generation_tf_utils.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  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.

import numpy as np
import tensorflow as tf

from .utils import logging


logger = logging.get_logger(__name__)


class TFGenerationMixin:
    """
    A class contraining all of the functions supporting generation, to be used as a mixin in
    :class:`~transfomers.TFPreTrainedModel`.
    """

    def prepare_inputs_for_generation(self, inputs, **kwargs):
        """
        Implement in subclasses of :class:`~transfomers.TFPreTrainedModel` for custom behavior to prepare inputs in the
        generate method.
        """
        return {"inputs": inputs}

    def _use_cache(self, outputs, use_cache):
        """During generation, decide whether to pass the `past` variable to the next forward pass."""
        if len(outputs) <= 1 or use_cache is False:
            return False
        if hasattr(self.config, "mem_len") and self.config.mem_len == 0:
            return False
        return True

    def generate(
        self,
        input_ids=None,
        max_length=None,
        min_length=None,
        do_sample=None,
        early_stopping=None,
        num_beams=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,
        use_cache=None,
    ):
        r"""
        Generates sequences for models with a language modeling head. The method currently supports greedy decoding,
        beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.

        Adapted in part from `Facebook's XLM beam search code
        <https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529>`__.

        Apart from :obj:`input_ids` and :obj:`attention_mask`, all the arguments below will default to the value of the
        attribute of the same name inside the :class:`~transformers.PretrainedConfig` of the model. The default values
        indicated are the default values of those config.

        Most of these parameters are explained in more detail in `this blog post
        <https://huggingface.co/blog/how-to-generate>`__.

        Parameters:

            input_ids (:obj:`tf.Tensor` of :obj:`dtype=tf.int32` and shape :obj:`(batch_size, sequence_length)`, `optional`):
                The sequence used as a prompt for the generation. If :obj:`None` the method initializes
                it as an empty :obj:`tf.Tensor` of shape :obj:`(1,)`.
            max_length (:obj:`int`, `optional`, defaults to 20):
                The maximum length of the sequence to be generated.
            min_length (:obj:`int`, `optional`, defaults to 10):
                The minimum length of the sequence to be generated.
            do_sample (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to use sampling ; use greedy decoding otherwise.
            early_stopping (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether to stop the beam search when at least ``num_beams`` sentences are finished per batch or not.
            num_beams (:obj:`int`, `optional`, defaults to 1):
                Number of beams for beam search. 1 means no beam search.
            temperature (:obj:`float`, `optional`, defaults to 1.0):
                The value used to module the next token probabilities.
            top_k (:obj:`int`, `optional`, defaults to 50):
                The number of highest probability vocabulary tokens to keep for top-k-filtering.
            top_p (:obj:`float`, `optional`, defaults to 1.0):
                If set to float < 1, only the most probable tokens with probabilities that add up to ``top_p`` or
                higher are kept for generation.
            repetition_penalty (:obj:`float`, `optional`, defaults to 1.0):
                The parameter for repetition penalty. 1.0 means no penalty. See `this paper
                <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details.
            pad_token_id (:obj:`int`, `optional`):
                The id of the `padding` token.
            bos_token_id (:obj:`int`, `optional`):
                The id of the `beginning-of-sequence` token.
            eos_token_id (:obj:`int`, `optional`):
                The id of the `end-of-sequence` token.
            length_penalty (:obj:`float`, `optional`, defaults to 1.0):
                Exponential penalty to the length. 1.0 means no penalty.

                Set to values < 1.0 in order to encourage the model to generate shorter sequences, to a value > 1.0 in
                order to encourage the model to produce longer sequences.
            no_repeat_ngram_size (:obj:`int`, `optional`, defaults to 0):
                If set to int > 0, all ngrams of that size can only occur once.
            bad_words_ids(:obj:`List[int]`, `optional`):
                List of token ids 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 :obj:`tokenizer.encode(bad_word, add_prefix_space=True)`.
            num_return_sequences(:obj:`int`, `optional`, defaults to 1):
                The number of independently computed returned sequences for each element in the batch.
            attention_mask (:obj:`tf.Tensor` of :obj:`dtype=tf.int32` and shape :obj:`(batch_size, sequence_length)`, `optional`):
                Mask to avoid performing attention on padding token indices. Mask values are in ``[0, 1]``, 1 for
                tokens that are not masked, and 0 for masked tokens.

                If not provided, will default to a tensor the same shape as :obj:`input_ids` that masks the pad token.

                `What are attention masks? <../glossary.html#attention-mask>`__
            decoder_start_token_id (:obj:`int`, `optional`):
                If an encoder-decoder model starts decoding with a different token than `bos`, the id of that token.
            use_cache: (:obj:`bool`, `optional`, defaults to :obj:`True`):
                Whether or not the model should use the past last key/values attentions (if applicable to the model) to
                speed up decoding.
            model_specific_kwargs:
                Additional model specific kwargs will be forwarded to the :obj:`forward` function of the model.

        Return:

            :obj:`tf.Tensor` of :obj:`dtype=tf.int32` and shape :obj:`(batch_size * num_return_sequences, sequence_length)`:
            The generated sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or
            shorter if all batches finished early due to the :obj:`eos_token_id`.

        Examples::

            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
            model = TFAutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
            outputs = model.generate(max_length=40)  # do greedy decoding
            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('openai-gpt')   # Initialize tokenizer
            model = TFAutoModelWithLMHead.from_pretrained('openai-gpt')    # Download model and configuration from S3 and cache.
            input_context = 'The dog'
            input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
            outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5)  # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
            for i in range(3): #  3 output sequences were generated
                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
            model = TFAutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
            input_context = 'The dog'
            input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
            outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3, do_sample=True)  # generate 3 candidates using sampling
            for i in range(3): #  3 output sequences were generated
                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('ctrl')   # Initialize tokenizer
            model = TFAutoModelWithLMHead.from_pretrained('ctrl')    # Download model and configuration from S3 and cache.
            input_context = 'Legal My neighbor is'  # "Legal" is one of the control codes for ctrl
            input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
            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'
            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
        if self.get_output_embeddings() is None:
            raise AttributeError(
                "You tried to generate sequences with a model that does not have a LM Head."
                "Please use another model class (e.g. `TFOpenAIGPTLMHeadModel`, `TFXLNetLMHeadModel`, `TFGPT2LMHeadModel`, `TFCTRLLMHeadModel`, `TFT5ForConditionalGeneration`, `TFTransfoXLLMHeadModel`)"
            )

        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
        early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        num_beams = num_beams if num_beams is not None else self.config.num_beams
        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
        )
        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
        )
        decoder_start_token_id = (
            decoder_start_token_id if decoder_start_token_id is not None else self.config.decoder_start_token_id
        )

        if input_ids is not None:
            batch_size = shape_list(input_ids)[0]  # overriden by the input batch_size
        else:
            batch_size = 1

        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictely positive integer."
        assert isinstance(min_length, int) and min_length >= 0, "`min_length` should be a positive integer."
        assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
        assert isinstance(early_stopping, bool), "`early_stopping` should be a boolean."
        assert isinstance(use_cache, bool), "`use_cache` should be a boolean."
        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
        assert temperature > 0, "`temperature` should be strictely positive."
        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."
        assert (eos_token_id is None) or (
            isinstance(eos_token_id, int) and (eos_token_id >= 0)
        ), "`eos_token_id` should be a positive integer."
        assert length_penalty > 0, "`length_penalty` should be strictely positive."
        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, (
                "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 = tf.fill((batch_size, 1), bos_token_id)
        else:
            assert len(shape_list(input_ids)) == 2, "Input prompt should be of shape (batch_size, sequence length)."

        # not allow to duplicate outputs when greedy decoding
        if do_sample is False:
            if num_beams == 1:
                # no_beam_search greedy generation conditions
                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 (
                    num_beams >= num_return_sequences
                ), "Greedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequences"

        # 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.numpy()):
            attention_mask = tf.cast(tf.math.not_equal(input_ids, pad_token_id), dtype=tf.int32)
        elif attention_mask is None:
            attention_mask = tf.ones_like(input_ids)

        if pad_token_id is None and eos_token_id is not None:
            logger.warning(
                "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_id)
            )
            pad_token_id = eos_token_id

        # current position and vocab size
        cur_len = shape_list(input_ids)[1]  # unused
        vocab_size = self.config.vocab_size

        # set effective batch size and effective batch multiplier according to do_sample
        if do_sample:
            effective_batch_size = batch_size * num_return_sequences
            effective_batch_mult = num_return_sequences
        else:
            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
        if num_return_sequences > 1 or num_beams > 1:
            input_ids_len = shape_list(input_ids)[-1]
            input_ids = tf.broadcast_to(
                tf.expand_dims(input_ids, 1), (batch_size, effective_batch_mult * num_beams, input_ids_len)
            )
            attention_mask = tf.broadcast_to(
                tf.expand_dims(attention_mask, 1), (batch_size, effective_batch_mult * num_beams, input_ids_len)
            )
            input_ids = tf.reshape(
                input_ids, (effective_batch_size * num_beams, input_ids_len)
            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
            attention_mask = tf.reshape(
                attention_mask, (effective_batch_size * num_beams, input_ids_len)
            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)

        if self.config.is_encoder_decoder:

            # create empty decoder_input_ids
            input_ids = (
                tf.ones(
                    (effective_batch_size * num_beams, 1),
                    dtype=tf.int32,
                )
                * decoder_start_token_id
            )
            cur_len = 1

            assert (
                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 = tf.reshape(
                tf.repeat(tf.expand_dims(tf.range(batch_size), -1), repeats=num_beams * effective_batch_mult, axis=1),
                shape=(-1,),
            )
            # expand encoder_outputs
            encoder_outputs = (tf.gather(encoder_outputs[0], expanded_batch_idxs, axis=0), *encoder_outputs[1:])

        else:
            encoder_outputs = None
            cur_len = shape_list(input_ids)[-1]

        assert (
            cur_len < max_length
        ), f"The context has {cur_len} number of tokens, but `max_length` is only {max_length}. Please make sure that `max_length` is bigger than the number of tokens, by setting either `generate(max_length=...,...)` or `config.max_length = ...`"

        if num_beams > 1:
            output = self._generate_beam_search(
                input_ids,
                cur_len=cur_len,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
                early_stopping=early_stopping,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
                bad_words_ids=bad_words_ids,
                pad_token_id=pad_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,
                cur_len=cur_len,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
                bad_words_ids=bad_words_ids,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                batch_size=effective_batch_size,
                vocab_size=vocab_size,
                encoder_outputs=encoder_outputs,
                attention_mask=attention_mask,
                use_cache=use_cache,
            )

        return output

    def _generate_no_beam_search(
        self,
        input_ids,
        cur_len,
        max_length,
        min_length,
        do_sample,
        temperature,
        top_k,
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
        bad_words_ids,
        pad_token_id,
        eos_token_id,
        batch_size,
        vocab_size,
        encoder_outputs,
        attention_mask,
        use_cache,
    ):
        """Generate sequences for each example without beam search (num_beams == 1).
        All returned sequence are generated independantly.
        """

        # length of generated sentences / unfinished sentences
        unfinished_sents = tf.ones_like(input_ids[:, 0])
        sent_lengths = tf.ones_like(input_ids[:, 0]) * max_length

        past = encoder_outputs  # 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, use_cache=use_cache
            )
            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._use_cache(outputs, use_cache):
                past = outputs[1]

            # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
                next_token_logits_penalties = _create_next_token_logits_penalties(
                    input_ids, next_token_logits, repetition_penalty
                )
                next_token_logits = tf.math.multiply(next_token_logits, next_token_logits_penalties)

            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_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:
                    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")
                )

            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
                is_token_logit_eos_token = tf.convert_to_tensor(
                    [True if token is eos_token_id else False for token in range(vocab_size)], dtype=tf.bool
                )
                eos_token_indices_mask = tf.broadcast_to(is_token_logit_eos_token, [batch_size, vocab_size])

                next_token_logits = set_tensor_by_indices_to_value(
                    next_token_logits, eos_token_indices_mask, -float("inf")
                )

            if do_sample:
                # Temperature (higher temperature => more likely to sample low probability tokens)
                if temperature != 1.0:
                    next_token_logits = next_token_logits / temperature
                # Top-p/top-k filtering
                next_token_logits = tf_top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
                # Sample
                next_token = tf.squeeze(
                    tf.random.categorical(next_token_logits, dtype=tf.int32, num_samples=1), axis=1
                )
            else:
                # Greedy decoding
                next_token = tf.math.argmax(next_token_logits, axis=-1, output_type=tf.int32)

            # update generations and finished sentences
            if eos_token_id is not None:
                # pad finished sentences if eos_token_id exist
                tokens_to_add = next_token * unfinished_sents + (pad_token_id) * (1 - unfinished_sents)
            else:
                tokens_to_add = next_token

            # add token and increase length by one
            input_ids = tf.concat([input_ids, tf.expand_dims(tokens_to_add, -1)], 1)
            cur_len = cur_len + 1

            if eos_token_id is not None:
                eos_in_sents = tokens_to_add == eos_token_id
                # if sentence is unfinished and the token to add is eos, sent_lengths is filled with current length
                is_sents_unfinished_and_token_to_add_is_eos = tf.math.multiply(
                    unfinished_sents, tf.cast(eos_in_sents, tf.int32)
                )
                sent_lengths = (
                    sent_lengths * (1 - is_sents_unfinished_and_token_to_add_is_eos)
                    + cur_len * is_sents_unfinished_and_token_to_add_is_eos
                )

                # unfinished_sents is set to zero if eos in sentence
                unfinished_sents -= is_sents_unfinished_and_token_to_add_is_eos

            # stop when there is a </s> in each sentence, or if we exceed the maximul length
            if tf.math.reduce_max(unfinished_sents) == 0:
                break

            # 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
                )

        # if there are different sentences lengths in the batch, some batches have to be padded
        min_sent_length = tf.math.reduce_min(sent_lengths)
        max_sent_length = tf.math.reduce_max(sent_lengths)
        if min_sent_length != max_sent_length:
            assert pad_token_id is not None, "`Pad_token_id` has to be defined if batches have different lengths"
            # finished sents are filled with pad_token
            padding = tf.ones([batch_size, max_sent_length.numpy()], dtype=tf.int32) * pad_token_id

            # create length masks for tf.where operation
            broad_casted_sent_lengths = tf.broadcast_to(
                tf.expand_dims(sent_lengths, -1), [batch_size, max_sent_length]
            )
            broad_casted_range = tf.transpose(
                tf.broadcast_to(tf.expand_dims(tf.range(max_sent_length), -1), [max_sent_length, batch_size])
            )

            decoded = tf.where(broad_casted_range < broad_casted_sent_lengths, input_ids, padding)
        else:
            decoded = input_ids

        return decoded

    def _generate_beam_search(
        self,
        input_ids,
        cur_len,
        max_length,
        min_length,
        do_sample,
        early_stopping,
        temperature,
        top_k,
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
        bad_words_ids,
        pad_token_id,
        eos_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."""

        # generated hypotheses
        generated_hyps = [
            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=early_stopping)
            for _ in range(batch_size)
        ]

        # for greedy decoding it is made sure that only tokens of the first beam are considered to avoid sampling the exact same tokens three times
        if do_sample is False:
            beam_scores_begin = tf.zeros((batch_size, 1), dtype=tf.float32)
            beam_scores_end = tf.ones((batch_size, num_beams - 1), dtype=tf.float32) * (-1e9)
            beam_scores = tf.concat([beam_scores_begin, beam_scores_end], -1)
        else:
            beam_scores = tf.zeros((batch_size, num_beams), dtype=tf.float32)

        beam_scores = tf.reshape(beam_scores, (batch_size * num_beams,))

        # cache compute states
        past = encoder_outputs
        # to stay similar to torch : past = (encoder_outputs, None) if encoder_outputs is not None else None

        # 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, use_cache=use_cache
            )
            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._use_cache(outputs, use_cache):
                past = outputs[1]

            # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
                next_token_logits_penalties = _create_next_token_logits_penalties(
                    input_ids, next_token_logits, repetition_penalty
                )
                next_token_logits = tf.math.multiply(next_token_logits, next_token_logits_penalties)

            # Temperature (higher temperature => more likely to sample low probability tokens)
            if temperature != 1.0:
                next_token_logits = next_token_logits / temperature

            #             calculate log softmax score
            scores = tf.nn.log_softmax(next_token_logits, axis=-1)  # (batch_size * num_beams, vocab_size)

            # 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
                num_batch_hypotheses = batch_size * num_beams

                is_token_logit_eos_token = tf.convert_to_tensor(
                    [True if token is eos_token_id else False for token in range(vocab_size)], dtype=tf.bool
                )
                eos_token_indices_mask = tf.broadcast_to(is_token_logit_eos_token, [num_batch_hypotheses, vocab_size])

                scores = set_tensor_by_indices_to_value(scores, eos_token_indices_mask, -float("inf"))

            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
                num_batch_hypotheses = batch_size * num_beams
                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:
                    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")
                )

            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:
                _scores = scores + tf.broadcast_to(
                    beam_scores[:, None], (batch_size * num_beams, vocab_size)
                )  # (batch_size * num_beams, vocab_size)

                # Top-p/top-k filtering
                _scores = tf_top_k_top_p_filtering(
                    _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
                )  # (batch_size * num_beams, vocab_size)
                # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search)
                _scores = tf.reshape(_scores, (batch_size, num_beams * vocab_size))

                next_tokens = sample_without_replacement(
                    _scores, num_samples=2 * num_beams
                )  # (batch_size, 2 * num_beams)
                # Compute next scores
                next_scores = tf.gather(_scores, next_tokens, batch_dims=1)  # (batch_size, 2 * num_beams)

                # sort the sampled vector to make sure that the first num_beams samples are the best
                next_scores_indices = tf.argsort(next_scores, direction="DESCENDING", axis=1)
                next_scores = tf.gather(next_scores, next_scores_indices, batch_dims=1)  # (batch_size, num_beams * 2)
                next_tokens = tf.gather(next_tokens, next_scores_indices, batch_dims=1)  # (batch_size, num_beams * 2)
            else:
                # Add the log prob of the new beams to the log prob of the beginning of the sequence (sum of logs == log of the product)
                next_scores = scores + tf.broadcast_to(
                    beam_scores[:, None], (batch_size * num_beams, vocab_size)
                )  # (batch_size * num_beams, vocab_size)

                # re-organize to group the beam together (we are keeping top hypothesis accross beams)
                next_scores = tf.reshape(
                    next_scores, (batch_size, num_beams * vocab_size)
                )  # (batch_size, num_beams * vocab_size)

                next_scores, next_tokens = tf.math.top_k(next_scores, k=2 * num_beams, sorted=True)

            assert shape_list(next_scores) == shape_list(next_tokens) == [batch_size, 2 * num_beams]

            # next batch beam content
            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
                    ), "Batch can only be done if at least {} beams have been generated".format(num_beams)
                    assert (
                        eos_token_id is not None and pad_token_id is not None
                    ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
                    continue

                # next sentence beam content
                next_sent_beam = []

                # next tokens for this sentence
                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() == 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:
                            continue
                        generated_hyps[batch_idx].add(
                            tf.identity(input_ids[effective_beam_id]), beam_token_score.numpy()
                        )
                    else:
                        # 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
                    if len(next_sent_beam) == num_beams:
                        break

                # Check if we are 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(), cur_len
                )

                # update next beam content
                assert len(next_sent_beam) == num_beams, "Beam should always be full"
                next_batch_beam.extend(next_sent_beam)
                assert len(next_batch_beam) == num_beams * (batch_idx + 1)

            # stop when we are done with each sentence
            if all(done):
                break

            # sanity check / prepare next batch
            assert len(next_batch_beam) == batch_size * num_beams
            beam_scores = tf.convert_to_tensor([x[0] for x in next_batch_beam], dtype=tf.float32)
            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)

            # re-order batch and update current length
            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)
            cur_len = cur_len + 1

            # re-order internal states
            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
                )

        # finalize all open beam hypotheses and end to generated hypotheses
        for batch_idx in range(batch_size):
            # Add all open beam hypothesis to generated_hyps
            if done[batch_idx]:
                continue
            # test that beam scores match previously calculated scores if not eos and batch_idx not done
            if eos_token_id is not None and all(
                (token_id % vocab_size).numpy().item() != eos_token_id for token_id in next_tokens[batch_idx]
            ):
                assert tf.reduce_all(
                    next_scores[batch_idx, :num_beams] == tf.reshape(beam_scores, (batch_size, num_beams))[batch_idx]
                ), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format(
                    next_scores[:, :num_beams][batch_idx], tf.reshape(beam_scores, (batch_size, num_beams))[batch_idx]
                )

            # need to add best num_beams hypotheses to generated hyps
            for beam_id in range(num_beams):
                effective_beam_id = batch_idx * num_beams + beam_id
                final_score = beam_scores[effective_beam_id].numpy().item()
                final_tokens = input_ids[effective_beam_id]
                generated_hyps[batch_idx].add(final_tokens, final_score)

        # depending on whether greedy generation is wanted or not define different output_batch_size and output_num_return_sequences_per_batch
        output_batch_size = batch_size if do_sample else batch_size * num_return_sequences
        output_num_return_sequences_per_batch = 1 if do_sample else num_return_sequences

        # select the best hypotheses
        sent_lengths_list = []
        best = []

        # retrieve best hypotheses
        for i, hypotheses in enumerate(generated_hyps):
            sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0])
            for j in range(output_num_return_sequences_per_batch):
                best_hyp = sorted_hyps.pop()[1]
                sent_lengths_list.append(len(best_hyp))
                best.append(best_hyp)
        assert output_batch_size == len(best), "Output batch size {} must match output beam hypotheses {}".format(
            output_batch_size, len(best)
        )

        sent_lengths = tf.convert_to_tensor(sent_lengths_list, dtype=tf.int32)

        # shorter batches are filled with pad_token
        if tf.reduce_min(sent_lengths).numpy() != tf.reduce_max(sent_lengths).numpy():
            assert pad_token_id is not None, "`Pad_token_id` has to be defined"
            sent_max_len = min(tf.reduce_max(sent_lengths).numpy() + 1, max_length)
            decoded_list = []

            # fill with hypothesis and eos_token_id if necessary
            for i, hypo in enumerate(best):
                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)

            decoded = tf.stack(decoded_list)
        else:
            # none of the hypotheses have an eos_token
            assert (len(hypo) == max_length for hypo in best)
            decoded = tf.stack(best)

        return decoded

    @staticmethod
    def _reorder_cache(past, beam_idx):
        return tuple(tf.gather(layer_past, beam_idx, axis=1) for layer_past in past)


def _create_next_token_logits_penalties(input_ids, logits, repetition_penalty):
    # create logit penalties for already seen input_ids
    token_penalties = np.ones(shape_list(logits))
    prev_input_ids = [np.unique(input_id) for input_id in input_ids.numpy()]
    for i, prev_input_id in enumerate(prev_input_ids):
        logit_penalized = logits[i].numpy()[prev_input_id]
        logit_penalties = np.zeros(logit_penalized.shape)
        # if previous logit score is < 0 then multiply repetition penalty else divide
        logit_penalties[logit_penalized < 0] = repetition_penalty
        logit_penalties[logit_penalized > 0] = 1 / repetition_penalty
        np.put(token_penalties[i], prev_input_id, logit_penalties)
    return tf.convert_to_tensor(token_penalties, dtype=tf.float32)


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
        return [[] for _ in range(num_hypos)]
    generated_ngrams = [{} for _ in range(num_hypos)]
    for idx in range(num_hypos):
        gen_tokens = prev_input_ids[idx].numpy().tolist()
        generated_ngram = generated_ngrams[idx]
        for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
            prev_ngram_tuple = tuple(ngram[:-1])
            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]

    def _get_generated_ngrams(hypo_idx):
        # Before decoding the next token, prevent decoding of ngrams that have already appeared
        start_idx = cur_len + 1 - no_repeat_ngram_size
        ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].numpy().tolist())
        return generated_ngrams[hypo_idx].get(ngram_idx, [])

    banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
    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_tokens):
            # if bad word tokens are longer than prev tokens 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:
        logits: logits distribution shape (batch size, vocabulary size)
        if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
        if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
            Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        Make sure we keep at least min_tokens_to_keep per batch example in the output
    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
    logits_shape = shape_list(logits)

    if top_k > 0:
        top_k = min(max(top_k, min_tokens_to_keep), logits_shape[-1])  # Safety check
        # Remove all tokens with a probability less than the last token of the top-k
        indices_to_remove = logits < tf.math.top_k(logits, k=top_k)[0][..., -1, None]
        logits = set_tensor_by_indices_to_value(logits, indices_to_remove, filter_value)

    if top_p < 1.0:
        sorted_indices = tf.argsort(logits, direction="DESCENDING")
        sorted_logits = tf.gather(
            logits, sorted_indices, axis=-1, batch_dims=1
        )  # expects logits to be of dim (batch_size, vocab_size)

        cumulative_probs = tf.math.cumsum(tf.nn.softmax(sorted_logits, axis=-1), axis=-1)

        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
        sorted_indices_to_remove = cumulative_probs > top_p

        if min_tokens_to_keep > 1:
            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
            sorted_indices_to_remove = tf.concat(
                [
                    tf.zeros_like(sorted_indices_to_remove[:, :min_tokens_to_keep]),
                    sorted_indices_to_remove[:, min_tokens_to_keep:],
                ],
                -1,
            )

        # Shift the indices to the right to keep also the first token above the threshold
        sorted_indices_to_remove = tf.roll(sorted_indices_to_remove, 1, axis=-1)
        sorted_indices_to_remove = tf.concat(
            [tf.zeros_like(sorted_indices_to_remove[:, :1]), sorted_indices_to_remove[:, 1:]],
            -1,
        )
        # scatter sorted tensors to original indexing
        indices_to_remove = scatter_values_on_batch_indices(sorted_indices_to_remove, sorted_indices)
        logits = set_tensor_by_indices_to_value(logits, indices_to_remove, filter_value)
    return logits


def scatter_values_on_batch_indices(values, batch_indices):
    shape = shape_list(batch_indices)
    # broadcast batch dim to shape
    broad_casted_batch_dims = tf.reshape(tf.broadcast_to(tf.expand_dims(tf.range(shape[0]), axis=-1), shape), [1, -1])
    # transform batch_indices to pair_indices
    pair_indices = tf.transpose(tf.concat([broad_casted_batch_dims, tf.reshape(batch_indices, [1, -1])], 0))
    # scatter values to pair indices
    return tf.scatter_nd(pair_indices, tf.reshape(values, [-1]), shape)


def set_tensor_by_indices_to_value(tensor, indices, value):
    # create value_tensor since tensor value assignment is not possible in TF
    value_tensor = tf.zeros_like(tensor) + value
    return tf.where(indices, value_tensor, tensor)


def sample_without_replacement(logits, num_samples):
    """
    categorical sampling witouth replacement is currently not implemented
    the gumbel-max trick will do for now
    see https://github.com/tensorflow/tensorflow/issues/9260 for more info
    """
    z = -tf.math.log(tf.random.uniform(shape_list(logits), 0, 1))
    _, indices = tf.nn.top_k(logits + z, num_samples)
    return indices


def shape_list(x):
    """Deal with dynamic shape in tensorflow cleanly."""
    static = x.shape.as_list()
    dynamic = tf.shape(x)
    return [dynamic[i] if s is None else s for i, s in enumerate(static)]


class BeamHypotheses(object):
    def __init__(self, num_beams, max_length, length_penalty, early_stopping):
        """
        Initialize n-best list of hypotheses.
        """
        self.max_length = max_length - 1  # ignoring bos_token
        self.length_penalty = length_penalty
        self.early_stopping = early_stopping
        self.num_beams = num_beams
        self.beams = []
        self.worst_score = 1e9

    def __len__(self):
        """
        Number of hypotheses in the list.
        """
        return len(self.beams)

    def add(self, hyp, sum_logprobs):
        """
        Add a new hypothesis to the list.
        """
        score = sum_logprobs / len(hyp) ** self.length_penalty
        if len(self) < self.num_beams or score > self.worst_score:
            self.beams.append((score, hyp))
            if len(self) > self.num_beams:
                sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
                del self.beams[sorted_scores[0][1]]
                self.worst_score = sorted_scores[1][0]
            else:
                self.worst_score = min(score, self.worst_score)

    def is_done(self, best_sum_logprobs, cur_len):
        """
        If there are enough hypotheses and that none of the hypotheses being generated
        can become better than the worst one in the heap, then we are done with this sentence.
        """

        if len(self) < self.num_beams:
            return False
        elif self.early_stopping:
            return True
        else:
            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
            ret = self.worst_score >= cur_score
            return ret