ByT5 model (#11971)

* allow tf to use uneven num of layers

* add tokenizer

* finish docs

* finish docs

* Apply suggestions from code review

* include in index

* finish

* Update docs/source/model_doc/byt5.rst
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* apply sylvais suggestions

* make style
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

README.md

@@ -206,6 +206,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Blenderbot](https://huggingface.co/transformers/model_doc/blenderbot.html)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/transformers/model_doc/blenderbot_small.html)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BORT](https://huggingface.co/transformers/model_doc/bort.html)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
+1. **[ByT5](https://huggingface.co/transformers/model_doc/byt5.html)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/transformers/model_doc/camembert.html)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CLIP](https://huggingface.co/transformers/model_doc/clip.html)** from (OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
 1. **[ConvBERT](https://huggingface.co/transformers/model_doc/convbert.html)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

docs/source/model_doc/byt5.rst

+.. 
+    Copyright 2021 The HuggingFace Team. All rights reserved.
+
+    Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+    the License. You may obtain a copy of the License at
+
+        http://www.apache.org/licenses/LICENSE-2.0
+
+    Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+    an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+    specific language governing permissions and limitations under the License.
+
+ByT5
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The ByT5 model was presented in `ByT5: Towards a token-free future with pre-trained byte-to-byte models
+<https://arxiv.org/abs/2105.13626>`_ by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir
+Kale, Adam Roberts, Colin Raffel.
+
+The abstract from the paper is the following:
+
+*Most widely-used pre-trained language models operate on sequences of tokens corresponding to word or subword units.
+Encoding text as a sequence of tokens requires a tokenizer, which is typically created as an independent artifact from
+the model. Token-free models that instead operate directly on raw text (bytes or characters) have many benefits: they
+can process text in any language out of the box, they are more robust to noise, and they minimize technical debt by
+removing complex and error-prone text preprocessing pipelines. Since byte or character sequences are longer than token
+sequences, past work on token-free models has often introduced new model architectures designed to amortize the cost of
+operating directly on raw text. In this paper, we show that a standard Transformer architecture can be used with
+minimal modifications to process byte sequences. We carefully characterize the trade-offs in terms of parameter count,
+training FLOPs, and inference speed, and show that byte-level models are competitive with their token-level
+counterparts. We also demonstrate that byte-level models are significantly more robust to noise and perform better on
+tasks that are sensitive to spelling and pronunciation. As part of our contribution, we release a new set of
+pre-trained byte-level Transformer models based on the T5 architecture, as well as all code and data used in our
+experiments.*
+
+This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__. The original code can be
+found `here <https://github.com/google-research/byt5>`__.
+
+
+ByT5's architecture is based on the T5 model, so one can refer to :doc:`T5's documentation page <t5>`.
+
+
+Example
+_______________________________________________________________________________________________________________________
+
+ByT5 works on raw UTF-8 bytes, so it can be used without a tokenizer:
+
+.. code-block::
+
+    from transformers import T5ForConditionalGeneration
+    import torch
+
+    model = T5ForConditionalGeneration.from_pretrained('google/byt5-small')
+
+    input_ids = torch.tensor([list("Life is like a box of chocolates.".encode("utf-8"))]) + 3  # add 3 for special tokens
+    labels = torch.tensor([list("La vie est comme une boîte de chocolat.".encode("utf-8"))]) + 3  # add 3 for special tokens
+
+    loss = model(input_ids, labels=labels).loss # forward pass
+
+
+For batched inference and training it is however recommended to make use of the tokenizer:
+
+.. code-block::
+
+    from transformers import T5ForConditionalGeneration, AutoTokenizer
+
+    model = T5ForConditionalGeneration.from_pretrained('google/byt5-small')
+    tokenizer = AutoTokenizer.from_pretrained('google/byt5-small')
+
+    model_inputs = tokenizer(["Life is like a box of chocolates.", "Today is Monday."], padding="longest", return_tensors="pt")
+    labels = tokenizer(["La vie est comme une boîte de chocolat.", "Aujourd'hui c'est lundi."], padding="longest", return_tensors="pt").input_ids
+
+    loss = model(**model_inputs, labels=labels).loss # forward pass
+
+ByT5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ByT5Tokenizer
+
+See :class:`~transformers.ByT5Tokenizer` for all details.

src/transformers/__init__.py

@@ -165,6 +165,7 @@ _import_structure = {
         "BlenderbotSmallConfig",
         "BlenderbotSmallTokenizer",
     ],
+    "models.byt5": ["ByT5Tokenizer"],
     "models.camembert": ["CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "CamembertConfig"],
     "models.clip": [
         "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -1636,6 +1637,7 @@ if TYPE_CHECKING:
         BlenderbotSmallConfig,
         BlenderbotSmallTokenizer,
     )
+    from .models.byt5 import ByT5Tokenizer
     from .models.camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, CamembertConfig
     from .models.clip import (
         CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,

src/transformers/models/auto/tokenization_auto.py

@@ -27,6 +27,7 @@ from ..bert_japanese.tokenization_bert_japanese import BertJapaneseTokenizer
 from ..bertweet.tokenization_bertweet import BertweetTokenizer
 from ..blenderbot.tokenization_blenderbot import BlenderbotTokenizer
 from ..blenderbot_small.tokenization_blenderbot_small import BlenderbotSmallTokenizer
+from ..byt5.tokenization_byt5 import ByT5Tokenizer
 from ..convbert.tokenization_convbert import ConvBertTokenizer
 from ..ctrl.tokenization_ctrl import CTRLTokenizer
 from ..deberta.tokenization_deberta import DebertaTokenizer
@@ -287,6 +288,7 @@ TOKENIZER_MAPPING = OrderedDict(
 NO_CONFIG_TOKENIZER = [
     BertJapaneseTokenizer,
     BertweetTokenizer,
+    ByT5Tokenizer,
     CpmTokenizer,
     HerbertTokenizer,
     HerbertTokenizerFast,

src/transformers/models/byt5/__init__.py

+# flake8: noqa
+# 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.
+
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...file_utils import _BaseLazyModule
+
+
+_import_structure = {
+    "tokenization_byt5": ["ByT5Tokenizer"],
+}
+
+
+if TYPE_CHECKING:
+    from .tokenization_byt5 import ByT5Tokenizer
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/byt5/convert_byt5_original_tf_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2018 The T5 authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert T5 checkpoint."""
+
+
+import argparse
+
+from transformers import T5Config, T5ForConditionalGeneration, load_tf_weights_in_t5
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = T5Config.from_json_file(config_file)
+    print(f"Building PyTorch model from configuration: {config}")
+    model = T5ForConditionalGeneration(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_t5(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print(f"Save PyTorch model to {pytorch_dump_path}")
+    model.save_pretrained(pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
+    )
+    parser.add_argument(
+        "--config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The config json file corresponding to the pre-trained T5 model. \n"
+        "This specifies the model architecture.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.config_file, args.pytorch_dump_path)

src/transformers/models/byt5/tokenization_byt5.py

+# coding=utf-8
+# Copyright 2021 T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Tokenization class for model ByT5."""
+
+
+import re
+import warnings
+from typing import Dict, List, Optional, Tuple
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class ByT5Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a ByT5 tokenizer. ByT5 simply uses raw bytes utf-8 encoding.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`"</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:`str`, `optional`, defaults to :obj:`"<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:`str`, `optional`, defaults to :obj:`"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        extra_ids (:obj:`int`, `optional`, defaults to 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 beginning ("<extra_id_0>" is the last token in the vocabulary
+            like in ByT5 preprocessing see `here
+            <https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117>`__).
+        additional_special_tokens (:obj:`List[str]`, `optional`):
+            Additional special tokens used by the tokenizer.
+    """
+
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        eos_token="</s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        extra_ids=125,
+        additional_special_tokens=None,
+        **kwargs
+    ) -> None:
+        # Add extra_ids to the special token list
+        if extra_ids > 0 and additional_special_tokens is None:
+            additional_special_tokens = [f"<extra_id_{i}>" for i in range(extra_ids)]
+        elif extra_ids > 0 and additional_special_tokens is not None:
+            # Check that we have the right number of extra_id special tokens
+            extra_tokens = len(set(filter(lambda x: bool("extra_id" in str(x)), additional_special_tokens)))
+            if extra_tokens != extra_ids:
+                raise ValueError(
+                    f"Both extra_ids ({extra_ids}) and additional_special_tokens ({additional_special_tokens}) are provided to ByT5Tokenizer. "
+                    "In this case the additional_special_tokens must include the extra_ids tokens"
+                )
+
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+
+        super().__init__(
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            extra_ids=extra_ids,
+            additional_special_tokens=additional_special_tokens,
+            **kwargs,
+        )
+
+        # define special tokens dict
+        self.special_tokens_encoder: Dict[int, str] = {
+            self.pad_token: 0,
+            self.eos_token: 1,
+            self.unk_token: 2,
+        }
+        self.special_tokens_decoder: Dict[str, int] = {v: k for k, v in self.special_tokens_encoder.items()}
+
+        self._num_special_tokens = len(self.special_tokens_encoder)
+        self._utf_vocab_size = 2 ** 8  # utf is 8 bits
+        self._extra_ids = extra_ids
+
+    @property
+    def vocab_size(self):
+        return self._utf_vocab_size + self._num_special_tokens + self._extra_ids
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        # normal case: some special tokens
+        if token_ids_1 is None:
+            return ([0] * len(token_ids_0)) + [1]
+        return ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+
+    def _add_eos_if_not_present(self, token_ids: List[int]) -> List[int]:
+        """Do not add eos again if user already added it."""
+        if len(token_ids) > 0 and token_ids[-1] == self.eos_token_id:
+            warnings.warn(
+                f"This sequence already has {self.eos_token}. In future versions this behavior may lead to duplicated eos tokens being added."
+            )
+            return token_ids
+        else:
+            return token_ids + [self.eos_token_id]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. ByT5 does not
+        make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of zeros.
+        """
+        eos = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + eos) * [0]
+        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A sequence has the following format:
+
+        - single sequence: ``X </s>``
+        - pair of sequences: ``A </s> B </s>``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        token_ids_0 = self._add_eos_if_not_present(token_ids_0)
+        if token_ids_1 is None:
+            return token_ids_0
+        else:
+            token_ids_1 = self._add_eos_if_not_present(token_ids_1)
+            return token_ids_0 + token_ids_1
+
+    def _tokenize(self, text: str) -> List[str]:
+        """Take as input a string and return a list of strings (tokens) for words/sub-words"""
+
+        def _sub_tokenize(sub_text):
+            character_list = list(sub_text)
+            utf_tokens_lists = [list(char.encode("utf-8")) for char in character_list]
+            sub_tokens = [chr(utf_token) for utf_tokens in utf_tokens_lists for utf_token in utf_tokens]
+            return sub_tokens
+
+        # split on special characters
+        pattern = f"({'|'.join(self.special_tokens_encoder.keys())})"
+        sub_texts = list(filter(None, re.split(pattern, text)))
+        tokens = []
+        for sub_text in sub_texts:
+            if sub_text in self.special_tokens_encoder.keys():
+                tokens += [sub_text]
+            else:
+                tokens += _sub_tokenize(sub_text)
+
+        return tokens
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        if token.startswith("<extra_id_"):
+            match = re.match(r"<extra_id_(\d+)>", token)
+            num = int(match.group(1))
+            token_id = self.vocab_size - num - 1
+        elif token in self.special_tokens_encoder:
+            token_id = self.special_tokens_encoder[token]
+        elif len(token) > 1:
+            # token of length > 1 must be newly added tokens => set them to unk token
+            token_id = self.unk_token_id
+        else:
+            token_id = ord(token) + self._num_special_tokens
+        return token_id
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        if index < self._num_special_tokens:
+            token = self.special_tokens_decoder[index]
+        elif index < self._utf_vocab_size + self._num_special_tokens:
+            token = chr(index - self._num_special_tokens)
+        else:
+            token = f"<extra_id_{self.vocab_size - 1 - index}>"
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+
+        def _convert_sub_string(sub_chars):
+            byte_string = bytes([ord(char) for char in sub_chars])
+            return byte_string.decode("utf-8", errors="ignore")
+
+        string = ""
+        sub_chars = []
+        for token in tokens:
+            # if is special token
+            if len(token) > 1:
+                string += _convert_sub_string(sub_chars)
+                string += token
+                sub_chars = []
+            else:
+                sub_chars.append(token)
+
+        # add remaining chars
+        string += _convert_sub_string(sub_chars)
+
+        return string
+
+    # ByT5Tokenizer has no vocab file
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        return ()

src/transformers/models/t5/modeling_tf_t5.py

@@ -1092,6 +1092,7 @@ class TFT5Model(TFT5PreTrainedModel):
 
         decoder_config = copy.deepcopy(config)
         decoder_config.is_decoder = True
+        decoder_config.num_layers = config.num_decoder_layers
         self.decoder = TFT5MainLayer(decoder_config, embed_tokens, name="decoder")
 
     def get_encoder(self):
@@ -1255,6 +1256,7 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
 
         decoder_config = copy.deepcopy(config)
         decoder_config.is_decoder = True
+        decoder_config.num_layers = config.num_decoder_layers
         self.decoder = TFT5MainLayer(decoder_config, embed_tokens, name="decoder")
 
         if not config.tie_word_embeddings:

tests/test_modeling_t5.py

@@ -30,7 +30,7 @@ from .test_modeling_common import ModelTesterMixin, ids_tensor
 if is_torch_available():
     import torch
 
-    from transformers import T5Config, T5EncoderModel, T5ForConditionalGeneration, T5Model, T5Tokenizer
+    from transformers import ByT5Tokenizer, T5Config, T5EncoderModel, T5ForConditionalGeneration, T5Model, T5Tokenizer
     from transformers.models.t5.modeling_t5 import T5_PRETRAINED_MODEL_ARCHIVE_LIST
 
 
@@ -846,6 +846,30 @@ class T5ModelIntegrationTests(unittest.TestCase):
         EXPECTED_SCORE = -59.0293
         self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
 
+    @slow
+    def test_small_byt5_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.9.1
+
+        >>> path_to_byt5_small_checkpoint = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_tf_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = t5.data.ByteVocabulary()
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = T5ForConditionalGeneration.from_pretrained("google/byt5-small").to(torch_device)
+        tokenizer = ByT5Tokenizer.from_pretrained("google/byt5-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="pt").input_ids
+        labels = tokenizer("Hi I am", return_tensors="pt").input_ids
+
+        loss = model(input_ids.to(torch_device), labels=labels.to(torch_device)).loss
+        mtf_score = -(labels.shape[-1] * loss.item())
+
+        EXPECTED_SCORE = -60.7397
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
+
     @slow
     def test_summarization(self):
         model = self.model

tests/test_modeling_tf_t5.py

@@ -26,7 +26,7 @@ from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
 if is_tf_available():
     import tensorflow as tf
 
-    from transformers import T5Tokenizer, TFT5EncoderModel, TFT5ForConditionalGeneration, TFT5Model
+    from transformers import ByT5Tokenizer, T5Tokenizer, TFT5EncoderModel, TFT5ForConditionalGeneration, TFT5Model
 
 
 class TFT5ModelTester:
@@ -499,6 +499,30 @@ class TFT5ModelIntegrationTests(unittest.TestCase):
         EXPECTED_SCORE = -59.0293
         self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
 
+    @slow
+    def test_small_byt5_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.9.1
+
+        >>> path_to_byt5_small_checkpoint = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_tf_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = t5.data.ByteVocabulary()
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = TFT5ForConditionalGeneration.from_pretrained("google/byt5-small")
+        tokenizer = ByT5Tokenizer.from_pretrained("google/byt5-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="tf").input_ids
+        labels = tokenizer("Hi I am", return_tensors="tf").input_ids
+
+        loss = model(input_ids, labels=labels).loss
+        mtf_score = -tf.math.reduce_sum(loss).numpy()
+
+        EXPECTED_SCORE = -60.7397
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
+
     @slow
     def test_summarization(self):
         model = self.model

tests/test_tokenization_byt5.py

+# coding=utf-8
+# Copyright 2020 Google T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import shutil
+import tempfile
+import unittest
+
+from transformers import BatchEncoding, ByT5Tokenizer
+from transformers.file_utils import cached_property, is_tf_available, is_torch_available
+
+from .test_tokenization_common import TokenizerTesterMixin
+
+
+if is_torch_available():
+    FRAMEWORK = "pt"
+elif is_tf_available():
+    FRAMEWORK = "tf"
+else:
+    FRAMEWORK = "jax"
+
+
+class ByT5TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = ByT5Tokenizer
+    test_rust_tokenizer = False
+
+    def setUp(self):
+        super().setUp()
+        tokenizer = ByT5Tokenizer()
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    @cached_property
+    def t5_base_tokenizer(self):
+        return ByT5Tokenizer.from_pretrained("google/byt5-small")
+
+    def get_tokenizer(self, **kwargs) -> ByT5Tokenizer:
+        return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
+
+    def test_eos_treatment(self):
+        tokenizer = self.t5_base_tokenizer
+        batch_with_eos_added = tokenizer(["hi</s>", "I went to the gym</s>", "</s>"])
+        batch_without_eos_added = tokenizer(["hi", "I went to the gym", ""])
+        self.assertListEqual(batch_with_eos_added["input_ids"], batch_without_eos_added["input_ids"])
+
+    def test_prepare_batch_integration(self):
+        tokenizer = self.t5_base_tokenizer
+        src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
+        # fmt: off
+        expected_src_tokens = [68, 35, 111, 114, 113, 106, 35, 115, 100, 117, 100, 106, 117, 100, 115, 107, 35, 105, 114, 117, 35, 118, 120, 112, 112, 100, 117, 108, 125, 100, 119, 108, 114, 113, 49, 1, 0]
+        # fmt: on
+        batch = tokenizer(src_text, padding=True, return_tensors=FRAMEWORK)
+        self.assertIsInstance(batch, BatchEncoding)
+
+        if FRAMEWORK != "jax":
+            result = list(batch.input_ids.numpy()[0])
+        else:
+            result = list(batch.input_ids.tolist()[0])
+
+        self.assertListEqual(expected_src_tokens, result)
+
+        self.assertEqual((2, 37), batch.input_ids.shape)
+        self.assertEqual((2, 37), batch.attention_mask.shape)
+
+    def test_empty_target_text(self):
+        tokenizer = self.t5_base_tokenizer
+        src_text = ["A long paragraph for summarization.", "Another paragraph for summarization."]
+        batch = tokenizer(src_text, padding=True, return_tensors=FRAMEWORK)
+        # check if input_ids are returned and no decoder_input_ids
+        self.assertIn("input_ids", batch)
+        self.assertIn("attention_mask", batch)
+        self.assertNotIn("decoder_input_ids", batch)
+        self.assertNotIn("decoder_attention_mask", batch)
+
+    def test_max_length_integration(self):
+        tokenizer = self.t5_base_tokenizer
+        tgt_text = [
+            "Summary of the text.",
+            "Another summary.",
+        ]
+        with tokenizer.as_target_tokenizer():
+            targets = tokenizer(
+                tgt_text, max_length=32, padding="max_length", truncation=True, return_tensors=FRAMEWORK
+            )
+        self.assertEqual(32, targets["input_ids"].shape[1])
+
+    def test_eos_in_input(self):
+        tokenizer = self.t5_base_tokenizer
+        src_text = ["A long paragraph for summarization. </s>"]
+        tgt_text = ["Summary of the text. </s>"]
+        # fmt: off
+        expected_src_tokens = [68, 35, 111, 114, 113, 106, 35, 115, 100, 117, 100, 106, 117, 100, 115, 107, 35, 105, 114, 117, 35, 118, 120, 112, 112, 100, 117, 108, 125, 100, 119, 108, 114, 113, 49, 35, 1]
+        expected_tgt_tokens = [86, 120, 112, 112, 100, 117, 124, 35, 114, 105, 35, 119, 107, 104, 35, 119, 104, 123, 119, 49, 35, 1]
+        # fmt: on
+
+        batch = tokenizer(src_text)
+        with tokenizer.as_target_tokenizer():
+            targets = tokenizer(tgt_text)
+
+        self.assertEqual(expected_src_tokens, batch["input_ids"][0])
+        self.assertEqual(expected_tgt_tokens, targets["input_ids"][0])
+
+    # cannot use default save_and_load_tokenzier test method because tokenzier has no vocab
+    def test_save_and_load_tokenizer(self):
+        # safety check on max_len default value so we are sure the test works
+        tokenizers = self.get_tokenizers()
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                self.assertNotEqual(tokenizer.model_max_length, 42)
+
+        # Now let's start the test
+        tokenizers = self.get_tokenizers()
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                # Isolate this from the other tests because we save additional tokens/etc
+                tmpdirname = tempfile.mkdtemp()
+
+                sample_text = " He is very happy, UNwant\u00E9d,running"
+                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
+                tokenizer.save_pretrained(tmpdirname)
+
+                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
+                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
+                self.assertListEqual(before_tokens, after_tokens)
+
+                shutil.rmtree(tmpdirname)
+
+        tokenizers = self.get_tokenizers(model_max_length=42)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                # Isolate this from the other tests because we save additional tokens/etc
+                tmpdirname = tempfile.mkdtemp()
+
+                sample_text = " He is very happy, UNwant\u00E9d,running"
+                tokenizer.add_tokens(["bim", "bambam"])
+                additional_special_tokens = tokenizer.additional_special_tokens
+                additional_special_tokens.append("new_additional_special_token")
+                tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
+                before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
+                tokenizer.save_pretrained(tmpdirname)
+
+                after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
+                after_tokens = after_tokenizer.encode(sample_text, add_special_tokens=False)
+                self.assertListEqual(before_tokens, after_tokens)
+                self.assertIn("new_additional_special_token", after_tokenizer.additional_special_tokens)
+                self.assertEqual(after_tokenizer.model_max_length, 42)
+
+                tokenizer = tokenizer.__class__.from_pretrained(tmpdirname, model_max_length=43)
+                self.assertEqual(tokenizer.model_max_length, 43)
+
+                shutil.rmtree(tmpdirname)
+
+    # tokenizer can be instantiated without any pretrained files, so no need for pretrained tokenizer list
+    def test_pretrained_model_lists(self):
+        pass
+
+    # tokenizer does not have vocabulary
+    def test_get_vocab(self):
+        pass
+
+    # inputs cannot be pretokenized since ids depend on whole input string and not just on single characters
+    def test_pretokenized_inputs(self):
+        pass
+
+    # tests all ids in vocab => vocab doesn't exist so unnecessary to test
+    def test_conversion_reversible(self):
+        pass