resolved_decode_with_prefix

decode_with_prefix
2020-04-02 14:19:12 -04:00 · 2020-04-02 14:14:11 -04:00 · 2020-04-02 14:09:31 -04:00 · 2020-04-02 14:07:08 -04:00 · 2020-04-02 15:18:33 +02:00 · 2020-04-02 12:34:05 +02:00
74 changed files with 5196 additions and 1024 deletions
@@ -47,6 +47,8 @@ Once you have setup `sphinx`, you can build the documentation by running the fol
 make html
 ```

+A folder called ``_build/html`` should have been created. You can now open the file ``_build/html/index.html`` in your browser. 
+
 ---
 **NOTE**

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.6.0'
+release = u'2.7.0'


 # -- General configuration ---------------------------------------------------
@@ -103,3 +103,4 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
    model_doc/xlmroberta
    model_doc/flaubert
    model_doc/bart
+    model_doc/t5
@@ -0,0 +1,101 @@
+T5
+----------------------------------------------------
+**DISCLAIMER:** This model is still a work in progress, if you see something strange,
+file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`_
+
+Overview
+~~~~~
+The T5 model was presented in `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer <https://arxiv.org/pdf/1910.10683.pdf>`_ by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu in 
+Here the abstract: 
+
+*Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. 
+In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. 
+Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. 
+By combining the insights from our exploration with scale and our new "Colossal Clean Crawled Corpus", we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. 
+To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.*
+
+The Authors' code can be found `here <https://github.com/google-research/text-to-text-transfer-transformer>`_ .
+
+Training
+~~~~~~~~~~~~~~~~~~~~
+T5 is an encoder-decoder model and converts all NLP problems into a text-to-text format. It is trained using teacher forcing.
+This means that for training we always need an input sequence and a target sequence. 
+The input sequence is fed to the model using ``input_ids``. The target sequence is shifted to the right, *i.e.* perprended by a start-sequence token and fed to the decoder using the `decoder_input_ids`. In teacher-forcing style, the target sequence is then appended by the EOS token and corresponds to the ``lm_labels``. The PAD token is hereby used as the start-sequence token.
+T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.
+
+- Unsupervised denoising training
+  In this setup spans of the input sequence are masked by so-called sentinel tokens (*a.k.a* unique mask tokens) 
+  and the output sequence is formed as a concatenation of the same sentinel tokens and the *real* masked tokens. 
+  Each sentinel tokens represents a unique mask token for this sentence and should start with ``<extra_id_1>``, ``<extrac_id_2>``, ... up to ``<extra_id_100>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
+  *E.g.* the sentence "The cute dog walks in the park" with the masks put on "cute dog" and "the" should be processed as follows: 
+
+::
+
+  input_ids = tokenizer.encode('The <extra_id_1> walks in <extra_id_2> park', return_tensors='pt')
+  lm_labels = tokenizer.encode('<extra_id_1> cute dog <extra_id_2> the <extra_id_3> </s>', return_tensors='pt')
+  # the forward function automatically creates the correct decoder_input_ids
+  model(input_ids=input_ids, lm_labels=lm_labels)
+
+- Supervised training
+  In this setup the input sequence and output sequence are standard sequence to sequence input output mapping.
+  In translation, *e.g.* the input sequence "The house is wonderful." and output sequence "Das Haus ist wunderbar." should 
+  be processed as follows:
+  
+::
+
+  input_ids = tokenizer.encode('translate English to German: The house is wonderful. </s>', return_tensors='pt')
+  lm_labels = tokenizer.encode('Das Haus ist wunderbar. </s>', return_tensors='pt')
+  # the forward function automatically creates the correct decoder_input_ids
+  model(input_ids=input_ids, lm_labels=lm_labels)
+
+Tips
+~~~~~~~~~~~~~~~~~~~~
+- T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised 
+  and supervised tasks and for which each task is converted into a text-to-text format.
+  T5 works well on a variety of tasks out-of-the-box by prepending a different prefix to the input corresponding to each task, e.g.: for translation: *translate English to German: ..., summarize: ...*.
+  For more information about which prefix to use, it is easiest to look into Appendix D of the `paper <https://arxiv.org/pdf/1910.10683.pdf>`_ .
+- For sequence to sequence generation, it is recommended to use ``T5ForConditionalGeneration.generate()``. The method takes care of feeding the encoded input via cross-attention layers to the decoder and auto-regressively generates the decoder output.
+- T5 uses relative scalar embeddings. Encoder input padding can be done on the left and on the right.
+
+
+T5Config
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Config
+    :members:
+
+
+T5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Tokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+T5Model
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5Model
+    :members:
+
+
+T5ForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5ForConditionalGeneration
+    :members:
+
+
+TFT5Model
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFT5Model
+    :members:
+
+
+TFT5ForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFT5ForConditionalGeneration
+    :members:
@@ -275,7 +275,6 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | | FlauBERT large architecture                                                                                                         |
 |                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Bart              | ``bart-large``                                             | | 12-layer, 1024-hidden, 16-heads, 406M parameters                                                                                    |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_)                                                       |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -285,6 +284,3 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   | ``bart-large-cnn``                                         | | 12-layer, 1024-hidden, 16-heads, 406M parameters       (same as base)                                                               |
 |                   |                                                            | | bart-large base architecture finetuned on cnn summarization task                                                                    |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-
-
-.. <https://huggingface.co/transformers/examples.html>`__
@@ -420,7 +420,7 @@ to generate the tokens following the initial sequence in PyTorch, and creating a
    sequence = f"Hugging Face is based in DUMBO, New York City, and is"

    input = tokenizer.encode(sequence, return_tensors="pt")
-    generated = model.generate(input, max_length=50)
+    generated = model.generate(input, max_length=50, do_sample=True)

    resulting_string = tokenizer.decode(generated.tolist()[0])
    print(resulting_string)
@@ -432,14 +432,10 @@ to generate the tokens following the initial sequence in PyTorch, and creating a
    model = TFAutoModelWithLMHead.from_pretrained("gpt2")

    sequence = f"Hugging Face is based in DUMBO, New York City, and is"
-    generated = tokenizer.encode(sequence)
+    input = tokenizer.encode(sequence, return_tensors="tf")
+    generated = model.generate(input, max_length=50, do_sample=True)

-    for i in range(50):
-        predictions = model(tf.constant([generated]))[0]
-        token = tf.argmax(predictions[0], axis=1)[-1].numpy()
-        generated += [token]
-
-    resulting_string = tokenizer.decode(generated)
+    resulting_string = tokenizer.decode(generated.tolist()[0])
    print(resulting_string)


@@ -594,4 +590,138 @@ following array should be the output:

 ::

-    [('[CLS]', 'O'), ('Hu', 'I-ORG'), ('##gging', 'I-ORG'), ('Face', 'I-ORG'), ('Inc', 'I-ORG'), ('.', 'O'), ('is', 'O'), ('a', 'O'), ('company', 'O'), ('based', 'O'), ('in', 'O'), ('New', 'I-LOC'), ('York', 'I-LOC'), ('City', 'I-LOC'), ('.', 'O'), ('Its', 'O'), ('headquarters', 'O'), ('are', 'O'), ('in', 'O'), ('D', 'I-LOC'), ('##UM', 'I-LOC'), ('##BO', 'I-LOC'), (',', 'O'), ('therefore', 'O'), ('very', 'O'), ('##c', 'O'), ('##lose', 'O'), ('to', 'O'), ('the', 'O'), ('Manhattan', 'I-LOC'), ('Bridge', 'I-LOC'), ('.', 'O'), ('[SEP]', 'O')]
+    [('[CLS]', 'O'), ('Hu', 'I-ORG'), ('##gging', 'I-ORG'), ('Face', 'I-ORG'), ('Inc', 'I-ORG'), ('.', 'O'), ('is', 'O'), ('a', 'O'), ('company', 'O'), ('based', 'O'), ('in', 'O'), ('New', 'I-LOC'), ('York', 'I-LOC'), ('City', 'I-LOC'), ('.', 'O'), ('Its', 'O'), ('headquarters', 'O'), ('are', 'O'), ('in', 'O'), ('D', 'I-LOC'), ('##UM', 'I-LOC'), ('##BO', 'I-LOC'), (',', 'O'), ('therefore', 'O'), ('very', 'O'), ('##c', 'O'), ('##lose', 'O'), ('to', 'O'), ('the', 'O'), ('Manhattan', 'I-LOC'), ('Bridge', 'I-LOC'), ('.', 'O'), ('[SEP]', 'O')]   
+Summarization
+----------------------------------------------------
+
+Summarization is the task of summarizing a text / an article into a shorter text.
+
+An example of a summarization dataset is the CNN / Daily Mail dataset, which consists of long news articles and was created for the task of summarization.
+If you would like to fine-tune a model on a summarization task, you may leverage the ``examples/summarization/bart/run_train.sh`` (leveraging pytorch-lightning) script.
+
+Here is an example using the pipelines do to summarization. 
+It leverages a Bart model that was fine-tuned on the CNN / Daily Mail data set.
+
+::
+
+    from transformers import pipeline
+
+    summarizer = pipeline("summarization")
+
+    ARTICLE = """ New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. 
+    A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband. 
+    Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. 
+    In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. 
+    Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 
+    2010 marriage license application, according to court documents. 
+    Prosecutors said the marriages were part of an immigration scam. 
+    On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. 
+    After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective 
+    Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002. 
+    All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. 
+    Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages. 
+    Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. 
+    The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s 
+    Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. 
+    Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. 
+    If convicted, Barrientos faces up to four years in prison.  Her next court appearance is scheduled for May 18.
+    """
+    
+    print(summarizer(ARTICLE, max_length=130, min_length=30))
+
+Because the summarization pipeline depends on the ``PretrainedModel.generate()`` method, we can override the default arguments 
+of ``PretrainedModel.generate()`` directly in the pipeline as is shown for ``max_length`` and ``min_length`` above.
+This outputs the following summary:
+
+::
+
+  Liana Barrientos has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. She pleaded not guilty at State Supreme Court in the Bronx on Friday.
+  
+Here is an example doing summarization using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. Summarization is usually done using an encoder-decoder model, such as ``Bart`` or ``T5``.
+- Define the article that should be summarizaed.
+- Leverage the ``PretrainedModel.generate()`` method.
+- Add the T5 specific prefix "summarize: ".
+
+Here Google`s T5 model is used that was only pre-trained on a multi-task mixed data set (including CNN / Daily Mail), but nevertheless yields very good results.
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+
+    model = AutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    # T5 uses a max_length of 512 so we cut the article to 512 tokens.
+    inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="pt", max_length=512)
+    outputs = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
+    print(outputs)
+    
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+
+    model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    # T5 uses a max_length of 512 so we cut the article to 512 tokens.
+    inputs = tokenizer.encode("summarize: " + ARTICLE, return_tensors="tf", max_length=512)
+    outputs = model.generate(inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4, early_stopping=True)
+    print(outputs)  
+Translation
+----------------------------------------------------
+
+Translation is the task of translating a text from one language to another.
+
+An example of a translation dataset is the WMT English to German dataset, which has English sentences as the input data 
+and German sentences as the target data.
+
+Here is an example using the pipelines do to translation. 
+It leverages a T5 model that was only pre-trained on a multi-task mixture dataset (including WMT), but yields impressive 
+translation results nevertheless.
+
+::
+
+    from transformers import pipeline
+
+    translator = pipeline("translation_en_to_de")
+    print(translator("Hugging Face is a technology company based in New York and Paris", max_length=40))
+
+Because the translation pipeline depends on the ``PretrainedModel.generate()`` method, we can override the default arguments 
+of ``PretrainedModel.generate()`` directly in the pipeline as is shown for ``max_length`` above.
+This outputs the following translation into German:
+
+::
+
+  Hugging Face ist ein Technologieunternehmen mit Sitz in New York und Paris.
+  
+Here is an example doing translation using a model and a tokenizer. The process is the following:
+
+- Instantiate a tokenizer and a model from the checkpoint name. Summarization is usually done using an encoder-decoder model, such as ``Bart`` or ``T5``.
+- Define the article that should be summarizaed.
+- Leverage the ``PretrainedModel.generate()`` method.
+- Add the T5 specific prefix "translate English to German: "
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoModelWithLMHead, AutoTokenizer
+
+    model = AutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="pt")
+    outputs = model.generate(inputs, max_length=40, num_beams=4, early_stopping=True)
+
+    print(outputs)
+    
+    ## TENSORFLOW CODE
+    from transformers import TFAutoModelWithLMHead, AutoTokenizer
+
+    model = TFAutoModelWithLMHead.from_pretrained("t5-base")
+    tokenizer = AutoTokenizer.from_pretrained("t5-base")
+
+    inputs = tokenizer.encode("translate English to German: Hugging Face is a technology company based in New York and Paris", return_tensors="tf")
+    outputs = model.generate(inputs, max_length=40, num_beams=4, early_stopping=True)
+
+    print(outputs)
@@ -68,7 +68,7 @@ class GLUETransformer(BaseTransformer):
                    output_mode=args.glue_output_mode,
                    pad_on_left=bool(args.model_type in ["xlnet"]),  # pad on the left for xlnet
                    pad_token=self.tokenizer.convert_tokens_to_ids([self.tokenizer.pad_token])[0],
-                    pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+                    pad_token_segment_id=self.tokenizer.pad_token_type_id,
                )
                logger.info("Saving features into cached file %s", cached_features_file)
                torch.save(features, cached_features_file)
@@ -192,5 +192,5 @@ if __name__ == "__main__":
    # Optionally, predict on dev set and write to output_dir
    if args.do_predict:
        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
-        GLUETransformer.load_from_checkpoint(checkpoints[-1])
+        model = model.load_from_checkpoint(checkpoints[-1])
        trainer.test(model)
@@ -342,8 +342,8 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
            max_length=args.max_seq_length,
            output_mode=output_mode,
            pad_on_left=bool(args.model_type in ["xlnet"]),  # pad on the left for xlnet
-            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-            pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+            pad_token=tokenizer.pad_token_id,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
        )
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
@@ -348,8 +348,8 @@ def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
            # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
            pad_on_left=bool(args.model_type in ["xlnet"]),
            # pad on the left for xlnet
-            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-            pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+            pad_token=tokenizer.pad_token_id,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
            pad_token_label_id=pad_token_label_id,
        )
        if args.local_rank in [-1, 0]:
@@ -64,8 +64,8 @@ class NERTransformer(BaseTransformer):
                    sep_token=self.tokenizer.sep_token,
                    sep_token_extra=bool(args.model_type in ["roberta"]),
                    pad_on_left=bool(args.model_type in ["xlnet"]),
-                    pad_token=self.tokenizer.convert_tokens_to_ids([self.tokenizer.pad_token])[0],
-                    pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+                    pad_token=self.tokenizer.pad_token_id,
+                    pad_token_segment_id=self.tokenizer.pad_token_type_id,
                    pad_token_label_id=self.pad_token_label_id,
                )
                logger.info("Saving features into cached file %s", cached_features_file)
@@ -192,5 +192,5 @@ if __name__ == "__main__":
        # https://github.com/PyTorchLightning/pytorch-lightning/blob/master\
        # /pytorch_lightning/callbacks/model_checkpoint.py#L169
        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
-        NERTransformer.load_from_checkpoint(checkpoints[-1])
+        model = model.load_from_checkpoint(checkpoints[-1])
        trainer.test(model)
@@ -157,7 +157,9 @@ def train(
    writer = tf.summary.create_file_writer("/tmp/mylogs")

    with strategy.scope():
-        loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
+        loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(
+            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
+        )
        optimizer = create_optimizer(args["learning_rate"], num_train_steps, args["warmup_steps"])

        if args["fp16"]:
@@ -205,11 +207,9 @@ def train(

            with tf.GradientTape() as tape:
                logits = model(train_features["input_ids"], **inputs)[0]
-                logits = tf.reshape(logits, (-1, len(labels) + 1))
-                active_loss = tf.reshape(train_features["input_mask"], (-1,))
-                active_logits = tf.boolean_mask(logits, active_loss)
-                train_labels = tf.reshape(train_labels, (-1,))
-                active_labels = tf.boolean_mask(train_labels, active_loss)
+                active_loss = tf.reshape(train_labels, (-1,)) != pad_token_label_id
+                active_logits = tf.boolean_mask(tf.reshape(logits, (-1, len(labels))), active_loss)
+                active_labels = tf.boolean_mask(tf.reshape(train_labels, (-1,)), active_loss)
                cross_entropy = loss_fct(active_labels, active_logits)
                loss = tf.reduce_sum(cross_entropy) * (1.0 / train_batch_size)
                grads = tape.gradient(loss, model.trainable_variables)
@@ -329,11 +329,9 @@ def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode)

        with strategy.scope():
            logits = model(eval_features["input_ids"], **inputs)[0]
-            tmp_logits = tf.reshape(logits, (-1, len(labels) + 1))
-            active_loss = tf.reshape(eval_features["input_mask"], (-1,))
-            active_logits = tf.boolean_mask(tmp_logits, active_loss)
-            tmp_eval_labels = tf.reshape(eval_labels, (-1,))
-            active_labels = tf.boolean_mask(tmp_eval_labels, active_loss)
+            active_loss = tf.reshape(eval_labels, (-1,)) != pad_token_label_id
+            active_logits = tf.boolean_mask(tf.reshape(logits, (-1, len(labels))), active_loss)
+            active_labels = tf.boolean_mask(tf.reshape(eval_labels, (-1,)), active_loss)
            cross_entropy = loss_fct(active_labels, active_logits)
            loss += tf.reduce_sum(cross_entropy) * (1.0 / eval_batch_size)

@@ -436,8 +434,8 @@ def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, batch_s
            # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
            pad_on_left=bool(args["model_type"] in ["xlnet"]),
            # pad on the left for xlnet
-            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-            pad_token_segment_id=4 if args["model_type"] in ["xlnet"] else 0,
+            pad_token=tokenizer.pad_token_id,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
            pad_token_label_id=pad_token_label_id,
        )
        logging.info("Saving features into cached file %s", cached_features_file)
@@ -497,8 +495,8 @@ def main(_):
    )

    labels = get_labels(args["labels"])
-    num_labels = len(labels) + 1
-    pad_token_label_id = 0
+    num_labels = len(labels)
+    pad_token_label_id = -1
    config = AutoConfig.from_pretrained(
        args["config_name"] if args["config_name"] else args["model_name_or_path"],
        num_labels=num_labels,
@@ -522,7 +520,6 @@ def main(_):
                config=config,
                cache_dir=args["cache_dir"] if args["cache_dir"] else None,
            )
-            model.layers[-1].activation = tf.keras.activations.softmax

        train_batch_size = args["per_device_train_batch_size"] * args["n_device"]
        train_dataset, num_train_examples = load_and_cache_examples(
@@ -112,12 +112,15 @@ def convert_examples_to_features(
        label_ids = []
        for word, label in zip(example.words, example.labels):
            word_tokens = tokenizer.tokenize(word)
-            tokens.extend(word_tokens)
-            # Use the real label id for the first token of the word, and padding ids for the remaining tokens
-            label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
+
+            # bert-base-multilingual-cased sometimes output "nothing ([]) when calling tokenize with just a space.
+            if len(word_tokens) > 0:
+                tokens.extend(word_tokens)
+                # Use the real label id for the first token of the word, and padding ids for the remaining tokens
+                label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))

        # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
-        special_tokens_count = 3 if sep_token_extra else 2
+        special_tokens_count = tokenizer.num_added_tokens()
        if len(tokens) > max_seq_length - special_tokens_count:
            tokens = tokens[: (max_seq_length - special_tokens_count)]
            label_ids = label_ids[: (max_seq_length - special_tokens_count)]
@@ -360,8 +360,8 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
            max_length=args.max_seq_length,
            output_mode=output_mode,
            pad_on_left=bool(args.model_type in ["xlnet"]),  # pad on the left for xlnet
-            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-            pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+            pad_token=tokenizer.pad_token_id,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
        )
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
@@ -624,6 +624,7 @@ def main():
        and os.listdir(args.output_dir)
        and args.do_train
        and not args.overwrite_output_dir
+        and not args.should_continue
    ):
        raise ValueError(
            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
@@ -361,7 +361,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
            args.max_seq_length,
            tokenizer,
            pad_on_left=bool(args.model_type in ["xlnet"]),  # pad on the left for xlnet
-            pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
        )
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
@@ -350,8 +350,8 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
            max_length=args.max_seq_length,
            output_mode=output_mode,
            pad_on_left=False,
-            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-            pad_token_segment_id=0,
+            pad_token=tokenizer.pad_token_id,
+            pad_token_segment_id=tokenizer.pad_token_type_id,
        )
        if args.local_rank in [-1, 0]:
            logger.info("Saving features into cached file %s", cached_features_file)
@@ -1,13 +1,15 @@
-### Get the CNN Data
+### Get Preprocessed CNN Data
 To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:

 ```bash
-tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/cnn_dm.tgz
+tar -xzvf cnn_dm.tgz
 ```
+
 this should make a directory called cnn_dm/ with files like `test.source`. 
 To use your own data, copy that files format. Each article to be summarized is on its own line.

-### Usage
+### Evaluation
 To create summaries for each article in dataset, run:
 ```bash
 python evaluate_cnn.py <path_to_test.source> cnn_test_summaries.txt
@@ -16,21 +18,12 @@ the default batch size, 8, fits in 16GB GPU memory, but may need to be adjusted


 ### Training
-
-
-
-After downloading the CNN and Daily Mail datasets, preprocess the dataset:
-```commandline
-git clone https://github.com/artmatsak/cnn-dailymail
-cd cnn-dailymail && python make_datafiles.py ../cnn/stories/ ../dailymail/stories/
-```
-
-Run the training script: `run_train.sh`
-
+Run/modify `run_train.sh`
+    
 ### Where is the code?
 The core model is in `src/transformers/modeling_bart.py`. This directory only contains examples.

-### (WIP) Rouge Scores
+## (WIP) Rouge Scores

 ### Stanford CoreNLP Setup
 ```
@@ -16,15 +16,17 @@ def chunks(lst, n):
        yield lst[i : i + n]


-def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
+def generate_summaries(
+    examples: list, out_file: str, model_name: str, batch_size: int = 8, device: str = DEFAULT_DEVICE
+):
    fout = Path(out_file).open("w")
-    model = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(device)
+    model = BartForConditionalGeneration.from_pretrained(model_name, output_past=True,).to(device)
    tokenizer = BartTokenizer.from_pretrained("bart-large")

    max_length = 140
    min_length = 55

-    for batch in tqdm(list(chunks(lns, batch_size))):
+    for batch in tqdm(list(chunks(examples, batch_size))):
        dct = tokenizer.batch_encode_plus(batch, max_length=1024, return_tensors="pt", pad_to_max_length=True)
        summaries = model.generate(
            input_ids=dct["input_ids"].to(device),
@@ -43,7 +45,7 @@ def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
            fout.flush()


-def _run_generate():
+def run_generate():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "source_path", type=str, help="like cnn_dm/test.source",
@@ -51,6 +53,9 @@ def _run_generate():
    parser.add_argument(
        "output_path", type=str, help="where to save summaries",
    )
+    parser.add_argument(
+        "model_name", type=str, default="bart-large-cnn", help="like bart-large-cnn",
+    )
    parser.add_argument(
        "--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.",
    )
@@ -58,9 +63,9 @@ def _run_generate():
        "--bs", type=int, default=8, required=False, help="batch size: how many to summarize at a time",
    )
    args = parser.parse_args()
-    lns = [" " + x.rstrip() for x in open(args.source_path).readlines()]
-    generate_summaries(lns, args.output_path, batch_size=args.bs, device=args.device)
+    examples = [" " + x.rstrip() for x in open(args.source_path).readlines()]
+    generate_summaries(examples, args.output_path, args.model_name, batch_size=args.bs, device=args.device)


 if __name__ == "__main__":
-    _run_generate()
+    run_generate()
@@ -5,7 +5,7 @@ import unittest
 from pathlib import Path
 from unittest.mock import patch

-from .evaluate_cnn import _run_generate
+from .evaluate_cnn import run_generate


 articles = [" New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]
@@ -19,10 +19,14 @@ class TestBartExamples(unittest.TestCase):
    def test_bart_cnn_cli(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
-        tmp = Path(tempfile.gettempdir()) / "utest_generations.hypo"
+        tmp = Path(tempfile.gettempdir()) / "utest_generations_bart_sum.hypo"
        with tmp.open("w") as f:
            f.write("\n".join(articles))
-        testargs = ["evaluate_cnn.py", str(tmp), "output.txt"]
+
+        output_file_name = Path(tempfile.gettempdir()) / "utest_output_bart_sum.hypo"
+
+        testargs = ["evaluate_cnn.py", str(tmp), str(output_file_name), "sshleifer/bart-tiny-random"]
+
        with patch.object(sys, "argv", testargs):
-            _run_generate()
-            self.assertTrue(Path("output.txt").exists())
+            run_generate()
+            self.assertTrue(Path(output_file_name).exists())
@@ -1,4 +1,4 @@
-***This script evaluates the the multitask pre-trained checkpoint for ``t5-large`` (see paper [here](https://arxiv.org/pdf/1910.10683.pdf)) on the CNN/Daily Mail test dataset. Please note that the results in the paper were attained using a model fine-tuned on summarization, so that results will be worse here by approx. 0.5 ROUGE points***
+***This script evaluates the the multitask pre-trained checkpoint for ``t5-base`` (see paper [here](https://arxiv.org/pdf/1910.10683.pdf)) on the CNN/Daily Mail test dataset. Please note that the results in the paper were attained using a model fine-tuned on summarization, so that results will be worse here by approx. 0.5 ROUGE points***

 ### Get the CNN Data
 First, you need to download the CNN data. It's about ~400 MB and can be downloaded by 
@@ -14,13 +14,13 @@ def chunks(lst, n):
        yield lst[i : i + n]


-def generate_summaries(lns, output_file_path, batch_size, device):
+def generate_summaries(lns, output_file_path, model_size, batch_size, device):
    output_file = Path(output_file_path).open("w")

-    model = T5ForConditionalGeneration.from_pretrained("t5-large")
+    model = T5ForConditionalGeneration.from_pretrained(model_size)
    model.to(device)

-    tokenizer = T5Tokenizer.from_pretrained("t5-large")
+    tokenizer = T5Tokenizer.from_pretrained(model_size)

    # update config with summarization specific params
    task_specific_params = model.config.task_specific_params
@@ -61,6 +61,12 @@ def calculate_rouge(output_lns, reference_lns, score_path):

 def run_generate():
    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "model_size",
+        type=str,
+        help="T5 model size, either 't5-small', 't5-base', 't5-large', 't5-3b', 't5-11b'. Defaults to 't5-base'.",
+        default="t5-base",
+    )
    parser.add_argument(
        "input_path", type=str, help="like cnn_dm/test_articles_input.txt",
    )
@@ -83,7 +89,7 @@ def run_generate():

    source_lns = [x.rstrip() for x in open(args.input_path).readlines()]

-    generate_summaries(source_lns, args.output_path, args.batch_size, args.device)
+    generate_summaries(source_lns, args.output_path, args.model_size, args.batch_size, args.device)

    output_lns = [x.rstrip() for x in open(args.output_path).readlines()]
    reference_lns = [x.rstrip() for x in open(args.reference_path).readlines()]
@@ -8,6 +8,9 @@ from unittest.mock import patch
 from .evaluate_cnn import run_generate


+output_file_name = "output_t5_sum.txt"
+score_file_name = "score_t5_sum.txt"
+
 articles = ["New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]

 logging.basicConfig(level=logging.DEBUG)
@@ -19,11 +22,23 @@ class TestT5Examples(unittest.TestCase):
    def test_t5_cli(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
-        tmp = Path(tempfile.gettempdir()) / "utest_generations.hypo"
+        tmp = Path(tempfile.gettempdir()) / "utest_generations_t5_sum.hypo"
        with tmp.open("w") as f:
            f.write("\n".join(articles))
-        testargs = ["evaluate_cnn.py", str(tmp), "output.txt", str(tmp), "score.txt"]
+
+        output_file_name = Path(tempfile.gettempdir()) / "utest_output_t5_sum.hypo"
+        score_file_name = Path(tempfile.gettempdir()) / "utest_score_t5_sum.hypo"
+
+        testargs = [
+            "evaluate_cnn.py",
+            "patrickvonplaten/t5-tiny-random",
+            str(tmp),
+            str(output_file_name),
+            str(tmp),
+            str(score_file_name),
+        ]
+
        with patch.object(sys, "argv", testargs):
            run_generate()
-            self.assertTrue(Path("output.txt").exists())
-            self.assertTrue(Path("score.txt").exists())
+            self.assertTrue(Path(output_file_name).exists())
+            self.assertTrue(Path(score_file_name).exists())
@@ -14,13 +14,13 @@ def chunks(lst, n):
        yield lst[i : i + n]


-def generate_translations(lns, output_file_path, batch_size, device):
+def generate_translations(lns, output_file_path, model_size, batch_size, device):
    output_file = Path(output_file_path).open("w")

-    model = T5ForConditionalGeneration.from_pretrained("t5-base")
+    model = T5ForConditionalGeneration.from_pretrained(model_size)
    model.to(device)

-    tokenizer = T5Tokenizer.from_pretrained("t5-base")
+    tokenizer = T5Tokenizer.from_pretrained(model_size)

    # update config with summarization specific params
    task_specific_params = model.config.task_specific_params
@@ -52,6 +52,12 @@ def calculate_bleu_score(output_lns, refs_lns, score_path):

 def run_generate():
    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "model_size",
+        type=str,
+        help="T5 model size, either 't5-small', 't5-base', 't5-large', 't5-3b', 't5-11b'. Defaults to 't5-base'.",
+        default="t5-base",
+    )
    parser.add_argument(
        "input_path", type=str, help="like wmt/newstest2013.en",
    )
@@ -78,7 +84,7 @@ def run_generate():

    input_lns = [x.strip().replace(dash_pattern[0], dash_pattern[1]) for x in open(args.input_path).readlines()]

-    generate_translations(input_lns, args.output_path, args.batch_size, args.device)
+    generate_translations(input_lns, args.output_path, args.model_size, args.batch_size, args.device)

    output_lns = [x.strip() for x in open(args.output_path).readlines()]
    refs_lns = [x.strip().replace(dash_pattern[0], dash_pattern[1]) for x in open(args.reference_path).readlines()]
@@ -8,7 +8,11 @@ from unittest.mock import patch
 from .evaluate_wmt import run_generate


-text = [" New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]
+text = ["When Liana Barrientos was 23 years old, she got married in Westchester County."]
+translation = ["Als Liana Barrientos 23 Jahre alt war, heiratete sie in Westchester County."]
+
+output_file_name = "output_t5_trans.txt"
+score_file_name = "score_t5_trans.txt"

 logging.basicConfig(level=logging.DEBUG)

@@ -19,10 +23,28 @@ class TestT5Examples(unittest.TestCase):
    def test_t5_cli(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)
-        tmp = Path(tempfile.gettempdir()) / "utest_generations.hypo"
-        with tmp.open("w") as f:
+
+        tmp_source = Path(tempfile.gettempdir()) / "utest_generations_t5_trans.hypo"
+        with tmp_source.open("w") as f:
            f.write("\n".join(text))
-        testargs = ["evaluate_cnn.py", str(tmp), "output.txt", str(tmp), "score.txt"]
+
+        tmp_target = Path(tempfile.gettempdir()) / "utest_generations_t5_trans.target"
+        with tmp_target.open("w") as f:
+            f.write("\n".join(translation))
+
+        output_file_name = Path(tempfile.gettempdir()) / "utest_output_trans.hypo"
+        score_file_name = Path(tempfile.gettempdir()) / "utest_score.hypo"
+
+        testargs = [
+            "evaluate_wmt.py",
+            "patrickvonplaten/t5-tiny-random",
+            str(tmp_source),
+            str(output_file_name),
+            str(tmp_target),
+            str(score_file_name),
+        ]
+
        with patch.object(sys, "argv", testargs):
            run_generate()
-            self.assertTrue(Path("output.txt").exists())
+            self.assertTrue(Path(output_file_name).exists())
+            self.assertTrue(Path(score_file_name).exists())
@@ -1,5 +1,6 @@
 ---
 language: german
+license: mit
 ---

 # 🤗 + 📚 dbmdz German BERT models
@@ -1,5 +1,6 @@
 ---
 language: german
+license: mit
 tags:
  - "historic german"
 ---
@@ -1,5 +1,6 @@
 ---
 language: german
+license: mit
 tags:
  - "historic german"
 ---
@@ -1,5 +1,6 @@
 ---
 language: german
+license: mit
 ---

 # 🤗 + 📚 dbmdz German BERT models
@@ -1,5 +1,6 @@
 ---
 language: italian
+license: mit
 ---

 # 🤗 + 📚 dbmdz BERT models
@@ -1,5 +1,6 @@
 ---
 language: italian
+license: mit
 ---

 # 🤗 + 📚 dbmdz BERT models
@@ -1,5 +1,6 @@
 ---
 language: italian
+license: mit
 ---

 # 🤗 + 📚 dbmdz BERT models
@@ -1,5 +1,6 @@
 ---
 language: italian
+license: mit
 ---

 # 🤗 + 📚 dbmdz BERT models
@@ -1,5 +1,6 @@
 ---
 language: turkish
+license: mit
 ---

 # 🤗 + 📚 dbmdz Turkish BERT model
@@ -1,5 +1,6 @@
 ---
 language: turkish
+license: mit
 ---

 # 🤗 + 📚 dbmdz Turkish BERT model
@@ -1,5 +1,6 @@
 ---
 language: turkish
+license: mit
 ---

 # 🤗 + 📚 dbmdz Turkish BERT model
@@ -1,5 +1,6 @@
 ---
 language: turkish
+license: mit
 ---

 # 🤗 + 📚 dbmdz Turkish BERT model
@@ -1,5 +1,6 @@
 ---
 language: turkish
+license: mit
 ---

 # 🤗 + 📚 dbmdz Distilled Turkish BERT model
@@ -0,0 +1,59 @@
+This language model is trained using sentence_transformers (https://github.com/UKPLab/sentence-transformers)
+Started with bert-base-nli-stsb-mean-tokens
+Continue training on quora questions deduplication dataset (https://www.kaggle.com/c/quora-question-pairs)
+See train_script.py for script used
+
+Below is the performance over the course of training
+epoch,steps,cosine_pearson,cosine_spearman,euclidean_pearson,euclidean_spearman,manhattan_pearson,manhattan_spearman,dot_pearson,dot_spearman
+0,1000,0.5944576426835938,0.6010801382777033,0.5942803776859142,0.5934485776801595,0.5939676679774666,0.593162725602328,0.5905591590826669,0.5921674789994058
+0,2000,0.6404080440207146,0.6416811632113405,0.6384419354012121,0.6352050423100778,0.6379917744471867,0.6347884067391001,0.6410544760582826,0.6379252046791412
+0,3000,0.6710168301884945,0.6676529324662036,0.6660195209784969,0.6618423144808695,0.6656461098096684,0.6615366331956389,0.6724401903484759,0.666073727723655
+0,4000,0.6886373265097949,0.6808948140300153,0.67907655686838,0.6714218133850957,0.6786809551564443,0.6711577956884357,0.6926435869763303,0.68190855298609
+0,5000,0.6991409753700026,0.6919630610321864,0.6991041519437052,0.6868961486499775,0.6987076032270729,0.6865385550504007,0.7035518148330993,0.6916275246101342
+0,6000,0.7120367327025509,0.6975005265298305,0.7065567493967201,0.6922375503495235,0.7060005509843024,0.6916475765570651,0.7147094303373102,0.6981390706722722
+0,7000,0.7254672394728687,0.7130118465900485,0.7261844956277705,0.7086213543110718,0.7257479964972307,0.7079315661881832,0.728729909455115,0.7122743793160531
+0,8000,0.7402421930101399,0.7216774208330149,0.7367901914441078,0.7166256588352043,0.7362607046874481,0.7158881916281887,0.7433902441373252,0.7220998491980078
+0,9000,0.7381005358120434,0.7197216844469877,0.7343228719349923,0.7139462687943793,0.7345247569255238,0.7145106206467152,0.7421843672419275,0.720686853053079
+0,10000,0.7465436564646095,0.7260327107480364,0.7467524239596304,0.7230195666847953,0.7467721566237211,0.7231367593302213,0.749792199122442,0.7263143296580317
+0,11000,0.7521805421706547,0.7323771570146701,0.7530672061250105,0.729223203496722,0.7530616532823367,0.7293818369675622,0.7552399002305836,0.7320808333541338
+0,12000,0.7579359969644401,0.7340677616737238,0.7570017235719905,0.7305965412825544,0.7570601853520393,0.730718189957289,0.7611254136080384,0.7351501229591327
+0,-1,0.7573407371218097,0.7329952035782198,0.755595312163209,0.7291445551777086,0.7557737117990928,0.7295404703700227,0.7607276219361719,0.7342415455980179
+1,1000,0.7619907683805341,0.7374667949734767,0.7629820517114324,0.7330364216044966,0.7628369522755882,0.7331912674450544,0.7658583898073758,0.7381503446695727
+1,2000,0.7618972640071228,0.7362151058969478,0.764582212425539,0.7335856230046062,0.7643125513700815,0.7334501607097152,0.7652852805583232,0.7369104639809163
+1,3000,0.7687362955240467,0.7404674623181671,0.7708304819979073,0.7380959815601529,0.7707835692712482,0.7379796800453193,0.772074854759756,0.7414513460702766
+1,4000,0.7685047787908202,0.7403088288815168,0.7703522257474043,0.7379787888808298,0.7701221475099808,0.7377898546753812,0.7713755359045312,0.7409415801952219
+1,5000,0.7696438109797803,0.7410393893292365,0.773270389327895,0.7392953127251652,0.7729880866533291,0.7389853982789335,0.7726236305835863,0.7416278035580925
+1,6000,0.7749538363837081,0.7436499342062207,0.774879168058157,0.7401827241766746,0.7745754601165837,0.739763415043146,0.7788801166152383,0.7446249060022169
+1,7000,0.7794560817870597,0.7480970176267153,0.7803506944510302,0.7453305130502859,0.7799867949176531,0.7447100155494814,0.7828208193123926,0.7486740690324809
+1,8000,0.7855844359073243,0.7496742172376921,0.7828816645965887,0.747176409009761,0.7827584875358967,0.7471037762845532,0.7879159073496309,0.7507349669102151
+1,9000,0.7844110753729492,0.7507746252693759,0.7847208586489722,0.7485172180290892,0.7846408087474059,0.748491818820158,0.7872061334510225,0.7514470349769437
+1,10000,0.7881311227435004,0.7530048509727403,0.7886917756879734,0.7508018068765787,0.7883332502188707,0.7505037008187275,0.7910707228932787,0.7537200382362567
+1,11000,0.7883300109606874,0.7513494487126553,0.7879329130497712,0.749818368689255,0.7876525616593218,0.7494872882301785,0.7911454269743292,0.7522843165147303
+1,12000,0.7853334933336618,0.7516809747712728,0.7893895316714998,0.749780492728257,0.7890075986655403,0.7494079715118533,0.7885959664070629,0.7523827940133203
+1,-1,0.7887529238148887,0.7534076729932393,0.7896864404801204,0.7513080079201105,0.7894077512343298,0.7510009899066772,0.7919617393746149,0.7542173273241598
+2,1000,0.7919209063905188,0.7550167329363414,0.7917464066515253,0.7523043685293455,0.7914371703225378,0.7520285423781206,0.7950297421784158,0.7562599556207076
+2,2000,0.7924507768792486,0.7542908512484463,0.7934519001953887,0.7517491515010692,0.7931885648751081,0.751521004535999,0.7951637852162545,0.7551495215642072
+2,3000,0.7937606244038364,0.755599577136169,0.7933633347508111,0.7527922999916203,0.7931581019714242,0.7527132061436363,0.797275652800117,0.7569827180764233
+2,4000,0.7938389298721445,0.7578716892320315,0.7963783770097079,0.7555928931784702,0.796150381773947,0.7555438771581088,0.7972911620482322,0.759178632650707
+2,5000,0.7935330563129844,0.7551129824372304,0.7970775059297484,0.7527285792572385,0.7967359830546507,0.7524478515463257,0.7966395126138969,0.756319220359678
+2,6000,0.7929852776759999,0.7525490026774382,0.7952484474454824,0.7503695753216607,0.7950784132079611,0.7503677929234961,0.7956152082976395,0.7535275392698093
+2,7000,0.794956504054517,0.756119591765251,0.7982025041673655,0.7532521587180684,0.7980261618830962,0.7532107179960499,0.7983222918908033,0.7571226363678287
+2,8000,0.7934568432535339,0.7538336661192452,0.797015698241178,0.7514773358161916,0.7968076980315735,0.7513458838811067,0.7960694134685949,0.754143803399873
+2,9000,0.7970040626682157,0.7576497805894974,0.7987855332059015,0.7550996144509958,0.7984693921009676,0.7548260162973456,0.7999509314900626,0.758347143906916
+2,10000,0.7979442987735523,0.7585338500791028,0.8018677081664496,0.7557412777548302,0.8015397301245205,0.7552916678886369,0.8007921348414564,0.7589772216225288
+2,11000,0.7985519561040211,0.7579986850302035,0.8021236875460913,0.7555826443181872,0.8019861620475348,0.7553763317660516,0.8009230128897853,0.7586541619907702
+2,12000,0.7986842143860736,0.7599570950134775,0.8029131054823838,0.7577678644678973,0.8027922603736795,0.7575152095990927,0.8020896747930555,0.7608540869254408
+2,-1,0.7994135319568432,0.7596286881516635,0.8022087183675333,0.7570593611974978,0.8020218401019292,0.7567291719729909,0.8026346812258125,0.7603928913647044
+3,1000,0.7985505039929134,0.7592588405681144,0.8023296699449267,0.7569345933969436,0.8023622066009718,0.7570237132696928,0.8013054275981851,0.759643838536062
+3,2000,0.7995482191699455,0.759205368623176,0.8026859405513612,0.7565709841358819,0.8024845263367439,0.7562920388231202,0.8021318586127523,0.7596496313300967
+3,3000,0.7991070423195897,0.7582027696555826,0.8016352550470427,0.7555585819429662,0.8014268261947898,0.7551838327642736,0.8013136081494014,0.7584429477727118
+3,4000,0.7999188836884763,0.7586764419322649,0.802987646214278,0.7561111254802977,0.8026549791861386,0.7556463650525692,0.8024068858366156,0.7591238238715613
+3,5000,0.7988075932525881,0.7583533823004922,0.8019498750207454,0.755792967372457,0.8016459824731964,0.7553834613587099,0.8015528810821693,0.7589527136833425
+3,6000,0.8003341798460688,0.7585432077405799,0.8032464035902267,0.7563722467405277,0.8028695045742804,0.7557626665682309,0.8027937010871594,0.7590404967573696
+3,7000,0.799187592384933,0.7579358555659604,0.8028413548398412,0.7555875459131398,0.8025187078191003,0.7551196665011402,0.8018680475193432,0.7585565756912578
+3,8000,0.797725037202641,0.757439012042047,0.802048241301358,0.7548888458326453,0.8017608103042271,0.7544606246736175,0.8005479449399782,0.758037452190282
+3,9000,0.7990232649360067,0.7573703896772077,0.8021375332910405,0.754873027155089,0.8018733796679427,0.7545680141630304,0.8016400687760605,0.7579461042843499
+3,10000,0.7994934439260372,0.758368978248884,0.8035693504115055,0.75619400688862,0.8032990505007025,0.7559016935896375,0.8022819185772518,0.7589558328445544
+3,11000,0.8002954591825011,0.758710753096932,0.8043310859792212,0.7566387152306694,0.8040865016706966,0.7564221538891368,0.8030873114870971,0.7592722085543488
+3,12000,0.8003726616196549,0.7588056657991931,0.8044000317617518,0.7566146528909147,0.8041705213966136,0.7563419459362758,0.8031760015719815,0.7593194421057111
+3,-1,0.8004926728141455,0.7587192194882135,0.8043340929890026,0.756546030526114,0.8041028559910275,0.7563103085106637,0.8032542493776693,0.7592325501951863
@@ -0,0 +1,38 @@
+# CovidBERT-NLI
+
+This is the model **CovidBERT** trained by DeepSet on AllenAI's [CORD19 Dataset](https://pages.semanticscholar.org/coronavirus-research) of scientific articles about coronaviruses.
+
+The model uses the original BERT wordpiece vocabulary and was subsequently fine-tuned on the [SNLI](https://nlp.stanford.edu/projects/snli/) and the [MultiNLI](https://www.nyu.edu/projects/bowman/multinli/) datasets using the [`sentence-transformers` library](https://github.com/UKPLab/sentence-transformers/) to produce universal sentence embeddings [1] using the **average pooling strategy** and a **softmax loss**.
+
+Parameter details for the original training on CORD-19 are available on [DeepSet's MLFlow](https://public-mlflow.deepset.ai/#/experiments/2/runs/ba27d00c30044ef6a33b1d307b4a6cba)
+
+**Base model**: `deepset/covid_bert_base` from HuggingFace's `AutoModel`.
+
+**Training time**: ~6 hours on the NVIDIA Tesla P100 GPU provided in Kaggle Notebooks.
+
+**Parameters**:
+
+| Parameter        | Value |
+|------------------|-------|
+| Batch size       | 64    |
+| Training steps   | 23000 |
+| Warmup steps     | 1450  |
+| Lowercasing      | True  |
+| Max. Seq. Length | 128   |
+
+**Performances**: The performance was evaluated on the test portion of the [STS dataset](http://ixa2.si.ehu.es/stswiki/index.php/STSbenchmark) using Spearman rank correlation and compared to the performances of similar models obtained with the same procedure to verify its performances.
+
+| Model                         | Score       |
+|-------------------------------|-------------|
+| `covidbert-nli` (this)        | 67.52       |
+| `gsarti/biobert-nli`          | 73.40       |
+| `gsarti/scibert-nli`          | 74.50       |
+| `bert-base-nli-mean-tokens`[2]| 77.12       |
+
+An example usage for similarity-based scientific paper retrieval is provided in the [Covid-19 Semantic Browser](https://github.com/gsarti/covid-papers-browser) repository.
+
+**References:**
+
+[1] A. Conneau et al., [Supervised Learning of Universal Sentence Representations from Natural Language Inference Data](https://www.aclweb.org/anthology/D17-1070/)
+
+[2] N. Reimers et I. Gurevych, [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://www.aclweb.org/anthology/D19-1410/)
@@ -0,0 +1,92 @@
+---
+language: malay
+---
+
+# Bahasa Tiny-BERT Model
+
+General Distilled Tiny BERT base language model for Malay and Indonesian. 
+
+## Pretraining Corpus
+
+`tiny-bert-bahasa-cased` model was distilled on ~1.8 Billion words. We distilled on both standard and social media language structures, and below is list of data we distilled on,
+
+1. [dumping wikipedia](https://github.com/huseinzol05/Malaya-Dataset#wikipedia-1).
+2. [local instagram](https://github.com/huseinzol05/Malaya-Dataset#instagram).
+3. [local twitter](https://github.com/huseinzol05/Malaya-Dataset#twitter-1).
+4. [local news](https://github.com/huseinzol05/Malaya-Dataset#public-news).
+5. [local parliament text](https://github.com/huseinzol05/Malaya-Dataset#parliament).
+6. [local singlish/manglish text](https://github.com/huseinzol05/Malaya-Dataset#singlish-text).
+7. [IIUM Confession](https://github.com/huseinzol05/Malaya-Dataset#iium-confession).
+8. [Wattpad](https://github.com/huseinzol05/Malaya-Dataset#wattpad).
+9. [Academia PDF](https://github.com/huseinzol05/Malaya-Dataset#academia-pdf).
+
+Preprocessing steps can reproduce from here, [Malaya/pretrained-model/preprocess](https://github.com/huseinzol05/Malaya/tree/master/pretrained-model/preprocess).
+
+## Distilling details
+
+- This model was distilled using huawei-noah Tiny-BERT's github [repository](https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/TinyBERT) on 3 Titan V100 32GB VRAM.
+- All steps can reproduce from here, [Malaya/pretrained-model/tiny-bert](https://github.com/huseinzol05/Malaya/tree/master/pretrained-model/tiny-bert).
+
+## Load Distilled Model
+
+You can use this model by installing `torch` or `tensorflow` and Huggingface library `transformers`. And you can use it directly by initializing it like this:  
+
+```python
+from transformers import AlbertTokenizer, BertModel
+
+model = BertModel.from_pretrained('huseinzol05/tiny-bert-bahasa-cased')
+tokenizer = AlbertTokenizer.from_pretrained(
+    'huseinzol05/tiny-base-bahasa-cased',
+    unk_token = '[UNK]',
+    pad_token = '[PAD]',
+    do_lower_case = False,
+)
+```
+
+We use [google/sentencepiece](https://github.com/google/sentencepiece) to train the tokenizer, so to use it, need to load from `AlbertTokenizer`.
+
+## Example using AutoModelWithLMHead
+
+```python
+from transformers import AlbertTokenizer, AutoModelWithLMHead, pipeline
+
+model = AutoModelWithLMHead.from_pretrained('huseinzol05/tiny-base-bahasa-cased')
+tokenizer = AlbertTokenizer.from_pretrained(
+    'huseinzol05/tiny-base-bahasa-cased',
+    unk_token = '[UNK]',
+    pad_token = '[PAD]',
+    do_lower_case = False,
+)
+fill_mask = pipeline('fill-mask', model = model, tokenizer = tokenizer)
+print(fill_mask('makan ayam dengan [MASK]'))
+```
+
+Output is,
+
+```text
+[{'sequence': '[CLS] makan ayam dengan berbual[SEP]',
+  'score': 0.00015769545279908925,
+  'token': 17859},
+ {'sequence': '[CLS] makan ayam dengan kembar[SEP]',
+  'score': 0.0001448775001335889,
+  'token': 8289},
+ {'sequence': '[CLS] makan ayam dengan memaklumkan[SEP]',
+  'score': 0.00013484008377417922,
+  'token': 6881},
+ {'sequence': '[CLS] makan ayam dengan Senarai[SEP]',
+  'score': 0.00013061291247140616,
+  'token': 11698},
+ {'sequence': '[CLS] makan ayam dengan Tiga[SEP]',
+  'score': 0.00012453157978598028,
+  'token': 4232}]
+```
+
+## Results
+
+For further details on the model performance, simply checkout accuracy page from Malaya, https://malaya.readthedocs.io/en/latest/Accuracy.html, we compared with traditional models.
+
+## Acknowledgement
+
+Thanks to [Im Big](https://www.facebook.com/imbigofficial/), [LigBlou](https://www.facebook.com/ligblou), [Mesolitica](https://mesolitica.com/) and [KeyReply](https://www.keyreply.com/) for sponsoring AWS, Google and GPU clouds to train BERT for Bahasa. 
+
+
@@ -50,7 +50,7 @@ tokenizer = XLNetTokenizer.from_pretrained(
    'huseinzol05/xlnet-base-bahasa-cased', do_lower_case = False
 )
 fill_mask = pipeline('fill-mask', model = model, tokenizer = tokenizer)
-print(fill_mask('makan ayam dengan [MASK]'))
+print(fill_mask('makan ayam dengan <mask>'))
 ```

 ## Results
@@ -0,0 +1,14 @@
+# BERT-IMDB
+
+## What is it?
+BERT (`bert-large-cased`) trained for sentiment classification on the [IMDB dataset](https://www.kaggle.com/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews).
+
+## Training setting
+
+The model was trained on 80% of the IMDB dataset for sentiment classification for three epochs with a learning rate of `1e-5` with the `simpletransformers` library. The library uses a learning rate schedule.
+
+## Result
+The model achieved 90% classification accuracy on the validation set.
+
+## Reference
+The full experiment is available in the [tlr repo](https://lvwerra.github.io/trl/03-bert-imdb-training/).
@@ -0,0 +1,18 @@
+# GPT2-IMDB-pos
+
+## What is it?
+A small GPT2 (`lvwerra/gpt2-imdb`) language model fine-tuned to produce positive movie reviews based the [IMDB dataset](https://www.kaggle.com/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews). The model is trained with rewards from a BERT sentiment classifier (`lvwerra/gpt2-imdb`) via PPO.
+
+## Training setting
+The model was trained for `100` optimisation steps with a batch size of `256` which corresponds to `25600` training samples. The full experiment setup can be found in the Jupyter notebook in the [trl repo](https://lvwerra.github.io/trl/04-gpt2-sentiment-ppo-training/).
+
+## Examples
+A few examples of the model response to a query before and after optimisation:
+
+| query | response (before) | response (after) | rewards (before) | rewards (after) |
+|-------|-------------------|------------------|------------------|-----------------|
+|I'd never seen a |heavier, woodier example of Victorian archite... |film of this caliber, and I think it's wonder... |3.297736 |4.158653|
+|I love John's work	|but I actually have to write language as in w... |and I hereby recommend this film. I am really... |-1.904006 |4.159198 |
+|I's a big struggle |to see anyone who acts in that way. by Jim Th... |, but overall I'm happy with the changes even ... |-1.595925 |2.651260|
+
+
@@ -0,0 +1,27 @@
+# GPT2-IMDB
+
+## What is it?
+A GPT2 (`gpt2`) language model fine-tuned on the [IMDB dataset](https://www.kaggle.com/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews).
+
+## Training setting
+
+The GPT2 language model was fine-tuned for 1 epoch on the IMDB dataset. All comments were joined into a single text file separated by the EOS token:
+
+```
+import pandas as pd
+df = pd.read_csv("imdb-dataset.csv")
+imdb_str = " <|endoftext|> ".join(df['review'].tolist())
+
+with open ('imdb.txt', 'w') as f:
+    f.write(imdb_str)
+```
+
+To train the model the `run_language_modeling.py` script in the `transformer` library was used:
+
+```
+python run_language_modeling.py 
+	--train_data_file imdb.txt 
+	--output_dir gpt2-imdb 
+	--model_type gpt2 
+	--model_name_or_path gpt2
+```
@@ -5,15 +5,16 @@ thumbnail:

 # GPT-2 + CORD19 dataset : 🦠 ✍ ⚕

-**GPT-2** fine-tuned on **biorxiv_medrxiv** and **comm_use_subset files** from [CORD-19](https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge) dataset.
+**GPT-2** fine-tuned on **biorxiv_medrxiv**, **comm_use_subset** and **custom_license** files from [CORD-19](https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge) dataset.


-## Datasets details:
+## Datasets details

 | Dataset                | # Files |
 | ---------------------- | ----- |
 | biorxiv_medrxiv        | 885  |
-| comm_use_subse         | 9K   |
+| comm_use_subset         | 9K   |
+| custom_license         | 20.6K   |

 ## Model training

@@ -37,7 +38,7 @@ python run_language_modeling.py \

 <img alt="training loss" src="https://svgshare.com/i/JTf.svg' title='GTP-2-finetuned-CORDS19-loss" width="600" height="300" />

-## Model in action / Example of usage: ✒
+## Model in action / Example of usage ✒

 You can get the following script [here](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py)

@@ -0,0 +1,62 @@
+---
+language: english
+thumbnail:
+---
+
+# GPT-2 + bio/medrxiv files from CORD19: 🦠 ✍ ⚕
+
+**GPT-2** fine-tuned on **biorxiv_medrxiv** files from [CORD-19](https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge) dataset.
+
+
+## Datasets details:
+
+| Dataset                | # Files |
+| ---------------------- | ----- |
+| biorxiv_medrxiv        | 885  |
+
+
+## Model training:
+
+The model was trained on a Tesla P100 GPU and 25GB of RAM with the following command:
+
+```bash
+
+export TRAIN_FILE=/path/to/dataset/train.txt
+
+python run_language_modeling.py \
+    --model_type gpt2 \
+    --model_name_or_path gpt2 \
+    --do_train \
+    --train_data_file $TRAIN_FILE \
+    --num_train_epochs 4 \
+    --output_dir model_output \
+    --overwrite_output_dir \
+    --save_steps 2000 \
+    --per_gpu_train_batch_size 3
+```
+
+## Model in action / Example of usage: ✒
+
+You can get the following script [here](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py)
+
+```bash
+python run_generation.py \
+    --model_type gpt2 \
+    --model_name_or_path mrm8488/GPT-2-finetuned-CORD19 \
+    --length 200
+```
+```txt
+👵👴🦠
+# Input: Old people with COVID-19 tends to suffer 
+# Output: === GENERATED SEQUENCE 1 ===
+Old people with COVID-19 tends to suffer more symptom onset time and death. It is well known that many people with COVID-19 have high homozygous ZIKV infection in the face of severe symptoms in both severe and severe cases.
+The origin of Wuhan Fever was investigated by Prof. Shen Jiang at the outbreak of Wuhan Fever [34]. As Huanan Province is the epicenter of this outbreak, Huanan, the epicenter of epidemic Wuhan Fever, is the most potential location for the direct transmission of infection (source: Zhongzhen et al., 2020). A negative risk ratio indicates more frequent underlying signs in the people in Huanan Province with COVID-19 patients. Further analysis of reported Huanan Fever onset data in the past two years indicated that the intensity of exposure is the key risk factor for developing MERS-CoV infection in this region, especially among children and elderly. To be continued to develop infected patients would be a very important area for
+```
+
+![Model in action](https://media.giphy.com/media/TgUdO72Iwk9h7hhm7G/giphy.gif)
+
+
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -66,6 +66,8 @@ nlp_ner = pipeline(
        {"use_fast": False}
 ))

+text = 'Mis amigos están pensando viajar a Londres este verano'
+
 nlp_ner(text)

 #Output: [{'entity': 'B-LOC', 'score': 0.9998720288276672, 'word': 'Londres'}]
@@ -0,0 +1,83 @@
+---
+language: spanish
+thumbnail:
+---
+
+# Spanish BERT (BETO) + Syntax POS tagging ✍🏷
+
+This model is a fine-tuned version of the Spanish BERT [(BETO)](https://github.com/dccuchile/beto) on Spanish **syntax** annotations in [CONLL CORPORA](https://www.kaggle.com/nltkdata/conll-corpora) dataset for **syntax POS** (Part of Speech tagging) downstream task.
+
+## Details of the downstream task (Syntax POS) - Dataset
+
+- [Dataset: CONLL Corpora ES](https://www.kaggle.com/nltkdata/conll-corpora)
+
+#### [Fine-tune script on NER dataset provided by Huggingface](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py)
+
+#### 21 Syntax annotations (Labels) covered:
+
+- \_
+- ATR
+- ATR.d
+- CAG
+- CC
+- CD
+- CD.Q
+- CI
+- CPRED
+- CPRED.CD
+- CPRED.SUJ
+- CREG
+- ET
+- IMPERS
+- MOD
+- NEG
+- PASS
+- PUNC
+- ROOT
+- SUJ
+- VOC
+
+## Metrics on test set 📋
+
+|  Metric   |  # score  |
+| :-------: | :-------: |
+|    F1     | **89.27** |
+| Precision | **89.44** |
+|  Recall   | **89.11** |
+
+## Model in action 🔨
+
+Fast usage with **pipelines** 🧪
+
+```python
+from transformers import pipeline
+
+nlp_pos_syntax = pipeline(
+    "ner",
+    model="mrm8488/bert-spanish-cased-finetuned-pos-syntax",
+    tokenizer="mrm8488/bert-spanish-cased-finetuned-pos-syntax"
+)
+
+text = 'Mis amigos están pensando viajar a Londres este verano.'
+
+nlp_pos_syntax(text)[1:len(nlp_pos_syntax(text))-1]
+```
+
+```json
+[
+  { "entity": "_", "score": 0.9999216794967651, "word": "Mis" },
+  { "entity": "SUJ", "score": 0.999882698059082, "word": "amigos" },
+  { "entity": "_", "score": 0.9998869299888611, "word": "están" },
+  { "entity": "ROOT", "score": 0.9980518221855164, "word": "pensando" },
+  { "entity": "_", "score": 0.9998420476913452, "word": "viajar" },
+  { "entity": "CD", "score": 0.999351978302002, "word": "a" },
+  { "entity": "_", "score": 0.999959409236908, "word": "Londres" },
+  { "entity": "_", "score": 0.9998968839645386, "word": "este" },
+  { "entity": "CC", "score": 0.99931401014328, "word": "verano" },
+  { "entity": "PUNC", "score": 0.9998534917831421, "word": "." }
+]
+```
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -21,7 +21,7 @@ I preprocessed the dataset and splitted it as train / dev (80/20)

 - [Fine-tune on NER script provided by Huggingface](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py)

- Labels covered:
+- **60** Labels covered:

 ```
 AO, AQ, CC, CS, DA, DD, DE, DI, DN, DP, DT, Faa, Fat, Fc, Fd, Fe, Fg, Fh, Fia, Fit, Fp, Fpa, Fpt, Fs, Ft, Fx, Fz, I, NC, NP, P0, PD, PI, PN, PP, PR, PT, PX, RG, RN, SP, VAI, VAM, VAN, VAP, VAS, VMG, VMI, VMM, VMN, VMP, VMS, VSG, VSI, VSM, VSN, VSP, VSS, Y and Z
@@ -74,6 +74,8 @@ nlp_pos(text)
 ```
 ![model in action](https://media.giphy.com/media/jVC9m1cNrdIWuAAtjy/giphy.gif)

+16 POS tags version also available [here](https://huggingface.co/mrm8488/bert-spanish-cased-finetuned-pos-16-tags)
+

 > Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)

@@ -0,0 +1,82 @@
+---
+language: multilingual
+thumbnail:
+---
+
+# DistilBERT multilingual fine-tuned on TydiQA (GoldP task) dataset for multilingual Q&A 😛🌍❓
+
+
+## Details of the language model
+
+[distilbert-base-multilingual-cased](https://huggingface.co/distilbert-base-multilingual-cased)
+
+
+## Details of the Tydi QA dataset
+
+TyDi QA contains 200k human-annotated question-answer pairs in 11 Typologically Diverse languages, written without seeing the answer and without the use of translation, and is designed for the **training and evaluation** of automatic question answering systems. This repository provides evaluation code and a baseline system for the dataset. https://ai.google.com/research/tydiqa
+
+
+## Details of the downstream task (Gold Passage or GoldP aka the secondary task)
+
+Given a passage that is guaranteed to contain the answer, predict the single contiguous span of characters that answers the question. the gold passage task differs from the [primary task](https://github.com/google-research-datasets/tydiqa/blob/master/README.md#the-tasks) in several ways:
+*   only the gold answer passage is provided rather than the entire Wikipedia article;
+*   unanswerable questions have been discarded, similar to MLQA and XQuAD;
+*   we evaluate with the SQuAD 1.1 metrics like XQuAD; and
+*   Thai and Japanese are removed since the lack of whitespace breaks some tools.
+
+
+## Model training 💪🏋️‍
+
+The model was fine-tuned on a Tesla P100 GPU and 25GB of RAM.
+The script is the following:
+
+```python
+python transformers/examples/run_squad.py \
+  --model_type distilbert \
+  --model_name_or_path distilbert-base-multilingual-cased \
+  --do_train \
+  --do_eval \
+  --train_file /path/to/dataset/train.json \
+  --predict_file /path/to/dataset/dev.json \
+  --per_gpu_train_batch_size 24 \
+  --per_gpu_eval_batch_size 24 \
+  --learning_rate 3e-5 \
+  --num_train_epochs 5 \
+  --max_seq_length 384 \
+  --doc_stride 128 \
+  --output_dir /content/model_output \
+  --overwrite_output_dir \
+  --save_steps 1000 \
+  --threads 400
+  ```
+
+## Global Results (dev set) 📝
+
+| Metric    | # Value     |
+| --------- | ----------- |
+| **EM**    | **63.85** |
+| **F1**    | **75.70** |
+
+## Specific Results (per language) 🌍📝 
+
+| Language    | # Samples     | # EM | # F1 |
+| --------- | ----------- |--------| ------ |
+| Arabic    | 1314  | 66.66 | 80.02 |
+| Bengali   | 180   | 53.09 | 63.50 |
+| English   | 654   | 62.42 | 73.12 |
+| Finnish   | 1031  | 64.57 | 75.15 |
+| Indonesian| 773   | 67.89 | 79.70 |
+| Korean    | 414   | 51.29 | 61.73 |
+| Russian   | 1079  | 55.42 | 70.08 |
+| Swahili   | 596   | 74.51 | 81.15 |
+| Telegu    | 874   | 66.21 | 79.85 |
+
+
+## Similar models
+
+You can also try [bert-multi-cased-finedtuned-xquad-tydiqa-goldp](https://huggingface.co/mrm8488/bert-multi-cased-finedtuned-xquad-tydiqa-goldp) that achieves **F1 = 82.16** and **EM = 71.06** (And of course better marks per language).
+
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain
@@ -0,0 +1,60 @@
+# ALECTRA-small-OWT
+
+This is an extension of
+[ELECTRA](https://openreview.net/forum?id=r1xMH1BtvB) small model, trained on the
+[OpenWebText corpus](https://skylion007.github.io/OpenWebTextCorpus/).
+The training task (discriminative LM / replaced-token-detection) can be generalized to any transformer type.  Here, we train an ALBERT model under the same scheme.
+
+## Pretraining task
+![electra task diagram](https://github.com/shoarora/lmtuners/raw/master/assets/electra.png)
+(figure from [Clark et al. 2020](https://openreview.net/pdf?id=r1xMH1BtvB))
+
+ELECTRA uses discriminative LM / replaced-token-detection for pretraining.
+This involves a generator (a Masked LM model) creating examples for a discriminator
+to classify as original or replaced for each token.
+
+The generator generalizes to any `*ForMaskedLM` model and the discriminator could be
+any `*ForTokenClassification` model.  Therefore, we can extend the task to ALBERT models,
+not just BERT as in the original paper.
+
+## Usage
+```python
+from transformers import AlbertForSequenceClassification, BertTokenizer
+
+# Both models use the bert-base-uncased tokenizer and vocab.
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+alectra = AlbertForSequenceClassification.from_pretrained('shoarora/alectra-small-owt')
+```
+NOTE: this ALBERT model uses a BERT WordPiece tokenizer.
+
+## Code
+The pytorch module that implements this task is available [here](https://github.com/shoarora/lmtuners/blob/master/lmtuners/lightning_modules/discriminative_lm.py).
+
+Further implementation information [here](https://github.com/shoarora/lmtuners/tree/master/experiments/disc_lm_small),
+and [here](https://github.com/shoarora/lmtuners/blob/master/experiments/disc_lm_small/train_alectra_small.py) is the script that created this model.
+
+This specific model was trained with the following params:
+- `batch_size: 512`
+- `training_steps: 5e5`
+- `warmup_steps: 4e4`
+- `learning_rate: 2e-3`
+
+
+## Downstream tasks
+#### GLUE Dev results
+| Model                    | # Params | CoLA | SST | MRPC | STS  | QQP  | MNLI | QNLI | RTE |
+| ---                      | ---      | ---  | --- | ---  | ---  | ---  | ---  | ---  | --- |
+| ELECTRA-Small++          | 14M      | 57.0 | 91. | 88.0 | 87.5 | 89.0 | 81.3 | 88.4 | 66.7|
+| ELECTRA-Small-OWT        | 14M      | 56.8 | 88.3| 87.4 | 86.8 | 88.3 | 78.9 | 87.9 | 68.5|
+| ELECTRA-Small-OWT (ours) | 17M      | 56.3 | 88.4| 75.0 | 86.1 | 89.1 | 77.9 | 83.0 | 67.1|
+| ALECTRA-Small-OWT (ours) |  4M      | 50.6 | 89.1| 86.3 | 87.2 | 89.1 | 78.2 | 85.9 | 69.6|
+
+
+#### GLUE Test results
+| Model                    | # Params | CoLA | SST | MRPC | STS  | QQP  | MNLI | QNLI | RTE |
+| ---                      | ---      | ---  | --- | ---  | ---  | ---  | ---  | ---  | --- |
+| BERT-Base                | 110M     | 52.1 | 93.5| 84.8 | 85.9 | 89.2 | 84.6 | 90.5 | 66.4|
+| GPT                      | 117M     | 45.4 | 91.3| 75.7 | 80.0 | 88.5 | 82.1 | 88.1 | 56.0|
+| ELECTRA-Small++          | 14M      | 57.0 | 91.2| 88.0 | 87.5 | 89.0 | 81.3 | 88.4 | 66.7|
+| ELECTRA-Small-OWT (ours) | 17M      | 57.4 | 89.3| 76.2 | 81.9 | 87.5 | 78.1 | 82.4 | 68.1|
+| ALECTRA-Small-OWT (ours) |  4M      | 43.9 | 87.9| 82.1 | 82.0 | 87.6 | 77.9 | 85.8 | 67.5|
@@ -0,0 +1,59 @@
+# ELECTRA-small-OWT
+
+This is an unnoficial implementation of an
+[ELECTRA](https://openreview.net/forum?id=r1xMH1BtvB) small model, trained on the
+[OpenWebText corpus](https://skylion007.github.io/OpenWebTextCorpus/).
+
+Differences from official ELECTRA models:
+ - we use a `BertForMaskedLM` as the generator and `BertForTokenClassification` as the discriminator
+ - they use an embedding projection layer, but Bert doesn't have one
+
+## Pretraining ttask
+![electra task diagram](https://github.com/shoarora/lmtuners/raw/master/assets/electra.png)
+(figure from [Clark et al. 2020](https://openreview.net/pdf?id=r1xMH1BtvB))
+
+ELECTRA uses discriminative LM / replaced-token-detection for pretraining.
+This involves a generator (a Masked LM model) creating examples for a discriminator
+to classify as original or replaced for each token.
+
+
+## Usage
+```python
+from transformers import BertForSequenceClassification, BertTokenizer
+
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+electra = BertForSequenceClassification.from_pretrained('shoarora/electra-small-owt')
+```
+
+## Code
+The pytorch module that implements this task is available [here](https://github.com/shoarora/lmtuners/blob/master/lmtuners/lightning_modules/discriminative_lm.py).
+
+Further implementation information [here](https://github.com/shoarora/lmtuners/tree/master/experiments/disc_lm_small),
+and [here](https://github.com/shoarora/lmtuners/blob/master/experiments/disc_lm_small/train_electra_small.py) is the script that created this model.
+
+This specific model was trained with the following params:
+- `batch_size: 512`
+- `training_steps: 5e5`
+- `warmup_steps: 4e4`
+- `learning_rate: 2e-3`
+
+
+## Downstream tasks
+#### GLUE Dev results
+| Model                    | # Params | CoLA | SST | MRPC | STS  | QQP  | MNLI | QNLI | RTE |
+| ---                      | ---      | ---  | --- | ---  | ---  | ---  | ---  | ---  | --- |
+| ELECTRA-Small++          | 14M      | 57.0 | 91. | 88.0 | 87.5 | 89.0 | 81.3 | 88.4 | 66.7|
+| ELECTRA-Small-OWT        | 14M      | 56.8 | 88.3| 87.4 | 86.8 | 88.3 | 78.9 | 87.9 | 68.5|
+| ELECTRA-Small-OWT (ours) | 17M      | 56.3 | 88.4| 75.0 | 86.1 | 89.1 | 77.9 | 83.0 | 67.1|
+| ALECTRA-Small-OWT (ours) |  4M      | 50.6 | 89.1| 86.3 | 87.2 | 89.1 | 78.2 | 85.9 | 69.6|
+
+- Table initialized from [ELECTRA github repo](https://github.com/google-research/electra)
+
+#### GLUE Test results
+| Model                    | # Params | CoLA | SST | MRPC | STS  | QQP  | MNLI | QNLI | RTE |
+| ---                      | ---      | ---  | --- | ---  | ---  | ---  | ---  | ---  | --- |
+| BERT-Base                | 110M     | 52.1 | 93.5| 84.8 | 85.9 | 89.2 | 84.6 | 90.5 | 66.4|
+| GPT                      | 117M     | 45.4 | 91.3| 75.7 | 80.0 | 88.5 | 82.1 | 88.1 | 56.0|
+| ELECTRA-Small++          | 14M      | 57.0 | 91.2| 88.0 | 87.5 | 89.0 | 81.3 | 88.4 | 66.7|
+| ELECTRA-Small-OWT (ours) | 17M      | 57.4 | 89.3| 76.2 | 81.9 | 87.5 | 78.1 | 82.4 | 68.1|
+| ALECTRA-Small-OWT (ours) |  4M      | 43.9 | 87.9| 82.1 | 82.0 | 87.6 | 77.9 | 85.8 | 67.5|
@@ -83,7 +83,7 @@ extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3", "sciki

 setup(
    name="transformers",
-    version="2.6.0",
+    version="2.7.0",
    author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, 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",
@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.

-__version__ = "2.6.0"
+__version__ = "2.7.0"

 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -26,6 +26,7 @@ BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/config.json",
    "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/config.json",
    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/config.json",
+    "bart-large-xsum": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-xsum/config.json",
 }


@@ -80,6 +80,7 @@ class PretrainedConfig(object):
        self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
        self.length_penalty = kwargs.pop("length_penalty", 1.0)
        self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
+        self.bad_words_ids = kwargs.pop("bad_words_ids", None)
        self.num_return_sequences = kwargs.pop("num_return_sequences", 1)

        # Fine-tuning task arguments
@@ -17,6 +17,7 @@

 import argparse
 import logging
+import os
 from pathlib import Path

 import fairseq
@@ -30,10 +31,11 @@ from transformers import (
    BartModel,
    BartTokenizer,
 )
+from transformers.modeling_bart import _make_linear_from_emb


-FAIRSEQ_MODELS = ["bart.large", "bart.large.mnli", "bart.large.cnn"]
-
+FAIRSEQ_MODELS = ["bart.large", "bart.large.mnli", "bart.large.cnn", "bart_xsum/model.pt"]
+extra_arch = {"bart.large": BartModel, "bart.large.mnli": BartForSequenceClassification}
 if version.parse(fairseq.__version__) < version.parse("0.9.0"):
    raise Exception("requires fairseq >= 0.9.0")

@@ -57,62 +59,79 @@ def rename_key(dct, old, new):
    dct[new] = val


-def convert_bart_checkpoint(checkpoint_path, pytorch_dump_folder_path):
+def load_xsum_checkpoint(checkpoint_path):
+    """Checkpoint path should end in model.pt"""
+    sd = torch.load(checkpoint_path, map_location="cpu")
+    hub_interface = torch.hub.load("pytorch/fairseq", "bart.large.cnn").eval()
+    hub_interface.model.load_state_dict(sd["model"])
+    return hub_interface
+
+
+@torch.no_grad()
+def convert_bart_checkpoint(checkpoint_path, pytorch_dump_folder_path, hf_checkpoint_name=None):
    """
    Copy/paste/tweak model's weights to our BERT structure.
    """
-    bart = torch.hub.load("pytorch/fairseq", checkpoint_path)
-    bart.eval()  # disable dropout
+    if not os.path.exists(checkpoint_path):
+        bart = torch.hub.load("pytorch/fairseq", checkpoint_path).eval()
+    else:
+        bart = load_xsum_checkpoint(checkpoint_path)
+
    bart.model.upgrade_state_dict(bart.model.state_dict())
-    hf_model_name = checkpoint_path.replace(".", "-")
-    config = BartConfig.from_pretrained(hf_model_name)
+    if hf_checkpoint_name is None:
+        hf_checkpoint_name = checkpoint_path.replace(".", "-")
+    config = BartConfig.from_pretrained(hf_checkpoint_name)
    tokens = bart.encode(SAMPLE_TEXT).unsqueeze(0)
-    tokens2 = BartTokenizer.from_pretrained(hf_model_name).encode(SAMPLE_TEXT, return_tensors="pt").unsqueeze(0)
+    tokens2 = BartTokenizer.from_pretrained(hf_checkpoint_name).encode(SAMPLE_TEXT, return_tensors="pt").unsqueeze(0)
    assert torch.eq(tokens, tokens2).all()

-    if checkpoint_path in ["bart.large", "bart.large.cnn"]:
-        state_dict = bart.model.state_dict()
-        for k in IGNORE_KEYS:
-            state_dict.pop(k, None)
-        state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"]
-        model = BartModel(config)
-        their_output = bart.extract_features(tokens)
-    else:  # MNLI Case
+    if checkpoint_path == "bart.large.mnli":
        state_dict = bart.state_dict()
-        for k in IGNORE_KEYS:
-            state_dict.pop(k, None)
+        remove_ignore_keys_(state_dict)
        state_dict["model.shared.weight"] = state_dict["model.decoder.embed_tokens.weight"]
        for src, dest in rename_keys:
            rename_key(state_dict, src, dest)
-        model = BartForSequenceClassification(config)
-        their_output = bart.predict("mnli", tokens, return_logits=True)
+        model = BartForSequenceClassification(config).eval()
+        model.load_state_dict(state_dict)
+        fairseq_output = bart.predict("mnli", tokens, return_logits=True)
+        new_model_outputs = model(tokens)[0]  # logits
+    else:  # no classification heads to worry about
+        state_dict = bart.model.state_dict()
+        remove_ignore_keys_(state_dict)
+        state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"]
+        fairseq_output = bart.extract_features(tokens)
+        if hf_checkpoint_name == "bart-large":
+            model = BartModel(config).eval()
+            model.load_state_dict(state_dict)
+            new_model_outputs = model(tokens).model[0]
+        else:
+            model = BartForConditionalGeneration(config).eval()  # an existing summarization ckpt
+            model.model.load_state_dict(state_dict)
+            if hasattr(model, "lm_head"):
+                model.lm_head = _make_linear_from_emb(model.model.shared)
+            new_model_outputs = model.model(tokens)[0]

-    # Load state dict
-    model.load_state_dict(state_dict)
-    model.eval()
    # Check results
-
-    if checkpoint_path == "bart.large.cnn":
-        model = BartForConditionalGeneration(config, base_model=model)
-        assert "lm_head.weight" in model.state_dict()
-        assert model.lm_head.out_features == config.max_position_embeddings
-        model.eval()
-        our_outputs = model.model(tokens)[0]
-    else:
-        our_outputs = model(tokens)[0]
-    assert their_output.shape == our_outputs.shape
-    assert (their_output == our_outputs).all().item()
+    assert fairseq_output.shape == new_model_outputs.shape
+    assert (fairseq_output == new_model_outputs).all().item()
    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
    model.save_pretrained(pytorch_dump_folder_path)


+def remove_ignore_keys_(state_dict):
+    for k in IGNORE_KEYS:
+        state_dict.pop(k, None)
+
+
 if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    # Required parameters
-    parser.add_argument("fairseq_path", choices=FAIRSEQ_MODELS, type=str, help="")
-
-    parser.add_argument("pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
-    args = parser.parse_args()
-    convert_bart_checkpoint(
-        args.fairseq_path, args.pytorch_dump_folder_path,
+    parser.add_argument(
+        "fairseq_path", type=str, help="bart.large, bart.large.cnn or a path to a model.pt on local filesystem."
    )
+    parser.add_argument("pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument(
+        "--hf_config", default=None, type=str, help="Which huggingface architecture to use: bart-large-xsum"
+    )
+    args = parser.parse_args()
+    convert_bart_checkpoint(args.fairseq_path, args.pytorch_dump_folder_path, hf_checkpoint_name=args.hf_config)
@@ -28,7 +28,7 @@ from ...file_utils import is_tf_available, is_torch_available
 logger = logging.getLogger(__name__)


-@dataclass(frozen=True)
+@dataclass(frozen=False)
 class InputExample:
    """
    A single training/test example for simple sequence classification.
@@ -34,6 +34,7 @@ BART_PRETRAINED_MODEL_ARCHIVE_MAP = {
    "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/pytorch_model.bin",
    "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/pytorch_model.bin",
    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/pytorch_model.bin",
+    "bart-large-xsum": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-xsum/pytorch_model.bin",
 }

 BART_START_DOCSTRING = r"""
@@ -72,6 +73,10 @@ BART_INPUTS_DOCSTRING = r"""
            Mask to avoid performing attention on padding token indices in input_ids.
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`, defaults to :obj:`None`):
+            Tuple consists of (`last_hidden_state`, `optional`: `hidden_states`, `optional`: `attentions`)
+            `last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`) is a sequence of hidden-states at the output of the last layer of the encoder.
+            Used in the cross-attention of the decoder.
        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`, defaults to :obj:`None`):
            Provide for translation and summarization training. By default, the model will create this tensor by shifting the input_ids right, following the paper.
        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`, defaults to :obj:`None`):
@@ -111,7 +116,6 @@ class PretrainedBartModel(PreTrainedModel):
    config_class = BartConfig
    base_model_prefix = "model"
    pretrained_model_archive_map = BART_PRETRAINED_MODEL_ARCHIVE_MAP
-    encoder_outputs_batch_dim_idx = 1  # outputs shaped (seq_len, bs, ...)

    def _init_weights(self, module):
        std = self.config.init_std
@@ -289,7 +293,10 @@ class BartEncoder(nn.Module):
        if self.output_hidden_states:
            encoder_states.append(x)

+        # T x B x C -> B x T x C
        encoder_states = [hidden_state.transpose(0, 1) for hidden_state in encoder_states]
+        x = x.transpose(0, 1)
+
        return x, encoder_states, all_attentions


@@ -433,7 +440,7 @@ class BartDecoder(nn.Module):
        # embed positions
        positions = self.embed_positions(input_ids, generation_mode=generation_mode)

-        if generation_mode:
+        if generation_mode and decoder_cached_states is not None:
            input_ids = input_ids[:, -1:]
            positions = positions[:, -1:]  # happens after we embed them
            assert input_ids.ne(self.padding_idx).any()
@@ -443,7 +450,11 @@ class BartDecoder(nn.Module):

        x = self.layernorm_embedding(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
-        x = x.transpose(0, 1)  # (seq_len, BS, model_dim)
+
+        # Convert to Bart output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
+        x = x.transpose(0, 1)
+        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)
+
        # decoder layers
        all_hidden_states = ()
        all_self_attns = ()
@@ -465,18 +476,19 @@ class BartDecoder(nn.Module):
                causal_mask=decoder_causal_mask,
            )

-            if self.output_past:
+            if generation_mode:
                next_decoder_cache.append(layer_past.copy())
            if self.output_hidden_states:
                all_hidden_states += (x,)
            if self.output_attentions:
                all_self_attns += (layer_self_attn,)

-        # Convert shapes from (seq_len, BS, model_dim) to (BS, seq_len, model_dim)
+        # Convert to standart output format: (seq_len, BS, model_dim) -> (BS, seq_len, model_dim)
        all_hidden_states = [hidden_state.transpose(0, 1) for hidden_state in all_hidden_states]
        x = x.transpose(0, 1)
+        encoder_hidden_states = encoder_hidden_states.transpose(0, 1)

-        if self.output_past:
+        if generation_mode:
            next_cache = ((encoder_hidden_states, encoder_padding_mask), next_decoder_cache)
        else:
            next_cache = None
@@ -800,6 +812,8 @@ class BartModel(PretrainedBartModel):

    def set_input_embeddings(self, value):
        self.shared = value
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared

    def get_output_embeddings(self):
        return _make_linear_from_emb(self.shared)  # make it on the fly
@@ -893,23 +907,18 @@ class BartForConditionalGeneration(PretrainedBartModel):
    def prepare_inputs_for_generation(self, decoder_input_ids, past, attention_mask, **kwargs):
        assert past is not None, "past has to be defined for encoder_outputs"

-        # first step, decoder_cached_states are empty
-        if not past[1]:
-            encoder_outputs, decoder_cached_states = past, None
-        else:
-            encoder_outputs, decoder_cached_states = past
+        # first step, decoder_cached_states are empty (None)
+        (encoder_outputs, encoder_attention_mask), decoder_cached_states = past
        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
+            "attention_mask": encoder_attention_mask,
            "decoder_cached_states": decoder_cached_states,
            "decoder_input_ids": decoder_input_ids,
-            "attention_mask": attention_mask,
            "generation_mode": True,
        }

    def prepare_scores_for_generation(self, scores, cur_len, max_length):
-        if cur_len == 1:
-            self._force_token_ids_generation(scores, self.config.bos_token_id)
        if cur_len == max_length - 1 and self.config.eos_token_id is not None:
            self._force_token_ids_generation(scores, self.config.eos_token_id)
        return scores
@@ -917,16 +926,19 @@ class BartForConditionalGeneration(PretrainedBartModel):
    @staticmethod
    def _reorder_cache(past, beam_idx):
        ((enc_out, enc_mask), decoder_cached_states) = past
-        reordered_past = []
-        for layer_past in decoder_cached_states:
-            # get the correct batch idx from decoder layer's batch dim for cross and self-attn
-            layer_past_new = {
-                attn_key: _reorder_buffer(attn_cache, beam_idx) for attn_key, attn_cache in layer_past.items()
-            }
-            # reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
-            # reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
-            reordered_past.append(layer_past_new)
-        new_enc_out = enc_out if enc_out is None else enc_out.index_select(1, beam_idx)
+
+        if decoder_cached_states is not None:
+            reordered_past = []
+            for layer_past in decoder_cached_states:
+                # get the correct batch idx from decoder layer's batch dim for cross and self-attn
+                layer_past_new = {
+                    attn_key: _reorder_buffer(attn_cache, beam_idx) for attn_key, attn_cache in layer_past.items()
+                }
+                reordered_past.append(layer_past_new)
+        else:
+            reordered_past = None
+
+        new_enc_out = enc_out if enc_out is None else (enc_out[0].index_select(1, beam_idx), *enc_out[1:])
        new_enc_mask = enc_mask if enc_mask is None else enc_mask.index_select(0, beam_idx)

        past = ((new_enc_out, new_enc_mask), reordered_past)
@@ -972,7 +984,7 @@ class BartForSequenceClassification(PretrainedBartModel):
    Returns:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BartConfig`) and inputs:
            loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
-                Classification  loss (cross entropy)
+                Classification loss (cross entropy)
            logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
                Classification (or regression if config.num_labels==1) scores (before SoftMax).
            hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
@@ -845,7 +845,7 @@ class BertForPreTraining(BertPreTrainedModel):
            Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 2)`):
+        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
            Prediction scores of the next sequence prediction (classification) head (scores of True/False
            continuation before SoftMax).
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
@@ -1048,7 +1048,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`next_sentence_label` is provided):
            Next sequence prediction (classification) loss.
-        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, 2)`):
+        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
@@ -27,7 +27,7 @@ from torch import nn
 from torch.nn import CrossEntropyLoss

 from .configuration_t5 import T5Config
-from .file_utils import DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings
+from .file_utils import DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_utils import PreTrainedModel, prune_linear_layer


@@ -457,7 +457,6 @@ class T5PreTrainedModel(PreTrainedModel):
    pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
    load_tf_weights = load_tf_weights_in_t5
    base_model_prefix = "transformer"
-    encoder_outputs_batch_dim_idx = 0  # outputs shaped (bs, ...)

    @property
    def dummy_inputs(self):
@@ -502,6 +501,27 @@ class T5PreTrainedModel(PreTrainedModel):
            if module.has_relative_attention_bias:
                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))

+    def _shift_right(self, input_ids):
+        decoder_start_token_id = self.config.decoder_start_token_id
+        pad_token_id = self.config.pad_token_id
+
+        assert (
+            decoder_start_token_id is not None
+        ), "self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id. See T5 docs for more information"
+
+        # shift inputs to the right
+        shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+        shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+        shifted_input_ids[..., 0] = decoder_start_token_id
+
+        assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+        # replace possible -100 values in lm_labels by `pad_token_id`
+        shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+        assert torch.all(shifted_input_ids >= 0).item(), "Verify that `lm_labels` has only positive values and -100"
+
+        return shifted_input_ids
+

 class T5Stack(T5PreTrainedModel):
    def __init__(self, config, embed_tokens=None):
@@ -696,30 +716,38 @@ T5_START_DOCSTRING = r"""    The T5 model was proposed in
 """

 T5_INPUTS_DOCSTRING = r"""
-    Inputs:
-        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary.
-            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
-
-            (a) For sequence pairs:
-
-                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-
-            (b) For single sequences:
-
-                ``tokens:         [CLS] the dog is hairy . [SEP]``
-
            T5 is a model with relative position embeddings so you should be able to pad the inputs on
-            the right or the left.
-
            Indices can be obtained using :class:`transformers.T5Tokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
-        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            To know more on how to prepare :obj:`input_ids` for pre-training take a look at
+            `T5 Training <./t5.html#training>`_ .
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            Mask to avoid performing attention on padding token indices.
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`, defaults to :obj:`None`):
+            Tuple consists of (`last_hidden_state`, `optional`: `hidden_states`, `optional`: `attentions`)
+            `last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`) is a sequence of hidden-states at the output of the last layer of the encoder.
+            Used in the cross-attention of the decoder.
+        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`, defaults to :obj:`None`):
+            Provide for sequence to sequence training. T5 uses the pad_token_id as the starting token for decoder_input_ids generation.
+            To know more on how to prepare :obj:`decoder_input_ids` for pre-training take a look at
+            `T5 Training <./t5.html#training>`_ .
+        decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`, defaults to :obj:`None`):
+            Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        head_mask: (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -729,31 +757,8 @@ T5_INPUTS_DOCSTRING = r"""
@add_start_docstrings(
    "The bare T5 Model transformer outputting raw hidden-states" "without any specific head on top.",
    T5_START_DOCSTRING,
-    T5_INPUTS_DOCSTRING,
 )
 class T5Model(T5PreTrainedModel):
-    r"""
-    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 output of the last layer of the model.
-        **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)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        tokenizer = T5Tokenizer.from_pretrained('t5-small')
-        model = T5Model.from_pretrained('t5-small')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids=input_ids)
-        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
-
-    """
-
    def __init__(self, config):
        super().__init__(config)
        self.shared = nn.Embedding(config.vocab_size, config.d_model)
@@ -783,6 +788,7 @@ class T5Model(T5PreTrainedModel):
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

+    @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -794,6 +800,34 @@ class T5Model(T5PreTrainedModel):
        decoder_inputs_embeds=None,
        head_mask=None,
    ):
+        r"""
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs.
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+            from transformers import T5Tokenizer, T5Model
+
+            tokenizer = T5Tokenizer.from_pretrained('t5-small')
+            model = T5Model.from_pretrained('t5-small')
+            input_ids = tokenizer.encode("Hello, my dog is cute", return_tensors="pt")  # Batch size 1
+            outputs = model(input_ids=input_ids, decoder_input_ids=input_ids)
+            last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+        """

        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
@@ -816,38 +850,8 @@ class T5Model(T5PreTrainedModel):
        return decoder_outputs + encoder_outputs


-@add_start_docstrings("""T5 Model with a `language modeling` head on top. """, T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """, T5_START_DOCSTRING)
 class T5ForConditionalGeneration(T5PreTrainedModel):
-    r"""
-        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
-            Labels for computing the masked language modeling loss.
-            Indices should either be in ``[0, ..., config.vocab_size]`` or -100 (see ``input_ids`` docstring).
-            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
-            in ``[0, ..., config.vocab_size]``.
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Masked 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).
-        **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)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        tokenizer = T5Tokenizer.from_pretrained('t5-small')
-        model = T5ForConditionalGeneration.from_pretrained('t5-small')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids=input_ids, lm_labels=input_ids)
-        loss, prediction_scores = outputs[:2]
-
-    """
-
    def __init__(self, config):
        super().__init__(config)
        self.model_dim = config.d_model
@@ -879,6 +883,7 @@ class T5ForConditionalGeneration(T5PreTrainedModel):
    def get_encoder(self):
        return self.encoder

+    @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
@@ -891,6 +896,44 @@ class T5ForConditionalGeneration(T5PreTrainedModel):
        decoder_inputs_embeds=None,
        head_mask=None,
    ):
+        r"""
+        lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+                Labels for computing the sequence classification/regression loss.
+                Indices should be in :obj:`[-100, 0, ..., config.vocab_size - 1]`.
+                All labels set to ``-100`` are ignored (masked), the loss is only
+                computed for labels in ``[0, ..., config.vocab_size]``
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs.
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`lm_label` is provided):
+            Classification loss (cross entropy).
+        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention.
+
+    Examples::
+
+        from transformers import T5Tokenizer, T5ForConditionalGeneration
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5ForConditionalGeneration.from_pretrained('t5-small')
+        input_ids = tokenizer.encode("Hello, my dog is cute", return_tensors="pt")  # Batch size 1
+        outputs = model(input_ids=input_ids, decoder_input_ids=input_ids, lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5ForConditionalGeneration.from_pretrained('t5-small')
+        input_ids = tokenizer.encode("summarize: Hello, my dog is cute", return_tensors="pt")  # Batch size 1
+        outputs = model.generate(input_ids)
+        """

        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
@@ -901,6 +944,10 @@ class T5ForConditionalGeneration(T5PreTrainedModel):

        hidden_states = encoder_outputs[0]

+        if lm_labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
+            # get decoder inputs from shifting lm labels to the right
+            decoder_input_ids = self._shift_right(lm_labels)
+
        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
@@ -919,10 +966,8 @@ class T5ForConditionalGeneration(T5PreTrainedModel):

        decoder_outputs = (lm_logits,) + decoder_outputs[1:]  # Add hidden states and attention if they are here
        if lm_labels is not None:
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = lm_labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss(ignore_index=-100)
-            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), lm_labels.view(-1))
            decoder_outputs = (
                loss,
            ) + decoder_outputs  # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
@@ -24,7 +24,7 @@ import math
 import tensorflow as tf

 from .configuration_t5 import T5Config
-from .file_utils import DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings
+from .file_utils import DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list


@@ -119,7 +119,7 @@ class TFT5Attention(tf.keras.layers.Layer):

        if self.has_relative_attention_bias:
            self.relative_attention_bias = tf.keras.layers.Embedding(
-                self.relative_attention_num_buckets, self.n_heads, name="relative_attention_bias"
+                self.relative_attention_num_buckets, self.n_heads, name="relative_attention_bias",
            )
        self.pruned_heads = set()

@@ -178,13 +178,15 @@ class TFT5Attention(tf.keras.layers.Layer):
        memory_position = tf.range(klen)[None, :]
        relative_position = memory_position - context_position  # shape (qlen, klen)
        rp_bucket = self._relative_position_bucket(
-            relative_position, bidirectional=not self.is_decoder, num_buckets=self.relative_attention_num_buckets
+            relative_position, bidirectional=not self.is_decoder, num_buckets=self.relative_attention_num_buckets,
        )
        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
        values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0)  # shape (1, num_heads, qlen, klen)
        return values

-    def call(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None, training=False):
+    def call(
+        self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None, training=False,
+    ):
        """
        Self-attention (if kv is None) or attention over source sentence (provided by kv).
        """
@@ -261,15 +263,17 @@ class TFT5LayerSelfAttention(tf.keras.layers.Layer):
    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
        super().__init__(**kwargs)
        self.SelfAttention = TFT5Attention(
-            config, has_relative_attention_bias=has_relative_attention_bias, name="SelfAttention"
+            config, has_relative_attention_bias=has_relative_attention_bias, name="SelfAttention",
        )
        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon, name="layer_norm")
        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)

-    def call(self, hidden_states, attention_mask=None, position_bias=None, head_mask=None, training=False):
+    def call(
+        self, hidden_states, attention_mask=None, position_bias=None, head_mask=None, training=False,
+    ):
        norm_x = self.layer_norm(hidden_states)
        attention_output = self.SelfAttention(
-            norm_x, mask=attention_mask, position_bias=position_bias, head_mask=head_mask, training=training
+            norm_x, mask=attention_mask, position_bias=position_bias, head_mask=head_mask, training=training,
        )
        y = attention_output[0]
        layer_output = hidden_states + self.dropout(y, training=training)
@@ -281,15 +285,17 @@ class TFT5LayerCrossAttention(tf.keras.layers.Layer):
    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
        super().__init__(**kwargs)
        self.EncDecAttention = TFT5Attention(
-            config, has_relative_attention_bias=has_relative_attention_bias, name="EncDecAttention"
+            config, has_relative_attention_bias=has_relative_attention_bias, name="EncDecAttention",
        )
        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon, name="layer_norm")
        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)

-    def call(self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None, training=False):
+    def call(
+        self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None, training=False,
+    ):
        norm_x = self.layer_norm(hidden_states)
        attention_output = self.EncDecAttention(
-            norm_x, mask=attention_mask, kv=kv, position_bias=position_bias, head_mask=head_mask, training=training
+            norm_x, mask=attention_mask, kv=kv, position_bias=position_bias, head_mask=head_mask, training=training,
        )
        y = attention_output[0]
        layer_output = hidden_states + self.dropout(y, training=training)
@@ -303,12 +309,12 @@ class TFT5Block(tf.keras.layers.Layer):
        self.is_decoder = config.is_decoder
        self.layer = []
        self.layer.append(
-            TFT5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias, name="layer_._0")
+            TFT5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias, name="layer_._0",)
        )
        if self.is_decoder:
            self.layer.append(
                TFT5LayerCrossAttention(
-                    config, has_relative_attention_bias=has_relative_attention_bias, name="layer_._1"
+                    config, has_relative_attention_bias=has_relative_attention_bias, name="layer_._1",
                )
            )
            self.layer.append(TFT5LayerFF(config, name="layer_._2"))
@@ -402,7 +408,7 @@ class TFT5MainLayer(tf.keras.layers.Layer):
        self.num_hidden_layers = config.num_layers

        self.block = [
-            TFT5Block(config, has_relative_attention_bias=bool(i == 0), name="block_._{}".format(i))
+            TFT5Block(config, has_relative_attention_bias=bool(i == 0), name="block_._{}".format(i),)
            for i in range(config.num_layers)
        ]
        self.final_layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon, name="final_layer_norm")
@@ -469,7 +475,7 @@ class TFT5MainLayer(tf.keras.layers.Layer):
            if self.config.is_decoder:
                seq_ids = tf.range(seq_length)
                causal_mask = tf.less_equal(
-                    tf.tile(seq_ids[None, None, :], (batch_size, seq_length, 1)), seq_ids[None, :, None]
+                    tf.tile(seq_ids[None, None, :], (batch_size, seq_length, 1)), seq_ids[None, :, None],
                )
                causal_mask = tf.cast(causal_mask, dtype=tf.float32)
                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
@@ -586,8 +592,8 @@ class TFT5PreTrainedModel(TFPreTrainedModel):
        input_ids = tf.constant(DUMMY_INPUTS)
        input_mask = tf.constant(DUMMY_MASK)
        dummy_inputs = {
+            "inputs": input_ids,
            "decoder_input_ids": input_ids,
-            "input_ids": input_ids,
            "decoder_attention_mask": input_mask,
        }
        return dummy_inputs
@@ -630,31 +636,41 @@ T5_START_DOCSTRING = r"""    The T5 model was proposed in
 """

 T5_INPUTS_DOCSTRING = r"""
-    Inputs:
-        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+    Args:
+        inputs are usually used as a `dict` (see T5 description above for more information) containing all the following.
+
+        inputs (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary.
-            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
-
-            (a) For sequence pairs:
-
-                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-
-            (b) For single sequences:
-
-                ``tokens:         [CLS] the dog is hairy . [SEP]``
-
-
            T5 is a model with relative position embeddings so you should be able to pad the inputs on
            the right or the left.
-
            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            To know more on how to prepare :obj:`input_ids` for pre-training take a look at
+            `T5 Training <./t5.html#training>`_ .
            See :func:`transformers.PreTrainedTokenizer.encode` and
            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
-        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+        decoder_input_ids (:obj:`tf.Tensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`, defaults to :obj:`None`):
+            Provide for sequence to sequence training. T5 uses the pad_token_id as the starting token for decoder_input_ids generation.
+        attention_mask (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            Mask to avoid performing attention on padding token indices.
            Mask values selected in ``[0, 1]``:
            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        encoder_outputs (:obj:`tuple(tuple(tf.FloatTensor)`, `optional`, defaults to :obj:`None`):
+            Tuple consists of (`last_hidden_state`, `optional`: `hidden_states`, `optional`: `attentions`)
+            `last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`) is a sequence of hidden-states at the output of the last layer of the encoder.
+            Used in the cross-attention of the decoder.
+        decoder_attention_mask (:obj:`tf.Tensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`, defaults to :obj:`None`):
+            Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.
+        inputs_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        decoder_inputs_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+            To know more on how to prepare :obj:`decoder_input_ids` for pre-training take a look at
+            `T5 Training <./t5.html#training>`_ .
+        head_mask: (:obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -664,34 +680,8 @@ T5_INPUTS_DOCSTRING = r"""
@add_start_docstrings(
    "The bare T5 Model transformer outputting raw hidden-states" "without any specific head on top.",
    T5_START_DOCSTRING,
-    T5_INPUTS_DOCSTRING,
 )
 class TFT5Model(TFT5PreTrainedModel):
-    r"""
-    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 output of the last layer of the model.
-        **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)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        import tensorflow as tf
-        from transformers import T5Tokenizer, TFT5Model
-
-        tokenizer = T5Tokenizer.from_pretrained('t5-small')
-        model = TFT5Model.from_pretrained('t5-small')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
-        outputs = model(input_ids=input_ids)
-        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
-
-    """
-
    def __init__(self, config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)
        self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model, name="shared")
@@ -715,15 +705,44 @@ class TFT5Model(TFT5PreTrainedModel):
    def get_output_embeddings(self):
        return self.shared

-    def call(self, decoder_input_ids, **kwargs):
+    @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
+    def call(self, inputs, **kwargs):
+        r"""
+    Return:
+        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs.
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

-        if isinstance(decoder_input_ids, dict):
-            kwargs.update(decoder_input_ids)
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+                :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        from transformers import T5Tokenizer, TFT5Model
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5Model.from_pretrained('t5-small')
+        input_ids = tokenizer.encode("Hello, my dog is cute", return_tensors="tf")  # Batch size 1
+        outputs = model(input_ids, input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+        """
+
+        if isinstance(inputs, dict):
+            kwargs.update(inputs)
        else:
-            kwargs["decoder_input_ids"] = decoder_input_ids
+            kwargs["inputs"] = inputs

        # retrieve arguments
-        input_ids = kwargs.get("input_ids", None)
+        input_ids = kwargs.get("inputs", None)
        decoder_input_ids = kwargs.get("decoder_input_ids", None)
        attention_mask = kwargs.get("attention_mask", None)
        encoder_outputs = kwargs.get("encoder_outputs", None)
@@ -735,7 +754,7 @@ class TFT5Model(TFT5PreTrainedModel):
        # Encode if needed (training, first prediction pass)
        if encoder_outputs is None:
            encoder_outputs = self.encoder(
-                input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask
+                input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask,
            )

        hidden_states = encoder_outputs[0]
@@ -753,33 +772,8 @@ class TFT5Model(TFT5PreTrainedModel):
        return decoder_outputs + encoder_outputs


-@add_start_docstrings("""T5 Model with a `language modeling` head on top. """, T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """, T5_START_DOCSTRING)
 class TFT5ForConditionalGeneration(TFT5PreTrainedModel):
-    r"""
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **prediction_scores**: ``Numpy array`` or ``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).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
-
-    Examples::
-
-        import tensorflow as tf
-        from transformers import T5Tokenizer, TFT5ForConditionalGeneration
-
-        tokenizer = T5Tokenizer.from_pretrained('t5-small')
-        model = TFT5ForConditionalGeneration.from_pretrained('t5-small')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
-        outputs = model(input_ids=input_ids)
-        prediction_scores = outputs[0]
-
-    """
-
    def __init__(self, config, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)
        self.model_dim = config.d_model
@@ -808,15 +802,50 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel):
    def get_encoder(self):
        return self.encoder

-    def call(self, decoder_input_ids, **kwargs):
+    @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
+    def call(self, inputs, **kwargs):
+        r"""
+    Return:
+        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs.
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`lm_label` is provided):
+            Classification loss (cross entropy).
+        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

-        if isinstance(decoder_input_ids, dict):
-            kwargs.update(decoder_input_ids)
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention.
+
+    Examples::
+
+        from transformers import T5Tokenizer, TFT5ForConditionalGeneration
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5ForConditionalGeneration.from_pretrained('t5-small')
+        input_ids = tokenizer.encode("Hello, my dog is cute", return_tensors="tf")  # Batch size 1
+        outputs = model(input_ids, input_ids=input_ids)
+        prediction_scores = outputs[0]
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5ForConditionalGeneration.from_pretrained('t5-small')
+        input_ids = tokenizer.encode("summarize: Hello, my dog is cute", return_tensors="tf")  # Batch size 1
+        model.generate(input_ids)
+
+        """
+
+        if isinstance(inputs, dict):
+            kwargs.update(inputs)
        else:
-            kwargs["decoder_input_ids"] = decoder_input_ids
+            kwargs["inputs"] = inputs

        # retrieve arguments
-        input_ids = kwargs.get("input_ids", None)
+        input_ids = kwargs.get("inputs", None)
        decoder_input_ids = kwargs.get("decoder_input_ids", None)
        attention_mask = kwargs.get("attention_mask", None)
        encoder_outputs = kwargs.get("encoder_outputs", None)
@@ -829,7 +858,7 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel):
        if encoder_outputs is None:
            # Convert encoder inputs in embeddings if needed
            encoder_outputs = self.encoder(
-                input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask
+                input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds, head_mask=head_mask,
            )

        hidden_states = encoder_outputs[0]
@@ -861,7 +890,8 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel):
            encoder_outputs = (past,)

        return {
-            "inputs": input_ids,
+            "inputs": None,  # inputs don't have to be defined, but still need to be passed to make Keras.layer.__call__ happy
+            "decoder_input_ids": input_ids,  # input_ids are the decoder_input_ids
            "encoder_outputs": encoder_outputs,
            "attention_mask": attention_mask,
        }
@@ -231,7 +231,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):

    def save_pretrained(self, save_directory):
        """ Save a model and its configuration file to a directory, so that it
-            can be re-loaded using the `:func:`~transformers.PreTrainedModel.from_pretrained`` class method.
+            can be re-loaded using the :func:`~transformers.PreTrainedModel.from_pretrained` class method.
        """
        assert os.path.isdir(
            save_directory
@@ -467,6 +467,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
        top_k=None,
        top_p=None,
        repetition_penalty=None,
+        bad_words_ids=None,
        bos_token_id=None,
        pad_token_id=None,
        eos_token_id=None,
@@ -523,8 +524,8 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            pad_token_id: (`optional`) int
                Pad token. Defaults to pad_token_id as defined in the models config.

-            eos_token_ids: (`optional`) int or list of int
-                End of sequence token or list of tokens to stop the generation. Default to 0.
+            eos_token_id: (`optional`) int
+                EOS token. Defaults to eos_token_id as defined in the models config.

            length_penalty: (`optional`) float
                Exponential penalty to the length. Default to 1.
@@ -532,6 +533,9 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            no_repeat_ngram_size: (`optional`) int
                If set to int > 0, all ngrams of size `no_repeat_ngram_size` can only occur once.

+            bad_words_ids: (`optional`) list of lists of int
+                `bad_words_ids` contains tokens that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use `tokenizer.encode(bad_word, add_prefix_space=True)`.
+
            num_return_sequences: (`optional`) int
                The number of independently computed returned sequences for each element in the batch. Default to 1.

@@ -541,7 +545,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
                Defaults to `None`.

-            `What are attention masks? <../glossary.html#attention-mask>`__
+                `What are attention masks? <../glossary.html#attention-mask>`__

            decoder_start_token_id=None: (`optional`) int
                If an encoder-decoder model starts decoding with a different token than BOS.
@@ -582,6 +586,12 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

+            tokenizer = AutoTokenizer.from_pretrained('gpt2')   # Initialize tokenizer
+            model = TFAutoModelWithLMHead.from_pretrained('gpt2')    # Download model and configuration from S3 and cache.
+            input_context = 'My cute dog'  # "Legal" is one of the control codes for ctrl
+            bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']]
+            input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids)  # generate sequences without allowing bad_words to be generated
        """

        # We cannot generate if the model does not have a LM head
@@ -607,6 +617,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
        no_repeat_ngram_size = (
            no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
        )
+        bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
        num_return_sequences = (
            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
        )
@@ -641,6 +652,9 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
        assert (
            isinstance(num_return_sequences, int) and num_return_sequences > 0
        ), "`num_return_sequences` should be a strictely positive integer."
+        assert (
+            bad_words_ids is None or isinstance(bad_words_ids, list) and isinstance(bad_words_ids[0], list)
+        ), "`bad_words_ids` is either `None` or a list of lists of tokens that should not be generated"

        if input_ids is None:
            assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
@@ -742,6 +756,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
@@ -766,6 +781,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
@@ -790,6 +806,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
+        bad_words_ids,
        bos_token_id,
        pad_token_id,
        eos_token_id,
@@ -828,7 +845,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            if no_repeat_ngram_size > 0:
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
-                banned_tokens = calc_banned_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
+                banned_tokens = calc_banned_ngram_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
                # create banned_tokens boolean mask
                banned_tokens_indices_mask = []
                for banned_tokens_slice in banned_tokens:
@@ -840,6 +857,20 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                    next_token_logits, tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf")
                )

+            if bad_words_ids is not None:
+                # calculate a list of banned tokens according to bad words
+                banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
+
+                banned_tokens_indices_mask = []
+                for banned_tokens_slice in banned_tokens:
+                    banned_tokens_indices_mask.append(
+                        [True if token in banned_tokens_slice else False for token in range(vocab_size)]
+                    )
+
+                next_token_logits = set_tensor_by_indices_to_value(
+                    next_token_logits, tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf")
+                )
+
            # set eos token prob to zero if min_length is not reached
            if eos_token_id is not None and cur_len < min_length:
                # create eos_token_id boolean mask
@@ -936,6 +967,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
+        bad_words_ids,
        bos_token_id,
        pad_token_id,
        decoder_start_token_id,
@@ -1012,7 +1044,9 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
                num_batch_hypotheses = batch_size * num_beams
-                banned_tokens = calc_banned_tokens(input_ids, num_batch_hypotheses, no_repeat_ngram_size, cur_len)
+                banned_tokens = calc_banned_ngram_tokens(
+                    input_ids, num_batch_hypotheses, no_repeat_ngram_size, cur_len
+                )
                # create banned_tokens boolean mask
                banned_tokens_indices_mask = []
                for banned_tokens_slice in banned_tokens:
@@ -1024,6 +1058,20 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                    scores, tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf")
                )

+            if bad_words_ids is not None:
+                # calculate a list of banned tokens according to bad words
+                banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
+
+                banned_tokens_indices_mask = []
+                for banned_tokens_slice in banned_tokens:
+                    banned_tokens_indices_mask.append(
+                        [True if token in banned_tokens_slice else False for token in range(vocab_size)]
+                    )
+
+                scores = set_tensor_by_indices_to_value(
+                    scores, tf.convert_to_tensor(banned_tokens_indices_mask, dtype=tf.bool), -float("inf")
+                )
+
            assert shape_list(scores) == [batch_size * num_beams, vocab_size]

            if do_sample:
@@ -1064,12 +1112,12 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            assert shape_list(next_scores) == shape_list(next_tokens) == [batch_size, 2 * num_beams]

            # next batch beam content
-            # list of (batch_size * num_beams) tuple(next hypothesis score, next token, current position in the batch)
            next_batch_beam = []

            # for each sentence
            for batch_idx in range(batch_size):

+                # if we are done with this sentence
                if done[batch_idx]:
                    assert (
                        len(generated_hyps[batch_idx]) >= num_beams
@@ -1087,14 +1135,13 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                for beam_token_rank, (beam_token_id, beam_token_score) in enumerate(
                    zip(next_tokens[batch_idx], next_scores[batch_idx])
                ):
-
                    # get beam and token IDs
                    beam_id = beam_token_id // vocab_size
                    token_id = beam_token_id % vocab_size

                    effective_beam_id = batch_idx * num_beams + beam_id
                    # add to generated hypotheses if end of sentence or last iteration
-                    if eos_token_id is not None and token_id.numpy() is eos_token_id:
+                    if (eos_token_id is not None) and (token_id.numpy() == eos_token_id):
                        # if beam_token does not belong to top num_beams tokens, it should not be added
                        is_beam_token_worse_than_top_num_beams = beam_token_rank >= num_beams
                        if is_beam_token_worse_than_top_num_beams:
@@ -1110,9 +1157,9 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
                    if len(next_sent_beam) == num_beams:
                        break

-                # if we are done with this sentence
+                # Check if were done so that we can save a pad step if all(done)
                done[batch_idx] = done[batch_idx] or generated_hyps[batch_idx].is_done(
-                    tf.reduce_max(next_scores[batch_idx]).numpy()
+                    tf.reduce_max(next_scores[batch_idx]).numpy(), cur_len=cur_len
                )

                # update next beam content
@@ -1130,6 +1177,8 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            beam_tokens = tf.convert_to_tensor([x[1] for x in next_batch_beam], dtype=tf.int32)
            beam_idx = tf.convert_to_tensor([x[2] for x in next_batch_beam], dtype=tf.int32)

+            print("Scores: {}-{}".format(cur_len, beam_scores.numpy()))
+
            # re-order batch
            input_ids = tf.stack([tf.identity(input_ids[x, :]) for x in beam_idx])
            input_ids = tf.concat([input_ids, tf.expand_dims(beam_tokens, 1)], axis=-1)
@@ -1137,6 +1186,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
            if past is not None:
                past = self._reorder_cache(past, beam_idx)

+            # extend attention_mask for new generated input if only decoder
            if self.config.is_encoder_decoder is False:
                attention_mask = tf.concat(
                    [attention_mask, tf.ones((shape_list(attention_mask)[0], 1), dtype=tf.int32)], axis=-1
@@ -1196,16 +1246,26 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):

            # fill with hypothesis and eos_token_id if necessary
            for i, hypo in enumerate(best):
-                padding = tf.ones((sent_max_len - shape_list(hypo)[0],), dtype=tf.int32) * pad_token_id
-                decoded_hypo = tf.concat([hypo, padding], axis=0)
+                assert sent_lengths[i] == shape_list(hypo)[0]
+                # if sent_length is max_len do not pad
+                if sent_lengths[i] == sent_max_len:
+                    decoded_slice = hypo
+                else:
+                    # else pad to sent_max_len
+                    num_pad_tokens = sent_max_len - sent_lengths[i]
+                    padding = pad_token_id * tf.ones((num_pad_tokens,), dtype=tf.int32)
+                    decoded_slice = tf.concat([hypo, padding], axis=-1)
+
+                    # finish sentence with EOS token
+                    if sent_lengths[i] < max_length:
+                        decoded_slice = tf.where(
+                            tf.range(sent_max_len, dtype=tf.int32) == sent_lengths[i],
+                            eos_token_id * tf.ones((sent_max_len,), dtype=tf.int32),
+                            decoded_slice,
+                        )
+                # add to list
+                decoded_list.append(decoded_slice)

-                if sent_lengths[i] < max_length:
-                    decoded_hypo = tf.where(
-                        tf.range(max_length) == sent_lengths[i],
-                        eos_token_id * tf.ones((sent_max_len,), dtype=tf.int32),
-                        decoded_hypo,
-                    )
-                decoded_list.append(decoded_hypo)
            decoded = tf.stack(decoded_list)
        else:
            # none of the hypotheses have an eos_token
@@ -1243,7 +1303,7 @@ def _create_next_token_logits_penalties(input_ids, logits, repetition_penalty):
    return tf.convert_to_tensor(token_penalties, dtype=tf.float32)


-def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len):
+def calc_banned_ngram_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len):
    # Copied from fairseq for no_repeat_ngram in beam_search"""
    if cur_len + 1 < no_repeat_ngram_size:
        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
@@ -1266,6 +1326,42 @@ def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len)
    return banned_tokens


+def calc_banned_bad_words_ids(prev_input_ids, bad_words_ids):
+    banned_tokens = []
+
+    def _tokens_match(prev_tokens, tokens):
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        if len(tokens) > len(prev_input_ids):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+
+        if prev_tokens[-len(tokens) :] == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    for prev_input_ids_slice in prev_input_ids:
+        banned_tokens_slice = []
+
+        for banned_token_seq in bad_words_ids:
+            assert len(banned_token_seq) > 0, "Banned words token sequences {} cannot have an empty list".format(
+                bad_words_ids
+            )
+
+            if _tokens_match(prev_input_ids_slice.numpy().tolist(), banned_token_seq[:-1]) is False:
+                # if tokens do not match continue
+                continue
+
+            banned_tokens_slice.append(banned_token_seq[-1])
+
+        banned_tokens.append(banned_tokens_slice)
+
+    return banned_tokens
+
+
 def tf_top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
@@ -658,23 +658,26 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
    def generate(
        self,
        input_ids=None,
-        max_length=None,
+        attention_mask=None,
+        encoder_input_ids=None,
+        decoder_input_ids=None,
        min_length=None,
-        do_sample=None,
+        max_length=None,
        early_stopping=None,
+        num_return_sequences=None,
        num_beams=None,
+        do_sample=None,
        temperature=None,
        top_k=None,
        top_p=None,
-        repetition_penalty=None,
+        bad_words_ids=None,
        bos_token_id=None,
        pad_token_id=None,
        eos_token_id=None,
        length_penalty=None,
        no_repeat_ngram_size=None,
-        num_return_sequences=None,
-        attention_mask=None,
-        decoder_start_token_id=None,
+        repetition_penalty=None,
+        use_cache=None,
    ):
        r""" Generates sequences for models with a LM head. The method currently supports greedy decoding, beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.

@@ -687,24 +690,42 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        Parameters:

            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
-                The sequence used as a prompt for the generation. If `None` the method initializes
-                it as an empty `torch.LongTensor` of shape `(1,)`.
+                Short-hand for either encoder_input_ids in seq2seq models or decoder_input_ids for language models.

-            max_length: (`optional`) int
-                The max length of the sequence to be generated.  Between `min_length` and infinity. Default to 20.
+            attention_mask (`optional`) obj: `torch.LongTensor` of same shape as `input_ids`
+                Mask to avoid performing attention on padding token indices.
+                Mask values selected in ``[0, 1]``:
+                ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+                Defaults to `None`.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+
+            encoder_input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
+                encoder inputs in the encoder-decoder setting (uses input_ids if None)
+
+            decoder_input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
+                The sequence used as a prompt for the generation in the encoder-decoder setting.
+                If `None` in the seq2seq setting, the method initializes it as a `torch.LongTensor` of shape `(batch_size, 1)` filled with bos_token_id.
+                If `None` in language modeling setting, uses input_ids if available, or initializes with bos_token_id otherwise.

            min_length: (`optional`) int
                The min length of the sequence to be generated.  Between 0 and infinity. Default to 0.

-            do_sample: (`optional`) bool
-                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.
+            max_length: (`optional`) int
+                The max length of the sequence to be generated.  Between `min_length` and infinity. Default to 20.

            early_stopping: (`optional`) bool
                if set to `True` beam search is stopped when at least `num_beams` sentences finished per batch. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.

+            num_return_sequences: (`optional`) int
+                The number of independently computed returned sequences for each element in the batch. Default to 1.
+
            num_beams: (`optional`) int
                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.

+            do_sample: (`optional`) bool
+                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.
+
            temperature: (`optional`) float
                The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.

@@ -714,20 +735,14 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            top_p: (`optional`) float
                The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.

-            repetition_penalty: (`optional`) float
-                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.
-
            pad_token_id: (`optional`) int
                Padding token. Default to specicic model pad_token_id or None if it does not exist.

            bos_token_id: (`optional`) int
-                BOS token. Defaults to bos_token_id as defined in the models config.
+                BOS token. Defaults to `bos_token_id` as defined in the models config.

-            pad_token_id: (`optional`) int
-                Pad token. Defaults to pad_token_id as defined in the models config.
-
-            eos_token_ids: (`optional`) int or list of int
-                End of sequence token or list of tokens to stop the generation. Default to eos_token_ids as defined in the models config.
+            eos_token_id: (`optional`) int
+                EOS token. Defaults to `eos_token_id` as defined in the models config.

            length_penalty: (`optional`) float
                Exponential penalty to the length. Default to 1.
@@ -735,20 +750,15 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            no_repeat_ngram_size: (`optional`) int
                If set to int > 0, all ngrams of size `no_repeat_ngram_size` can only occur once.

-            num_return_sequences: (`optional`) int
-                The number of independently computed returned sequences for each element in the batch. Default to 1.
+            repetition_penalty: (`optional`) float
+                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.

-            attention_mask (`optional`) obj: `torch.LongTensor` of same shape as `input_ids`
-                Mask to avoid performing attention on padding token indices.
-                Mask values selected in ``[0, 1]``:
-                ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-                Defaults to `None`.
+            bad_words_ids: (`optional`) list of lists of int
+                `bad_words_ids` contains tokens that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use `tokenizer.encode(bad_word, add_prefix_space=True)`.

-            `What are attention masks? <../glossary.html#attention-mask>`__
+            use_cache: (`optional`) bool
+                If set to `True` the model re-uses pre-computed decoder hidden states from one time step to the next.

-            decoder_start_token_id=None: (`optional`) int
-                If an encoder-decoder model starts decoding with a different token than BOS.
-                Defaults to `None` and is changed to `BOS` later.

        Return:

@@ -785,6 +795,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

+            tokenizer = AutoTokenizer.from_pretrained('gpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('gpt2')    # Download model and configuration from S3 and cache.
+            input_context = 'My cute dog'  # "Legal" is one of the control codes for ctrl
+            bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']]
+            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids)  # generate sequences without allowing bad_words to be generated
        """

        # We cannot generate if the model does not have a LM head
@@ -794,31 +810,31 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                "Please use another model class (e.g. `OpenAIGPTLMHeadModel`, `XLNetLMHeadModel`, `GPT2LMHeadModel`, `CTRLLMHeadModel`, `T5WithLMHeadModel`, `TransfoXLLMHeadModel`, `XLMWithLMHeadModel`, `BartForConditionalGeneration` )"
            )

-        max_length = max_length if max_length is not None else self.config.max_length
        min_length = min_length if min_length is not None else self.config.min_length
-        do_sample = do_sample if do_sample is not None else self.config.do_sample
+        max_length = max_length if max_length is not None else self.config.max_length
        early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
+        num_return_sequences = (
+            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
+        )
        num_beams = num_beams if num_beams is not None else self.config.num_beams
+        do_sample = do_sample if do_sample is not None else self.config.do_sample
        temperature = temperature if temperature is not None else self.config.temperature
        top_k = top_k if top_k is not None else self.config.top_k
        top_p = top_p if top_p is not None else self.config.top_p
        repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
-        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
        pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
        length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
        no_repeat_ngram_size = (
            no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
        )
-        num_return_sequences = (
-            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
-        )
-        decoder_start_token_id = (
-            decoder_start_token_id if decoder_start_token_id is not None else self.config.decoder_start_token_id
-        )
+        bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
+

        if input_ids is not None:
            batch_size = input_ids.shape[0]  # overriden by the input batch_size
+        elif encoder_input_ids is not None:
+            batch_size = encoder_input_ids.shape[0]
        else:
            batch_size = 1

@@ -831,9 +847,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
        assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
-        assert input_ids is not None or (
-            isinstance(bos_token_id, int) and bos_token_id >= 0
-        ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
        assert pad_token_id is None or (
            isinstance(pad_token_id, int) and (pad_token_id >= 0)
        ), "`pad_token_id` should be a positive integer."
@@ -847,17 +860,36 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        assert (
            isinstance(num_return_sequences, int) and num_return_sequences > 0
        ), "`num_return_sequences` should be a strictly positive integer."
+        assert (
+            bad_words_ids is None or isinstance(bad_words_ids, list) and isinstance(bad_words_ids[0], list)
+        ), "`bad_words_ids` is either `None` or a list of lists of tokens that should not be generated"

-        if input_ids is None:
+        # different requirements fod decoder-only and encoder-decoder
+        if self.config.is_encoder_decoder:
+            assert hasattr(self, "get_encoder"), "{} should have a 'get_encoder' function defined".format(self)
+            assert callable(self.get_encoder), "{} should be a method".format(self.get_encoder)
+            assert input_ids is not None or encoder_input_ids is not None, "the encoder inputs should to be provided"
+            bos_token_id = bos_token_id if bos_token_id is not None else self.config.decoder_start_token_id
+            encoder_ids = encoder_input_ids if encoder_input_ids is not None else input_ids # for back-compatibility, esp summarization examples
+            decoder_ids = decoder_input_ids
+        else:
+            decoder_ids = decoder_input_ids if decoder_input_ids is not None else input_ids
+
+        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
+        
+        assert decoder_ids is not None or (
+            isinstance(bos_token_id, int) and bos_token_id >= 0
+        ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
+        if decoder_ids is None:
            assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
                "you should either supply a context to complete as `input_ids` input "
                "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
            )
-            input_ids = torch.full(
+            decoder_ids = torch.full(
                (batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device,
            )
        else:
-            assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
+            assert decoder_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."

        # not allow to duplicate outputs when greedy decoding
        if do_sample is False:
@@ -866,7 +898,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                assert (
                    num_return_sequences == 1
                ), "Greedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1"
-
            else:
                # beam_search greedy generation conditions
                assert (
@@ -875,10 +906,23 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):

        # create attention mask if necessary
        # TODO (PVP): this should later be handled by the forward fn() in each model in the future see PR 3140
-        if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in input_ids):
-            attention_mask = input_ids.ne(pad_token_id).long()
-        elif attention_mask is None:
-            attention_mask = input_ids.new_ones(input_ids.shape)
+        if self.config.is_encoder_decoder:
+            if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in encoder_ids):
+                enc_attention_mask = encoder_ids.ne(pad_token_id).long()
+            elif attention_mask is None:
+                enc_attention_mask = encoder_ids.new_ones(decoder_ids.shape)
+            else:
+                enc_attention_mask = attention_mask
+            # TODO (Yacine): implement behavior for decoder side in encoder-decoder
+            dec_attention_mask = None
+        else:
+            enc_attention_mask = None
+            if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in decoder_ids):
+                dec_attention_mask = decoder_ids.ne(pad_token_id).long()
+            elif attention_mask is None:
+                dec_attention_mask = decoder_ids.new_ones(decoder_ids.shape)
+            else:
+                dec_attention_mask = attention_mask

        # set pad_token_id to eos_token_id if not set. Important that this is done after
        # attention_mask is created
@@ -899,65 +943,54 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            effective_batch_size = batch_size
            effective_batch_mult = 1

-        if self.config.is_encoder_decoder:
-            if decoder_start_token_id is None:
-                decoder_start_token_id = bos_token_id
-
-            assert (
-                decoder_start_token_id is not None
-            ), "decoder_start_token_id or bos_token_id has to be defined for encoder-decoder generation"
-            assert hasattr(self, "get_encoder"), "{} should have a 'get_encoder' function defined".format(self)
-            assert callable(self.get_encoder), "{} should be a method".format(self.get_encoder)
-
-            # get encoder and store encoder outputs
-            encoder = self.get_encoder()
-
-            encoder_outputs = encoder(input_ids, attention_mask=attention_mask)
-
-        # Expand input ids if num_beams > 1 or num_return_sequences > 1
+        # Expand decoder_ids if num_beams > 1 or num_return_sequences > 1
        if num_return_sequences > 1 or num_beams > 1:
-            input_ids_len = input_ids.shape[-1]
-            input_ids = input_ids.unsqueeze(1).expand(batch_size, effective_batch_mult * num_beams, input_ids_len)
-            attention_mask = attention_mask.unsqueeze(1).expand(
-                batch_size, effective_batch_mult * num_beams, input_ids_len
-            )
-
-            input_ids = input_ids.contiguous().view(
-                effective_batch_size * num_beams, input_ids_len
-            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
-            attention_mask = attention_mask.contiguous().view(
-                effective_batch_size * num_beams, input_ids_len
+            decoder_ids_len = decoder_ids.shape[-1]
+            decoder_ids = decoder_ids.unsqueeze(1).expand(batch_size, effective_batch_mult * num_beams, decoder_ids_len)
+            decoder_ids = decoder_ids.contiguous().view(
+                effective_batch_size * num_beams, decoder_ids_len
            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)

+
+        # compute encoder outputs once if necessary, and expand to match number of generated sequences
        if self.config.is_encoder_decoder:
-            # create empty decoder_input_ids
-            input_ids = torch.full(
-                (effective_batch_size * num_beams, 1),
-                decoder_start_token_id,
-                dtype=torch.long,
-                device=next(self.parameters()).device,
-            )
-            cur_len = 1
-            batch_idx = self.encoder_outputs_batch_dim_idx
+            encoder = self.get_encoder()
+            encoder_outputs = encoder(encoder_ids, attention_mask=enc_attention_mask)
+
            assert (
-                batch_size == encoder_outputs[0].shape[batch_idx]
-            ), f"expected encoder_outputs[0] to have 1st dimension bs={batch_size}, got {encoder_outputs[0].shape[1]} "
-            expanded_idx = (
+                batch_size == encoder_outputs[0].shape[0]
+            ), f"expected encoder_outputs[0] to have 1st dimension bs={batch_size}, got {encoder_outputs[0].shape[0]} "
+
+            # expand batch_idx to assign correct encoder output for expanded input_ids (due to num_beams > 1 and num_return_sequences > 1)
+            expanded_batch_idxs = (
                torch.arange(batch_size)
                .view(-1, 1)
                .repeat(1, num_beams * effective_batch_mult)
                .view(-1)
-                .to(input_ids.device)
+                .to(decoder_ids.device)
            )
-            encoder_outputs = (encoder_outputs[0].index_select(batch_idx, expanded_idx), *encoder_outputs[1:])
+            encoder_outputs = (encoder_outputs[0].index_select(0, expanded_batch_idxs), *encoder_outputs[1:]) # (x, encoder_states, all_attentions)
+            enc_attention_mask = enc_attention_mask.index_select(0, expanded_batch_idxs)
+            # TODO (Yacine): attention_mask was prepared for the encoder, not the decoder
+
        else:
            encoder_outputs = None
-            cur_len = input_ids.shape[-1]
+            dec_attention_mask = dec_attention_mask.unsqueeze(1).expand(
+                batch_size, effective_batch_mult * num_beams, decoder_ids_len
+            )
+            dec_attention_mask = dec_attention_mask.contiguous().view(
+                effective_batch_size * num_beams, decoder_ids_len
+            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
+        
+        cur_len = decoder_ids.shape[-1]

        if num_beams > 1:
            output = self._generate_beam_search(
-                input_ids,
+                input_ids=decoder_ids,
                cur_len=cur_len,
+                encoder_outputs=encoder_outputs,
+                encoder_attention_mask=enc_attention_mask,
+                decoder_attention_mask=dec_attention_mask,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
@@ -967,22 +1000,24 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
-                decoder_start_token_id=decoder_start_token_id,
                eos_token_id=eos_token_id,
                batch_size=effective_batch_size,
                num_return_sequences=num_return_sequences,
                length_penalty=length_penalty,
                num_beams=num_beams,
                vocab_size=vocab_size,
-                encoder_outputs=encoder_outputs,
-                attention_mask=attention_mask,
+                use_cache=use_cache,
            )
        else:
            output = self._generate_no_beam_search(
-                input_ids,
+                input_ids=decoder_ids,
                cur_len=cur_len,
+                encoder_outputs=encoder_outputs,
+                encoder_attention_mask=enc_attention_mask,
+                decoder_attention_mask=dec_attention_mask,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
@@ -991,13 +1026,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
-                decoder_start_token_id=decoder_start_token_id,
                eos_token_id=eos_token_id,
                batch_size=effective_batch_size,
-                encoder_outputs=encoder_outputs,
-                attention_mask=attention_mask,
+                use_cache=use_cache,
            )

        return output
@@ -1006,6 +1040,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        self,
        input_ids,
        cur_len,
+        encoder_outputs,
+        encoder_attention_mask,
+        decoder_attention_mask,
        max_length,
        min_length,
        do_sample,
@@ -1014,13 +1051,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
+        bad_words_ids,
        bos_token_id,
        pad_token_id,
        eos_token_id,
-        decoder_start_token_id,
        batch_size,
-        encoder_outputs,
-        attention_mask,
+        use_cache,
    ):
        """ Generate sequences for each example without beam search (num_beams == 1).
            All returned sequence are generated independantly.
@@ -1029,17 +1065,18 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        unfinished_sents = input_ids.new(batch_size).fill_(1)
        sent_lengths = input_ids.new(batch_size).fill_(max_length)

-        past = encoder_outputs  # defined for encoder-decoder models, None for decoder-only models
+        past = None if encoder_outputs is None else ((encoder_outputs, encoder_attention_mask), None) # defined for encoder-decoder models, None for decoder-only models

        while cur_len < max_length:
-            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=attention_mask)
+            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=decoder_attention_mask)

            outputs = self(**model_inputs)
            next_token_logits = outputs[0][:, -1, :]

            # if model has past, then set the past variable to speed up decoding
-            if self._do_output_past(outputs):
-                past = outputs[1]
+            if use_cache:
+                _, decoder_cache = outputs[1]
+                past = ((encoder_outputs, encoder_attention_mask), decoder_cache)

            # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
@@ -1048,7 +1085,14 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            if no_repeat_ngram_size > 0:
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
-                banned_tokens = calc_banned_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
+                banned_tokens = calc_banned_ngram_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
+                for batch_idx in range(batch_size):
+                    next_token_logits[batch_idx, banned_tokens[batch_idx]] = -float("inf")
+
+            if bad_words_ids is not None:
+                # calculate a list of banned tokens according to bad words
+                banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
+
                for batch_idx in range(batch_size):
                    next_token_logits[batch_idx, banned_tokens[batch_idx]] = -float("inf")

@@ -1115,6 +1159,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        self,
        input_ids,
        cur_len,
+        encoder_outputs,
+        encoder_attention_mask,
+        decoder_attention_mask,
        max_length,
        min_length,
        do_sample,
@@ -1124,17 +1171,16 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
+        bad_words_ids,
        bos_token_id,
        pad_token_id,
        eos_token_id,
-        decoder_start_token_id,
        batch_size,
        num_return_sequences,
        length_penalty,
        num_beams,
        vocab_size,
-        encoder_outputs,
-        attention_mask,
+        use_cache,
    ):
        """ Generate sequences for each example with beam search.
        """
@@ -1154,19 +1200,20 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)

        # cache compute states
-        past = encoder_outputs  # defined for encoder-decoder models, None for decoder-only models
+        past = None if encoder_outputs is None else ((encoder_outputs, encoder_attention_mask), None) # defined for encoder-decoder models, None for decoder-only models

        # done sentences
        done = [False for _ in range(batch_size)]

        while cur_len < max_length:
-            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=attention_mask)
+            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=decoder_attention_mask)
            outputs = self(**model_inputs)  # (batch_size * num_beams, cur_len, vocab_size)
            next_token_logits = outputs[0][:, -1, :]  # (batch_size * num_beams, vocab_size)

            # if model has past, then set the past variable to speed up decoding
-            if self._do_output_past(outputs):
-                past = outputs[1]
+            if use_cache:
+                _, decoder_cache = outputs[1]
+                past = ((encoder_outputs, encoder_attention_mask), decoder_cache)

            # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
@@ -1190,12 +1237,19 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                num_batch_hypotheses = batch_size * num_beams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
-                banned_batch_tokens = calc_banned_tokens(
+                banned_batch_tokens = calc_banned_ngram_tokens(
                    input_ids, num_batch_hypotheses, no_repeat_ngram_size, cur_len
                )
                for i, banned_tokens in enumerate(banned_batch_tokens):
                    scores[i, banned_tokens] = -float("inf")

+            if bad_words_ids is not None:
+                # calculate a list of banned tokens according to bad words
+                banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
+
+                for i, banned_tokens in enumerate(banned_tokens):
+                    scores[i, banned_tokens] = -float("inf")
+
            assert scores.shape == (batch_size * num_beams, vocab_size), "Shapes of scores: {} != {}".format(
                scores.shape, (batch_size * num_beams, vocab_size)
            )
@@ -1256,14 +1310,13 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                for beam_token_rank, (beam_token_id, beam_token_score) in enumerate(
                    zip(next_tokens[batch_idx], next_scores[batch_idx])
                ):
-                    # get beam and word IDs
+                    # get beam and token IDs
                    beam_id = beam_token_id // vocab_size
                    token_id = beam_token_id % vocab_size

                    effective_beam_id = batch_idx * num_beams + beam_id
-
-                    # add to generated hypotheses if end of sentence
-                    if (eos_token_id is not None) and (token_id.item() is eos_token_id):
+                    # add to generated hypotheses if end of sentence or last iteration
+                    if (eos_token_id is not None) and (token_id.item() == eos_token_id):
                        # if beam_token does not belong to top num_beams tokens, it should not be added
                        is_beam_token_worse_than_top_num_beams = beam_token_rank >= num_beams
                        if is_beam_token_worse_than_top_num_beams:
@@ -1272,7 +1325,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                            input_ids[effective_beam_id].clone(), beam_token_score.item(),
                        )
                    else:
-                        # add next predicted word if it is not eos_token
+                        # add next predicted token if it is not eos_token
                        next_sent_beam.append((beam_token_score, token_id, effective_beam_id))

                    # the beam for next step is full
@@ -1302,7 +1355,6 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
            # re-order batch
            input_ids = input_ids[beam_idx, :]
            input_ids = torch.cat([input_ids, beam_tokens.unsqueeze(1)], dim=-1)
-
            # re-order internal states
            if past is not None:
                past = self._reorder_cache(past, beam_idx)
@@ -1400,7 +1452,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
        return past


-def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len):
+def calc_banned_ngram_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len):
    # Copied from fairseq for no_repeat_ngram in beam_search"""
    if cur_len + 1 < no_repeat_ngram_size:
        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
@@ -1423,6 +1475,42 @@ def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len)
    return banned_tokens


+def calc_banned_bad_words_ids(prev_input_ids, bad_words_ids):
+    banned_tokens = []
+
+    def _tokens_match(prev_tokens, tokens):
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        if len(tokens) > len(prev_input_ids):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+
+        if prev_tokens[-len(tokens) :] == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    for prev_input_ids_slice in prev_input_ids:
+        banned_tokens_slice = []
+
+        for banned_token_seq in bad_words_ids:
+            assert len(banned_token_seq) > 0, "Banned words token sequences {} cannot have an empty list".format(
+                bad_words_ids
+            )
+
+            if _tokens_match(prev_input_ids_slice.tolist(), banned_token_seq[:-1]) is False:
+                # if tokens do not match continue
+                continue
+
+            banned_tokens_slice.append(banned_token_seq[-1])
+
+        banned_tokens.append(banned_tokens_slice)
+
+    return banned_tokens
+
+
 def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
@@ -214,7 +214,7 @@ class GradientAccumulator(object):
            raise ValueError("Expected %s gradients, but got %d" % (len(self._gradients), len(gradients)))

        for accum_gradient, gradient in zip(self._get_replica_gradients(), gradients):
-            if accum_gradient is not None:
+            if accum_gradient is not None and gradient is not None:
                accum_gradient.assign_add(gradient)

        self._accum_steps.assign_add(1)
@@ -241,6 +241,7 @@ class GradientAccumulator(object):
            return (
                gradient.device_map.select_for_current_replica(gradient.values, replica_context)
                for gradient in self._gradients
+                if gradient is not None
            )
        else:
            return self._gradients
@@ -1235,17 +1235,19 @@ class SummarizationPipeline(Pipeline):
            elif self.framework == "tf":
                input_length = tf.shape(inputs["input_ids"])[-1].numpy()

-            if input_length < self.model.config.min_length // 2:
+            min_length = generate_kwargs.get("min_length", self.model.config.min_length)
+            if input_length < min_length // 2:
                logger.warning(
                    "Your min_length is set to {}, but you input_length is only {}. You might consider decreasing min_length manually, e.g. summarizer('...', min_length=10)".format(
-                        self.model.config.min_length, input_length
+                        min_length, input_length
                    )
                )

-            if input_length < self.model.config.max_length:
+            max_length = generate_kwargs.get("max_length", self.model.config.max_length)
+            if input_length < max_length:
                logger.warning(
                    "Your max_length is set to {}, but you input_length is only {}. You might consider decreasing max_length manually, e.g. summarizer('...', max_length=50)".format(
-                        self.model.config.max_length, input_length
+                        max_length, input_length
                    )
                )

@@ -1349,10 +1351,11 @@ class TranslationPipeline(Pipeline):
            elif self.framework == "tf":
                input_length = tf.shape(inputs["input_ids"])[-1].numpy()

-            if input_length > 0.9 * self.model.config.max_length:
+            max_length = generate_kwargs.get("max_length", self.model.config.max_length)
+            if input_length > 0.9 * max_length:
                logger.warning(
                    "Your input_length: {} is bigger than 0.9 * max_length: {}. You might consider increasing your max_length manually, e.g. translator('...', max_length=400)".format(
-                        input_length, self.model.config.max_length
+                        input_length, max_length
                    )
                )

@@ -19,7 +19,7 @@ from .tokenization_roberta import RobertaTokenizer
 # vocab and merges same as roberta
 vocab_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-vocab.json"
 merges_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-merges.txt"
-_all_bart_models = ["bart-large", "bart-large-mnli", "bart-large-cnn"]
+_all_bart_models = ["bart-large", "bart-large-mnli", "bart-large-cnn", "bart-large-xsum"]


 class BartTokenizer(RobertaTokenizer):
@@ -61,14 +61,34 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {

 class T5Tokenizer(PreTrainedTokenizer):
    """
-        SentencePiece based tokenizer. Peculiarities:
+        Constructs an XLNet tokenizer. Based on `SentencePiece <https://github.com/google/sentencepiece>`__ .

-            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
-            - `extra_ids` add a number of extra ids added to the end of the vocabulary for use as sentinels.
-                These tokens are accessible as `<extra_id_{%d}>` where `{%d}` is a number between 0 and extra_ids-1.
-                Extra tokens are indexed from the end of the vocabulary up to beginnning (<extra_id_0> is the last token in the vocabulary)
-                (like in T5 preprocessing
+        This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the methods. Users
+        should refer to the superclass for more information regarding methods.
+
+        Args:
+            vocab_file (:obj:`string`):
+                `SentencePiece <https://github.com/google/sentencepiece>`__ file (generally has a `.spm` extension) that
+                contains the vocabulary necessary to instantiate a tokenizer.
+            eos_token (:obj:`string`, `optional`, defaults to "</s>"):
+                The end of sequence token.
+
+                .. note::
+
+                    When building a sequence using special tokens, this is not the token that is used for the end
+                    of sequence. The token used is the :obj:`sep_token`.
+            unk_token (:obj:`string`, `optional`, defaults to "<unk>"):
+                The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+                token instead.
+            pad_token (:obj:`string`, `optional`, defaults to "<pad>"):
+                The token used for padding, for example when batching sequences of different lengths.
+            extra_ids (:obj:`List[str]`, `optional`, defaults to :obj:`100`):
+                Add a number of extra ids added to the end of the vocabulary for use as sentinels.
+                These tokens are accessible as "<extra_id_{%d}>" where "{%d}" is a number between 0 and extra_ids-1.
+                Extra tokens are indexed from the end of the vocabulary up to beginnning ("<extra_id_0>" is the last token in the vocabulary like in T5 preprocessing
                see: https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117)
+            additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`None`):
+                Additional special tokens used by the tokenizer.
    """

    vocab_files_names = VOCAB_FILES_NAMES
@@ -27,7 +27,9 @@ from .utils import CACHE_DIR, require_torch, slow, torch_device
 if is_torch_available():
    import torch
    from transformers import (
+        AutoModel,
        AutoModelForSequenceClassification,
+        AutoTokenizer,
        BartModel,
        BartForConditionalGeneration,
        BartForSequenceClassification,
@@ -183,6 +185,15 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase):
    def test_inputs_embeds(self):
        pass

+    def test_tiny_model(self):
+        model_name = "sshleifer/bart-tiny-random"
+        tiny = AutoModel.from_pretrained(model_name)  # same vocab size
+        tok = AutoTokenizer.from_pretrained(model_name)  # same tokenizer
+        inputs_dict = tok.batch_encode_plus(["Hello my friends"], return_tensors="pt")
+
+        with torch.no_grad():
+            tiny(**inputs_dict)
+

@require_torch
 class BartHeadTests(unittest.TestCase):
@@ -450,6 +461,38 @@ class BartModelIntegrationTests(unittest.TestCase):
            model = BartModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
            self.assertIsNotNone(model)

+    @slow
+    def test_xsum_summarization_same_as_fairseq(self):
+        model = BartForConditionalGeneration.from_pretrained("bart-large-xsum").to(torch_device)
+        tok = BartTokenizer.from_pretrained("bart-large")
+
+        PGE_ARTICLE = """ PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."""
+        EXPECTED_SUMMARY = "California's largest power company has begun shutting off power to tens of thousands of homes and businesses in the state."
+        dct = tok.batch_encode_plus([PGE_ARTICLE], max_length=1024, pad_to_max_length=True, return_tensors="pt",)
+
+        hypotheses_batch = model.generate(
+            input_ids=dct["input_ids"].to(torch_device),
+            attention_mask=dct["attention_mask"].to(torch_device),
+            num_beams=2,
+            max_length=62,
+            min_length=11,
+            length_penalty=1.0,
+            no_repeat_ngram_size=3,
+            early_stopping=True,
+            decoder_start_token_id=model.config.eos_token_id,
+        )
+
+        decoded = [
+            tok.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in hypotheses_batch
+        ]
+        self.assertEqual(EXPECTED_SUMMARY, decoded[0])
+
+    def test_xsum_config_generation_params(self):
+        config = BartConfig.from_pretrained("bart-large-xsum")
+        expected_params = dict(num_beams=6, do_sample=False, early_stopping=True, length_penalty=1.0)
+        config_params = {k: getattr(config, k, "MISSING") for k, v in expected_params.items()}
+        self.assertDictEqual(expected_params, config_params)
+
    @slow
    def test_cnn_summarization_same_as_fairseq(self):
        hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
@@ -641,14 +641,14 @@ class ModelTesterMixin:
                with self.assertRaises(AssertionError):
                    model.generate(do_sample=True, max_length=5)
                # batch_size = 1
-                self._check_generated_tokens(model.generate(input_ids, do_sample=True))
+                self._check_generated_ids(model.generate(input_ids, do_sample=True))
                # batch_size = 1, num_beams > 1
-                self._check_generated_tokens(model.generate(input_ids, do_sample=True, num_beams=3))
+                self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=3))
            else:
                # batch_size = 1
-                self._check_generated_tokens(model.generate(do_sample=True, max_length=5))
+                self._check_generated_ids(model.generate(do_sample=True, max_length=5))
                # batch_size = 1, num_beams > 1
-                self._check_generated_tokens(model.generate(do_sample=True, max_length=5, num_beams=3))
+                self._check_generated_ids(model.generate(do_sample=True, max_length=5, num_beams=3))

            with self.assertRaises(AssertionError):
                # generating multiple sequences when greedy no beam generation
@@ -660,24 +660,52 @@ class ModelTesterMixin:
                model.generate(input_ids, do_sample=False, num_return_sequences=3, num_beams=2)

            # batch_size > 1, sample
-            self._check_generated_tokens(model.generate(input_ids, do_sample=True, num_return_sequences=3))
+            self._check_generated_ids(model.generate(input_ids, do_sample=True, num_return_sequences=3))
            # batch_size > 1, greedy
-            self._check_generated_tokens(model.generate(input_ids, do_sample=False))
+            self._check_generated_ids(model.generate(input_ids, do_sample=False))

            # batch_size > 1, num_beams > 1, sample
-            self._check_generated_tokens(
-                model.generate(input_ids, do_sample=True, num_beams=3, num_return_sequences=3,)
-            )
+            self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=3, num_return_sequences=3,))
            # batch_size > 1, num_beams > 1, greedy
-            self._check_generated_tokens(
-                model.generate(input_ids, do_sample=False, num_beams=3, num_return_sequences=3)
-            )
+            self._check_generated_ids(model.generate(input_ids, do_sample=False, num_beams=3, num_return_sequences=3))

-    def _check_generated_tokens(self, output_ids):
+            # check bad words tokens language generation
+            bad_words_ids = [
+                ids_tensor((1, 1), self.model_tester.vocab_size).squeeze(-1).tolist(),
+                ids_tensor((2, 1), self.model_tester.vocab_size).squeeze(-1).tolist(),
+            ]
+
+            # sampling
+            output_tokens = model.generate(
+                input_ids, do_sample=True, bad_words_ids=bad_words_ids, num_return_sequences=3
+            )
+            generated_ids = output_tokens[:, input_ids.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.tolist(), bad_words_ids))
+
+            # beam search
+            output_tokens = model.generate(
+                input_ids, do_sample=False, bad_words_ids=bad_words_ids, num_beams=3, num_return_sequences=3
+            )
+            generated_ids = output_tokens[:, input_ids.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.tolist(), bad_words_ids))
+
+    def _check_generated_ids(self, output_ids):
        for token_id in output_ids[0].tolist():
            self.assertGreaterEqual(token_id, 0)
            self.assertLess(token_id, self.model_tester.vocab_size)

+    def _check_match_tokens(self, generated_ids, bad_words_ids):
+        # for all bad word tokens
+        for bad_word_ids in bad_words_ids:
+            # for all slices in batch
+            for generated_ids_slice in generated_ids:
+                # for all word idx
+                for i in range(len(bad_word_ids), len(generated_ids_slice)):
+                    # if tokens match
+                    if generated_ids_slice[i - len(bad_word_ids) : i] == bad_word_ids:
+                        return True
+        return False
+

 global_rng = random.Random()

@@ -162,6 +162,10 @@ class TFModelTesterMixin:
            pt_inputs_dict = dict(
                (name, torch.from_numpy(key.numpy()).to(torch.long)) for name, key in inputs_dict.items()
            )
+            # need to rename encoder-decoder "inputs" for PyTorch
+            if "inputs" in pt_inputs_dict and self.is_encoder_decoder:
+                pt_inputs_dict["input_ids"] = pt_inputs_dict.pop("inputs")
+
            with torch.no_grad():
                pto = pt_model(**pt_inputs_dict)
            tfo = tf_model(inputs_dict, training=False)
@@ -201,6 +205,10 @@ class TFModelTesterMixin:
            pt_inputs_dict = dict(
                (name, torch.from_numpy(key.numpy()).to(torch.long)) for name, key in inputs_dict.items()
            )
+            # need to rename encoder-decoder "inputs" for PyTorch
+            if "inputs" in pt_inputs_dict and self.is_encoder_decoder:
+                pt_inputs_dict["input_ids"] = pt_inputs_dict.pop("inputs")
+
            with torch.no_grad():
                pto = pt_model(**pt_inputs_dict)
            tfo = tf_model(inputs_dict)
@@ -223,7 +231,7 @@ class TFModelTesterMixin:
        if self.is_encoder_decoder:
            input_ids = {
                "decoder_input_ids": tf.keras.Input(batch_shape=(2, 2000), name="decoder_input_ids", dtype="int32"),
-                "input_ids": tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32"),
+                "inputs": tf.keras.Input(batch_shape=(2, 2000), name="inputs", dtype="int32"),
            }
        else:
            input_ids = tf.keras.Input(batch_shape=(2, 2000), name="input_ids", dtype="int32")
@@ -259,7 +267,7 @@ class TFModelTesterMixin:
            outputs_dict = model(inputs_dict)

            inputs_keywords = copy.deepcopy(inputs_dict)
-            input_ids = inputs_keywords.pop("input_ids" if not self.is_encoder_decoder else "decoder_input_ids", None,)
+            input_ids = inputs_keywords.pop("input_ids" if not self.is_encoder_decoder else "inputs", None,)
            outputs_keywords = model(input_ids, **inputs_keywords)

            output_dict = outputs_dict[0].numpy()
@@ -395,9 +403,9 @@ class TFModelTesterMixin:
            input_ids = inputs_dict["input_ids"]
            del inputs_dict["input_ids"]
        else:
-            encoder_input_ids = inputs_dict["input_ids"]
+            encoder_input_ids = inputs_dict["inputs"]
            decoder_input_ids = inputs_dict["decoder_input_ids"]
-            del inputs_dict["input_ids"]
+            del inputs_dict["inputs"]
            del inputs_dict["decoder_input_ids"]

        for model_class in self.all_model_classes:
@@ -415,7 +423,7 @@ class TFModelTesterMixin:
    def test_lm_head_model_random_generate(self):

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        input_ids = inputs_dict["input_ids"]
+        input_ids = inputs_dict["input_ids"] if "input_ids" in inputs_dict else inputs_dict["inputs"]

        if self.is_encoder_decoder:
            config.output_past = True  # needed for Bart TODO: might have to update for other encoder-decoder models
@@ -427,14 +435,14 @@ class TFModelTesterMixin:
                with self.assertRaises(AssertionError):
                    model.generate(do_sample=True, max_length=5)
                # batch_size = 1
-                self._check_generated_tokens(model.generate(input_ids, do_sample=True))
+                self._check_generated_ids(model.generate(input_ids, do_sample=True))
                # batch_size = 1, num_beams > 1
-                self._check_generated_tokens(model.generate(input_ids, do_sample=True, num_beams=3))
+                self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=3))
            else:
                # batch_size = 1
-                self._check_generated_tokens(model.generate(do_sample=True, max_length=5))
+                self._check_generated_ids(model.generate(do_sample=True, max_length=5))
                # batch_size = 1, num_beams > 1
-                self._check_generated_tokens(model.generate(do_sample=True, max_length=5, num_beams=3))
+                self._check_generated_ids(model.generate(do_sample=True, max_length=5, num_beams=3))

            with self.assertRaises(AssertionError):
                # generating multiple sequences when greedy no beam generation
@@ -446,24 +454,52 @@ class TFModelTesterMixin:
                model.generate(input_ids, do_sample=False, num_return_sequences=3, num_beams=2)

            # batch_size > 1, sample
-            self._check_generated_tokens(model.generate(input_ids, do_sample=True, num_return_sequences=3))
+            self._check_generated_ids(model.generate(input_ids, do_sample=True, num_return_sequences=3))
            # batch_size > 1, greedy
-            self._check_generated_tokens(model.generate(input_ids, do_sample=False))
+            self._check_generated_ids(model.generate(input_ids, do_sample=False))

            # batch_size > 1, num_beams > 1, sample
-            self._check_generated_tokens(
-                model.generate(input_ids, do_sample=True, num_beams=3, num_return_sequences=3,)
-            )
+            self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=3, num_return_sequences=3,))
            # batch_size > 1, num_beams > 1, greedy
-            self._check_generated_tokens(
-                model.generate(input_ids, do_sample=False, num_beams=3, num_return_sequences=3)
-            )
+            self._check_generated_ids(model.generate(input_ids, do_sample=False, num_beams=3, num_return_sequences=3))

-    def _check_generated_tokens(self, output_ids):
+            # check bad words tokens language generation
+            bad_words_ids = [
+                tf.squeeze(ids_tensor((1, 1), self.model_tester.vocab_size), -1).numpy().tolist(),
+                tf.squeeze(ids_tensor((2, 1), self.model_tester.vocab_size), -1).numpy().tolist(),
+            ]
+
+            # sampling
+            output_tokens = model.generate(
+                input_ids, do_sample=True, bad_words_ids=bad_words_ids, num_return_sequences=3
+            )
+            generated_ids = output_tokens[:, input_ids.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
+
+            # beam search
+            output_tokens = model.generate(
+                input_ids, do_sample=False, bad_words_ids=bad_words_ids, num_beams=3, num_return_sequences=3
+            )
+            generated_ids = output_tokens[:, input_ids.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
+
+    def _check_generated_ids(self, output_ids):
        for token_id in output_ids[0].numpy().tolist():
            self.assertGreaterEqual(token_id, 0)
            self.assertLess(token_id, self.model_tester.vocab_size)

+    def _check_match_tokens(self, generated_ids, bad_words_ids):
+        # for all bad word tokens
+        for bad_word_ids in bad_words_ids:
+            # for all slices in batch
+            for generated_ids_slice in generated_ids:
+                # for all word idx
+                for i in range(len(bad_word_ids), len(generated_ids_slice)):
+                    # if tokens match
+                    if generated_ids_slice[i - len(bad_word_ids) : i] == bad_word_ids:
+                        return True
+        return False
+

 def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
    """Creates a random int32 tensor of the shape within the vocab size."""
Author	SHA1	Message	Date
yjernite	d1406e9a63	resolved_decode_with_prefix	2020-04-02 14:19:12 -04:00
yjernite	76f5cd8b7c	resolved_decode_with_prefix	2020-04-02 14:14:11 -04:00
yjernite	ae98046330	decode_with_prefix	2020-04-02 14:09:31 -04:00
Mark Kockerbeck	1b10159950	Adding should_continue check for retraining (#3509 )	2020-04-02 14:07:08 -04:00
Patrick von Platen	390c128592	[Encoder-Decoder] Force models outputs to always have batch_size as their first dim (#3536 ) * solve conflicts * improve comments	2020-04-02 15:18:33 +02:00
Patrick von Platen	ab5d06a094	[T5, examples] replace heavy t5 models with tiny random models (#3556 ) * replace heavy t5 models with tiny random models as was done by sshleifer * fix isort	2020-04-02 12:34:05 +02:00
Patrick von Platen	a4ee4da18a	[T5, TF 2.2] change tf t5 argument naming (#3547 ) * change tf t5 argument naming for TF 2.2 * correct bug in testing	2020-04-01 22:04:20 +02:00
Patrick von Platen	06dd597552	fix bug in warnings T5 pipelines (#3545 )	2020-04-01 21:59:12 +02:00
Anirudh Srinivasan	9de9ceb6c5	Correct output shape for Bert NSP models in docs (#3482 )	2020-04-01 15:04:38 -04:00
Patrick von Platen	b815edf69f	[T5, Testst] Add extensive hard-coded integration tests and make sure PT and TF give equal results (#3550 ) * add some t5 integration tests * finish summarization and translation integration tests for T5 - results loook good * add tf test * fix == vs is bug * fix tf beam search error and make tf t5 tests pass	2020-04-01 18:01:33 +02:00
HUSEIN ZOLKEPLI	8538ce9044	Add tiny-bert-bahasa-cased model card (#3567 ) * add bert bahasa readme * update readme * update readme * added xlnet * added tiny-bert and fix xlnet readme	2020-04-01 07:15:00 -04:00
Manuel Romero	c1a6252be1	Create model card (#3557 ) Create model card for: distilbert-multi-finetuned-for-xqua-on-tydiqa	2020-04-01 07:14:23 -04:00
Julien Chaumond	50e15c825c	Tokenizers: Start cleaning examples a little (#3455 ) * Start cleaning examples * Fixup	2020-04-01 07:13:40 -04:00
Patrick von Platen	b38d552a92	[Generate] Add bad words list argument to the generate function (#3367 ) * add bad words list * make style * add bad_words_tokens * make style * better naming * make style * fix typo	2020-03-31 18:42:31 +02:00
Patrick von Platen	ae6834e028	[Examples] Clean summarization and translation example testing files for T5 and Bart (#3514 ) * fix conflicts * add model size argument to summarization * correct wrong import * fix isort * correct imports * other isort make style * make style	2020-03-31 17:54:13 +02:00
Manuel Romero	0373b60c4c	Update README.md (#3552 ) - Show that the last uploaded version was trained on more data (custom_license files)	2020-03-31 10:40:34 -04:00
Patrick von Platen	83d1fbcff6	[Docs] Add usage examples for translation and summarization (#3538 )	2020-03-31 09:36:03 -04:00
Patrick von Platen	55bcae7f25	remove useless and confusing lm_labels line (#3531 )	2020-03-31 09:32:25 -04:00
Patrick von Platen	42e1e3c67f	Update usage doc regarding generate fn (#3504 )	2020-03-31 09:31:46 -04:00
Patrick von Platen	57b0fab692	Add better explanation to check `docs` locally. (#3459 )	2020-03-31 09:30:17 -04:00
Manuel Romero	a8d4dff0a1	Update README.md (#3470 ) Fix typo	2020-03-31 08:01:09 -04:00
Manuel Romero	4a5663568f	Create card for the model: GPT-2-finetuned-covid-bio-medrxiv (#3453 )	2020-03-31 08:01:03 -04:00
Branden Chan	bbedb59675	Create README.md (#3393 ) * Create README.md * Update README.md	2020-03-31 08:00:35 -04:00
Manuel Romero	c2cf192943	Add link to 16 POS tags model (#3465 )	2020-03-31 08:00:00 -04:00
Gabriele Sarti	c82ef72158	Added CovidBERT-NLI model card (#3477 )	2020-03-31 07:59:49 -04:00
Manuel Romero	b48a1f08c1	Add text shown in example of usage (#3464 )	2020-03-31 07:59:36 -04:00
Manuel Romero	99833a9cbf	Create model card (#3487 )	2020-03-31 07:59:22 -04:00
Sho Arora	ebceeeacda	Add electra and alectra model cards (#3524 )	2020-03-31 07:58:48 -04:00
Leandro von Werra	a6c4ee27fd	Add model cards (#3537 ) * feat: add model card bert-imdb * feat: add model card gpt2-imdb-pos * feat: add model card gpt2-imdb	2020-03-31 07:54:45 -04:00
Ethan Perez	e5c393dceb	[Bug fix] Using loaded checkpoint with --do_predict (instead of… (#3437 ) * Using loaded checkpoint with --do_predict Without this fix, I'm getting near-random validation performance for a trained model, and the validation performance differs per validation run. I think this happens since the `model` variable isn't set with the loaded checkpoint, so I'm using a randomly initialized model. Looking at the model activations, they differ each time I run evaluation (but they don't with this fix). * Update checkpoint loading * Fixing model loading	2020-03-30 17:06:08 -04:00
Sam Shleifer	8deff3acf2	[bart-tiny-random] Put a 5MB model on S3 to allow faster exampl… (#3488 )	2020-03-30 12:28:27 -04:00
dougianandIoannis Douratsos	1f72865726	[BART] Update encoder and decoder on set_input_embedding (#3501 ) Co-authored-by: Ioannis Douratsos <ioannisd@amazon.com>	2020-03-30 12:20:37 -04:00
Julien Chaumond	cc598b312b	[InputExample] Unfreeze for now, cf. #3423	2020-03-30 10:41:49 -04:00
Julien PluandJulien Plu	d38bbb225f	Update the NER TF script (#3511 ) * Update the NER TF script to remove the softmax and make the pad token label id to -1 * Reformat the quality and style Co-authored-by: Julien Plu <julien.plu@adevinta.com>	2020-03-30 09:50:12 -04:00
LysandreJik	eff757f2e3	Re-pin isort version	2020-03-30 09:00:47 -04:00
LysandreJik	a009d751c2	Un-pin isort for v2.7.0 pypi	2020-03-30 08:55:10 -04:00
LysandreJik	6f5a12a583	Release: v2.7.0	2020-03-30 08:49:24 -04:00
Patrick von Platen	296252c49e	fix lm lables in docstring (#3529 )	2020-03-30 14:26:24 +02:00
Patrick von Platen	75ec6c9e3a	[T5] make decoder input ids optional for t5 training (#3521 ) * make decoder input ids optional for t5 training * lm_lables should not be shifted in t5 * add tests * finish shift right functionality for PT T5 * move shift right to correct class * cleaner code * replace -100 values with pad token id * add assert statement * remove unnecessary for loop * make style	2020-03-30 13:45:26 +02:00
Patrick von Platen	5b44e0a31b	[T5] Add training documenation (#3507 ) * Add clear description of how to train T5 * correct docstring in T5 * correct typo * correct docstring format * update t5 model docs * implement collins feedback * fix typo and add more explanation for sentinal tokens * delete unnecessary todos	2020-03-30 13:35:53 +02:00
Sam Shleifer	33ef7002e1	[Docs] examples/summarization/bart: Simplify CNN/DM preprocessi… (#3516 )	2020-03-29 13:25:42 -04:00
Sam Shleifer	f6a23d1911	[BART] add bart-large-xsum weights (#3422 )	2020-03-29 10:51:13 -04:00
Stefan Schweter	601ac5b1dc	[model_cards]: use MIT license for all dbmdz models	2020-03-27 18:06:25 -04:00
Patrick von Platen	17dceae7a1	Fix circle ci flaky fail of wmt example (#3485 ) * force bleu * fix wrong file name * rename file * different filenames for each example test * test files should clean up after themselves * test files should clean up after themselves * do not force bleu * correct typo * fix isort	2020-03-27 13:01:28 -04:00
Patrick von Platen	00ea100e96	add summarization and translation to notebook (#3478 )	2020-03-27 11:05:37 -04:00
Funtowicz Morgan	b08259a120	run_ner.py / bert-base-multilingual-cased can output empty tokens (#2991 ) * Use tokenizer.num_added_tokens to count number of added special_tokens instead of hardcoded numbers. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * run_ner.py - Do not add a label to the labels_ids if word_tokens is empty. This can happen when using bert-base-multilingual-cased with an input containing an unique space. In this case, the tokenizer will output just an empty word_tokens thus leading to an non-consistent behavior over the labels_ids tokens adding one more tokens than tokens vector. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>	2020-03-27 10:59:55 -04:00
Patrick von Platen	f4f4946836	Rename `t5-large` to `t5-base` in README.md	2020-03-27 15:57:58 +01:00
Patrick von Platen	fa9af2468a	Add T5 to docs (#3461 ) * add t5 docs basis * improve docs * add t5 docs * improve t5 docstring * add t5 tokenizer docstring * finish docstring * make style * add pretrained models * correct typo * make examples work * finalize docs	2020-03-27 10:57:16 -04:00
Lysandre Debut	ff80b73157	Add option to choose T5 model size. (#3480 ) T5-small in test isort	2020-03-27 15:56:59 +01:00
LysandreJik	e2c05f06ef	Correct indentation in docstring For some reason Sphinx extremely dislikes this and crashes.	2020-03-27 09:28:52 -04:00