1])
+        bd = torch.einsum('ibnd,jbnd->bnij', q_head + self.r_r_bias, k_head_r)
+        bd = self.rel_shift_bnij(bd, klen=ac.shape[3])
 
         # segment based attention score
         if seg_mat is None:
             ef = 0
         else:
             ef = torch.einsum('ibnd,snd->ibns', q_head + self.r_s_bias, self.seg_embed)
-            ef = torch.einsum('ijbs,ibns->ijbn', seg_mat, ef)
+            ef = torch.einsum('ijbs,ibns->bnij', seg_mat, ef)
 
         # merge attention scores and perform masking
         attn_score = (ac + bd + ef) * self.scale
         if attn_mask is not None:
             # attn_score = attn_score * (1 - attn_mask) - 1e30 * attn_mask
             if attn_mask.dtype == torch.float16:
-                attn_score = attn_score - 65500 * attn_mask
+                attn_score = attn_score - 65500 * torch.einsum('ijbn->bnij', attn_mask)
             else:
-                attn_score = attn_score - 1e30 * attn_mask
+                attn_score = attn_score - 1e30 * torch.einsum('ijbn->bnij', attn_mask)
 
         # attention probability
-        attn_prob = F.softmax(attn_score, dim=1)
+        attn_prob = F.softmax(attn_score, dim=3)
         attn_prob = self.dropout(attn_prob)
 
         # Mask heads if we want to
         if head_mask is not None:
-            attn_prob = attn_prob * head_mask
+            attn_prob = attn_prob * torch.einsum('ijbn->bnij', head_mask)
 
         # attention output
-        attn_vec = torch.einsum('ijbn,jbnd->ibnd', attn_prob, v_head_h)
+        attn_vec = torch.einsum('bnij,jbnd->ibnd', attn_prob, v_head_h)
 
         if self.output_attentions:
-            return attn_vec, attn_prob
+            return attn_vec, torch.einsum('bnij->ijbn', attn_prob)
 
         return attn_vec
 
@@ -555,7 +570,7 @@ class XLNetModel(XLNetPreTrainedModel):
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -581,6 +596,7 @@ class XLNetModel(XLNetPreTrainedModel):
         super(XLNetModel, self).__init__(config)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.output_past = config.output_past
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -637,16 +653,13 @@ class XLNetModel(XLNetPreTrainedModel):
 
     def cache_mem(self, curr_out, prev_mem):
         """cache hidden states into memory."""
-        if self.mem_len is None or self.mem_len == 0:
-            return None
-        else:
-            if self.reuse_len is not None and self.reuse_len > 0:
-                curr_out = curr_out[:self.reuse_len]
+        if self.reuse_len is not None and self.reuse_len > 0:
+            curr_out = curr_out[:self.reuse_len]
 
-            if prev_mem is None:
-                new_mem = curr_out[-self.mem_len:]
-            else:
-                new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len:]
+        if prev_mem is None:
+            new_mem = curr_out[-self.mem_len:]
+        else:
+            new_mem = torch.cat([prev_mem, curr_out], dim=0)[-self.mem_len:]
 
         return new_mem.detach()
 
@@ -817,8 +830,9 @@ class XLNetModel(XLNetPreTrainedModel):
         attentions = []
         hidden_states = []
         for i, layer_module in enumerate(self.layer):
-            # cache new mems
-            new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
+            if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+                # cache new mems
+                new_mems = new_mems + (self.cache_mem(output_h, mems[i]),)
             if self.output_hidden_states:
                 hidden_states.append((output_h, output_g) if output_g is not None else output_h)
 
@@ -836,7 +850,11 @@ class XLNetModel(XLNetPreTrainedModel):
         output = self.dropout(output_g if output_g is not None else output_h)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (output.permute(1, 0, 2).contiguous(), new_mems)
+        outputs = (output.permute(1, 0, 2).contiguous(),)
+
+        if self.mem_len is not None and self.mem_len > 0 and self.output_past:
+            outputs = outputs + (new_mems,)
+
         if self.output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(h.permute(1, 0, 2).contiguous() for hs in hidden_states for h in hs)
@@ -847,7 +865,7 @@ class XLNetModel(XLNetPreTrainedModel):
             attentions = tuple(t.permute(2, 3, 0, 1).contiguous() for t in attentions)
             outputs = outputs + (attentions,)
 
-        return outputs  # outputs, new_mems, (hidden_states), (attentions)
+        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a language modeling head on top
@@ -867,7 +885,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
             Language modeling loss.
         **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -918,7 +936,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
+                                               input_mask=input_mask,
                                                head_mask=head_mask)
 
         logits = self.lm_loss(transformer_outputs[0])
@@ -932,7 +950,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
                             labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a sequence classification/regression head on top (a linear layer on top of
@@ -951,7 +969,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
             Classification (or regression if config.num_labels==1) loss.
         **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
             Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -992,7 +1010,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
+                                               input_mask=input_mask,
                                                head_mask=head_mask)
         output = transformer_outputs[0]
 
@@ -1011,7 +1029,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
                 loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 @add_start_docstrings("""XLNet Model with a multiple choice classification head on top (a linear layer on top of
     the pooled output and a softmax) e.g. for RACE/SWAG tasks. """,
@@ -1046,6 +1064,11 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
         **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above).
             Classification scores (before SoftMax).
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
+            list of ``torch.FloatTensor`` (one for each layer):
+            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -1102,7 +1125,7 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
             loss = loss_fct(reshaped_logits, labels.view(-1))
             outputs = (loss,) + outputs
 
-        return outputs  # return (loss), logits, mems, (hidden states), (attentions)
+        return outputs  # return (loss), logits, (mems), (hidden states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1126,7 +1149,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             Span-start scores (before SoftMax).
         **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
             Span-end scores (before SoftMax).
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1169,7 +1192,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
                                     perm_mask=perm_mask,
                                     target_mapping=target_mapping,
                                     token_type_ids=token_type_ids,
-                                    input_mask=input_mask, 
+                                    input_mask=input_mask,
                                     head_mask=head_mask)
 
         sequence_output = outputs[0]
@@ -1197,7 +1220,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             total_loss = (start_loss + end_loss) / 2
             outputs = (total_loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        return outputs  # (loss), start_logits, end_logits, (mems), (hidden_states), (attentions)
 
 
 @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
@@ -1239,7 +1262,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
         **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
             ``torch.FloatTensor`` of shape ``(batch_size,)``
             Log probabilities for the ``is_impossible`` label of the answers.
-        **mems**:
+        **mems**: (`optional`, returned when ``config.mem_len > 0``)
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
@@ -1284,7 +1307,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
                                                perm_mask=perm_mask,
                                                target_mapping=target_mapping,
                                                token_type_ids=token_type_ids,
-                                               input_mask=input_mask, 
+                                               input_mask=input_mask,
                                                head_mask=head_mask)
         hidden_states = transformer_outputs[0]
         start_logits = self.start_logits(hidden_states, p_mask=p_mask)

transformers/tests/modeling_common_test.py

@@ -17,8 +17,10 @@ from __future__ import division
 from __future__ import print_function
 
 import copy
+import sys
 import os
 import shutil
+import tempfile
 import json
 import random
 import uuid
@@ -31,6 +33,7 @@ from transformers import is_torch_available
 
 if is_torch_available():
     import torch
+    import numpy as np
 
     from transformers import (PretrainedConfig, PreTrainedModel,
                                     BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -38,6 +41,20 @@ if is_torch_available():
 else:
     pytestmark = pytest.mark.skip("Require Torch")
 
+if sys.version_info[0] == 2:
+    import cPickle as pickle
+
+    class TemporaryDirectory(object):
+        """Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
+        def __enter__(self):
+            self.name = tempfile.mkdtemp()
+            return self.name
+        def __exit__(self, exc_type, exc_value, traceback):
+            shutil.rmtree(self.name)
+else:
+    import pickle
+    TemporaryDirectory = tempfile.TemporaryDirectory
+    unicode = str
 
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
@@ -57,6 +74,29 @@ class CommonTestCases:
         test_resize_embeddings = True
         test_head_masking = True
 
+        def test_save_load(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                model.eval()
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
+
+                with TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+                    with torch.no_grad():
+                        after_outputs = model(**inputs_dict)
+
+                    # Make sure we don't have nans
+                    out_1 = after_outputs[0].numpy()
+                    out_2 = outputs[0].numpy()
+                    out_1 = out_1[~np.isnan(out_1)]
+                    out_2 = out_2[~np.isnan(out_2)]
+                    max_diff = np.amax(np.abs(out_1 - out_2))
+                    self.assertLessEqual(max_diff, 1e-5)
+
         def test_initialization(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

transformers/tests/modeling_tf_common_test.py

@@ -22,6 +22,7 @@ import random
 import shutil
 import unittest
 import uuid
+import tempfile
 
 import pytest
 import sys
@@ -36,6 +37,20 @@ if is_tf_available():
 else:
     pytestmark = pytest.mark.skip("Require TensorFlow")
 
+if sys.version_info[0] == 2:
+    import cPickle as pickle
+
+    class TemporaryDirectory(object):
+        """Context manager for tempfile.mkdtemp() so it's usable with "with" statement."""
+        def __enter__(self):
+            self.name = tempfile.mkdtemp()
+            return self.name
+        def __exit__(self, exc_type, exc_value, traceback):
+            shutil.rmtree(self.name)
+else:
+    import pickle
+    TemporaryDirectory = tempfile.TemporaryDirectory
+    unicode = str
 
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
@@ -66,13 +81,31 @@ class TFCommonTestCases:
             #             self.assertIn(param.data.mean().item(), [0.0, 1.0],
             #             msg="Parameter {} of model {} seems not properly initialized".format(name, model_class))
 
+        def test_save_load(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                outputs = model(inputs_dict)
+
+                with TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+                    after_outputs = model(inputs_dict)
+
+                    # Make sure we don't have nans
+                    out_1 = after_outputs[0].numpy()
+                    out_2 = outputs[0].numpy()
+                    out_1 = out_1[~np.isnan(out_1)]
+                    out_2 = out_2[~np.isnan(out_2)]
+                    max_diff = np.amax(np.abs(out_1 - out_2))
+                    self.assertLessEqual(max_diff, 1e-5)
 
         def test_pt_tf_model_equivalence(self):
             if not is_torch_available():
                 return
 
             import torch
-            import numpy as np
             import transformers
 
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -99,6 +132,34 @@ class TFCommonTestCases:
                 max_diff = np.amax(np.abs(tfo[0].numpy() - pto[0].numpy()))
                 self.assertLessEqual(max_diff, 2e-2)
 
+        def test_compile_tf_model(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            input_ids = tf.keras.Input(batch_shape=(2, 2000), name='input_ids', dtype='int32')
+            optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
+            loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+            metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
+
+            for model_class in self.all_model_classes:
+                # Prepare our model
+                model = model_class(config)
+                
+                # Let's load it from the disk to be sure we can use pretrained weights
+                with TemporaryDirectory() as tmpdirname:
+                    outputs = model(inputs_dict)  # build the model
+                    model.save_pretrained(tmpdirname)
+                    model = model_class.from_pretrained(tmpdirname)
+
+                outputs_dict = model(input_ids)
+                hidden_states = outputs_dict[0]
+
+                # Add a dense layer on top to test intetgration with other keras modules
+                outputs = tf.keras.layers.Dense(2, activation='softmax', name='outputs')(hidden_states)
+
+                # Compile extended model
+                extended_model = tf.keras.Model(inputs=[input_ids], outputs=[outputs])
+                extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
+
         def test_keyword_and_dict_args(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

transformers/tests/modeling_tf_xlnet_test.py

@@ -161,6 +161,11 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                 "outputs": outputs.numpy(),
             }
 
+            config.mem_len = 0
+            model = TFXLNetModel(config)
+            no_mems_outputs = model(inputs)
+            self.parent.assertEqual(len(no_mems_outputs), 1)
+
             self.parent.assertListEqual(
                 list(result["outputs"].shape),
                 [self.batch_size, self.seq_length, self.hidden_size])

transformers/tests/modeling_xlnet_test.py

@@ -150,6 +150,12 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 "outputs": outputs,
             }
 
+            config.mem_len = 0
+            model = XLNetModel(config)
+            model.eval()
+            no_mems_outputs = model(input_ids_1)
+            self.parent.assertEqual(len(no_mems_outputs), 1)
+
             self.parent.assertListEqual(
                 list(result["outputs"].size()),
                 [self.batch_size, self.seq_length, self.hidden_size])

transformers/tokenization_bert.py

@@ -44,6 +44,8 @@ PRETRAINED_VOCAB_FILES_MAP = {
         'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-vocab.txt",
         'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-vocab.txt",
         'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-vocab.txt",
+        'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-vocab.txt",
+        'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-vocab.txt",
     }
 }
 
@@ -61,6 +63,8 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': 512,
     'bert-large-cased-whole-word-masking-finetuned-squad': 512,
     'bert-base-cased-finetuned-mrpc': 512,
+    'bert-base-german-dbmdz-cased': 512,
+    'bert-base-german-dbmdz-uncased': 512,
 }
 
 PRETRAINED_INIT_CONFIGURATION = {
@@ -77,6 +81,8 @@ PRETRAINED_INIT_CONFIGURATION = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': {'do_lower_case': True},
     'bert-large-cased-whole-word-masking-finetuned-squad': {'do_lower_case': False},
     'bert-base-cased-finetuned-mrpc': {'do_lower_case': False},
+    'bert-base-german-dbmdz-cased': {'do_lower_case': False},
+    'bert-base-german-dbmdz-uncased': {'do_lower_case': True},
 }
 
 

transformers/tokenization_ctrl.py

@@ -22,8 +22,6 @@ import os
 import regex as re
 from io import open
 
-from .tokenization_bert import BasicTokenizer
-
 from .tokenization_utils import PreTrainedTokenizer
 
 logger = logging.getLogger(__name__)
@@ -48,39 +46,11 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'ctrl': 256,
 }
 
-def text_standardize(text):
-    """
-    fixes some issues the spacy tokenizer had on books corpus
-    also does some whitespace standardization
-    """
-    text = text.replace('—', '-')
-    text = text.replace('–', '-')
-    text = text.replace('―', '-')
-    text = text.replace('…', '...')
-    text = text.replace('´', "'")
-    text = re.sub(r'''(-+|~+|!+|"+|;+|\?+|\++|,+|\)+|\(+|\\+|\/+|\*+|\[+|\]+|}+|{+|\|+|_+)''', r' \1 ', text)
-    text = re.sub(r'\s*\n\s*', ' \n ', text)
-    text = re.sub(r'[^\S\n]+', ' ', text)
-    return text.strip()
-
-
 def get_pairs(word):
     """Return set of symbol pairs in a word.
 
     Word is represented as tuple of symbols (symbols being variable-length strings).
     """
-    # pairs = []
-    # prev_char = word[0]
-    # for i, char in enumerate(word[1:]):
-    #     #_i = i + 1
-    #     #if word[_i+1:] == tuple('</w>'):
-    #     #    pairs.append((prev_char, char+'</w>'))
-    #     #    break
-    #     #else:
-    #     if True:
-    #         pairs.append((prev_char, char))
-    #         prev_char = char
-
     pairs = set()
     prev_char = word[0]
     for char in word[1:]:
@@ -108,17 +78,6 @@ class CTRLTokenizer(PreTrainedTokenizer):
         self.max_len_single_sentence = self.max_len # no default special tokens - you can update this value if you add special tokens
         self.max_len_sentences_pair = self.max_len # no default special tokens - you can update this value if you add special tokens
 
-        try:
-            import ftfy
-            from spacy.lang.en import English
-            _nlp = English()
-            self.nlp = _nlp.Defaults.create_tokenizer(_nlp)
-            self.fix_text = ftfy.fix_text
-        except ImportError:
-            logger.warning("ftfy or spacy is not installed using BERT BasicTokenizer instead of SpaCy & ftfy.")
-            self.nlp = BasicTokenizer(do_lower_case=True)
-            self.fix_text = None
-
         self.encoder = json.load(open(vocab_file, encoding="utf-8"))
         self.decoder = {v:k for k,v in self.encoder.items()}
         merges = open(merges_file, encoding='utf-8').read().split('\n')[1:-1]
@@ -177,22 +136,11 @@ class CTRLTokenizer(PreTrainedTokenizer):
         """ Tokenize a string.
         """
         split_tokens = []
-        if self.fix_text is None:
-            # Using BERT's BasicTokenizer
-            text = self.nlp.tokenize(text)
-            for token in text:
-                split_tokens.extend([t for t in self.bpe(token).split(' ')])
-        else:
-            # Using SpaCy & ftfy (original tokenization process of OpenAI GPT)
-            text = self.nlp(text_standardize(self.fix_text(text)))
-            for token in text:
-                split_tokens.extend([t for t in self.bpe(token.text.lower()).split(' ')])
-        # for token in text.split():
-        #     if sys.version_info[0] == 2:
-        #         token = ''.join(self.byte_encoder[ord(b)] for b in token) # Maps all our bytes to unicode strings, avoiding controle tokens of the BPE (spaces in our case)
-        #     else:
-        #         token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) # Maps all our bytes to unicode strings, avoiding controle tokens of the BPE (spaces in our case)
-        #     bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(' '))
+
+        text = text.split(' ')
+
+        for token in text:
+            split_tokens.extend([t for t in self.bpe(token).split(' ')])
         return split_tokens
 
     def _convert_token_to_id(self, token):

transformers/tokenization_utils.py

@@ -337,13 +337,13 @@ class PreTrainedTokenizer(object):
                 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 ({}). "
-                    "We assumed '{}' was a path or url but couldn't find tokenizer files"
-                    "at this path or url.".format(
+                raise EnvironmentError(
+                    "Model name '{}' was not found in tokenizers model name list ({}). "
+                    "We assumed '{}' was a path or url to a directory containing vocabulary files "
+                    "named {} but couldn't find such vocabulary files at this path or url.".format(
                         pretrained_model_name_or_path, ', '.join(s3_models),
-                        pretrained_model_name_or_path, ))
-                return None
+                        pretrained_model_name_or_path, 
+                        list(cls.vocab_files_names.values())))
 
         # Get files from url, cache, or disk depending on the case
         try:
@@ -353,17 +353,18 @@ class PreTrainedTokenizer(object):
                     resolved_vocab_files[file_id] = None
                 else:
                     resolved_vocab_files[file_id] = cached_path(file_path, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
-        except EnvironmentError as e:
+        except EnvironmentError:
             if pretrained_model_name_or_path in s3_models:
-                logger.error("Couldn't reach server to download vocabulary.")
+                msg = "Couldn't reach server at '{}' to download vocabulary files."
             else:
-                logger.error(
-                    "Model name '{}' was not found in model name list ({}). "
-                    "We assumed '{}' was a path or url but couldn't find files {} "
-                    "at this path or url.".format(
+                msg = "Model name '{}' was not found in tokenizers model name list ({}). " \
+                    "We assumed '{}' was a path or url to a directory containing vocabulary files " \
+                    "named {}, but couldn't find such vocabulary files at this path or url.".format(
                         pretrained_model_name_or_path, ', '.join(s3_models),
-                        pretrained_model_name_or_path, str(vocab_files.keys())))
-            raise e
+                        pretrained_model_name_or_path,
+                        list(cls.vocab_files_names.values()))
+
+            raise EnvironmentError(msg)
 
         for file_id, file_path in vocab_files.items():
             if file_path == resolved_vocab_files[file_id]: