transformers/examples/research_projects/bertabs/modeling_bertabs.py

# MIT License

# Copyright (c) 2019 Yang Liu and the HuggingFace team

# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:

# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.

# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import copy
import math

import numpy as np
import torch
from configuration_bertabs import BertAbsConfig
from torch import nn
from torch.nn.init import xavier_uniform_

from transformers import BertConfig, BertModel, PreTrainedModel


MAX_SIZE = 5000

BERTABS_FINETUNED_MODEL_ARCHIVE_LIST = [
    "remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization",
]


class BertAbsPreTrainedModel(PreTrainedModel):
    config_class = BertAbsConfig
    load_tf_weights = False
    base_model_prefix = "bert"


class BertAbs(BertAbsPreTrainedModel):
    def __init__(self, args, checkpoint=None, bert_extractive_checkpoint=None):
        super().__init__(args)
        self.args = args
        self.bert = Bert()

        # If pre-trained weights are passed for Bert, load these.
        load_bert_pretrained_extractive = True if bert_extractive_checkpoint else False
        if load_bert_pretrained_extractive:
            self.bert.model.load_state_dict(
                {n[11:]: p for n, p in bert_extractive_checkpoint.items() if n.startswith("bert.model")},
                strict=True,
            )

        self.vocab_size = self.bert.model.config.vocab_size

        if args.max_pos > 512:
            my_pos_embeddings = nn.Embedding(args.max_pos, self.bert.model.config.hidden_size)
            my_pos_embeddings.weight.data[:512] = self.bert.model.embeddings.position_embeddings.weight.data
            my_pos_embeddings.weight.data[512:] = self.bert.model.embeddings.position_embeddings.weight.data[-1][
                None, :
            ].repeat(args.max_pos - 512, 1)
            self.bert.model.embeddings.position_embeddings = my_pos_embeddings
        tgt_embeddings = nn.Embedding(self.vocab_size, self.bert.model.config.hidden_size, padding_idx=0)

        tgt_embeddings.weight = copy.deepcopy(self.bert.model.embeddings.word_embeddings.weight)

        self.decoder = TransformerDecoder(
            self.args.dec_layers,
            self.args.dec_hidden_size,
            heads=self.args.dec_heads,
            d_ff=self.args.dec_ff_size,
            dropout=self.args.dec_dropout,
            embeddings=tgt_embeddings,
            vocab_size=self.vocab_size,
        )

        gen_func = nn.LogSoftmax(dim=-1)
        self.generator = nn.Sequential(nn.Linear(args.dec_hidden_size, args.vocab_size), gen_func)
        self.generator[0].weight = self.decoder.embeddings.weight

        load_from_checkpoints = False if checkpoint is None else True
        if load_from_checkpoints:
            self.load_state_dict(checkpoint)

    def init_weights(self):
        for module in self.decoder.modules():
            if isinstance(module, (nn.Linear, nn.Embedding)):
                module.weight.data.normal_(mean=0.0, std=0.02)
            elif isinstance(module, nn.LayerNorm):
                module.bias.data.zero_()
                module.weight.data.fill_(1.0)
            if isinstance(module, nn.Linear) and module.bias is not None:
                module.bias.data.zero_()
        for p in self.generator.parameters():
            if p.dim() > 1:
                xavier_uniform_(p)
            else:
                p.data.zero_()

    def forward(
        self,
        encoder_input_ids,
        decoder_input_ids,
        token_type_ids,
        encoder_attention_mask,
        decoder_attention_mask,
    ):
        encoder_output = self.bert(
            input_ids=encoder_input_ids,
            token_type_ids=token_type_ids,
            attention_mask=encoder_attention_mask,
        )
        encoder_hidden_states = encoder_output[0]
        dec_state = self.decoder.init_decoder_state(encoder_input_ids, encoder_hidden_states)
        decoder_outputs, _ = self.decoder(decoder_input_ids[:, :-1], encoder_hidden_states, dec_state)
        return decoder_outputs


class Bert(nn.Module):
    """This class is not really necessary and should probably disappear."""

    def __init__(self):
        super().__init__()
        config = BertConfig.from_pretrained("bert-base-uncased")
        self.model = BertModel(config)

    def forward(self, input_ids, attention_mask=None, token_type_ids=None, **kwargs):
        self.eval()
        with torch.no_grad():
            encoder_outputs, _ = self.model(
                input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, **kwargs
            )
        return encoder_outputs


class TransformerDecoder(nn.Module):
    """
    The Transformer decoder from "Attention is All You Need".

    Args:
       num_layers (int): number of encoder layers.
       d_model (int): size of the model
       heads (int): number of heads
       d_ff (int): size of the inner FF layer
       dropout (float): dropout parameters
       embeddings (:obj:`onmt.modules.Embeddings`):
          embeddings to use, should have positional encodings
       attn_type (str): if using a separate copy attention
    """

    def __init__(self, num_layers, d_model, heads, d_ff, dropout, embeddings, vocab_size):
        super().__init__()

        # Basic attributes.
        self.decoder_type = "transformer"
        self.num_layers = num_layers
        self.embeddings = embeddings
        self.pos_emb = PositionalEncoding(dropout, self.embeddings.embedding_dim)

        # Build TransformerDecoder.
        self.transformer_layers = nn.ModuleList(
            [TransformerDecoderLayer(d_model, heads, d_ff, dropout) for _ in range(num_layers)]
        )

        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)

    # forward(input_ids, attention_mask, encoder_hidden_states, encoder_attention_mask)
    # def forward(self, input_ids, state, attention_mask=None, memory_lengths=None,
    # step=None, cache=None, encoder_attention_mask=None, encoder_hidden_states=None, memory_masks=None):
    def forward(
        self,
        input_ids,
        encoder_hidden_states=None,
        state=None,
        attention_mask=None,
        memory_lengths=None,
        step=None,
        cache=None,
        encoder_attention_mask=None,
    ):
        """
        See :obj:`onmt.modules.RNNDecoderBase.forward()`
        memory_bank = encoder_hidden_states
        """
        # Name conversion
        tgt = input_ids
        memory_bank = encoder_hidden_states
        memory_mask = encoder_attention_mask

        # src_words = state.src
        src_words = state.src
        src_batch, src_len = src_words.size()

        padding_idx = self.embeddings.padding_idx

        # Decoder padding mask
        tgt_words = tgt
        tgt_batch, tgt_len = tgt_words.size()
        tgt_pad_mask = tgt_words.data.eq(padding_idx).unsqueeze(1).expand(tgt_batch, tgt_len, tgt_len)

        # Encoder padding mask
        if memory_mask is not None:
            src_len = memory_mask.size(-1)
            src_pad_mask = memory_mask.expand(src_batch, tgt_len, src_len)
        else:
            src_pad_mask = src_words.data.eq(padding_idx).unsqueeze(1).expand(src_batch, tgt_len, src_len)

        # Pass through the embeddings
        emb = self.embeddings(input_ids)
        output = self.pos_emb(emb, step)
        assert emb.dim() == 3  # len x batch x embedding_dim

        if state.cache is None:
            saved_inputs = []

        for i in range(self.num_layers):
            prev_layer_input = None
            if state.cache is None:
                if state.previous_input is not None:
                    prev_layer_input = state.previous_layer_inputs[i]

            output, all_input = self.transformer_layers[i](
                output,
                memory_bank,
                src_pad_mask,
                tgt_pad_mask,
                previous_input=prev_layer_input,
                layer_cache=state.cache["layer_{}".format(i)] if state.cache is not None else None,
                step=step,
            )
            if state.cache is None:
                saved_inputs.append(all_input)

        if state.cache is None:
            saved_inputs = torch.stack(saved_inputs)

        output = self.layer_norm(output)

        if state.cache is None:
            state = state.update_state(tgt, saved_inputs)

        # Decoders in transformers return a tuple. Beam search will fail
        # if we don't follow this convention.
        return output, state  # , state

    def init_decoder_state(self, src, memory_bank, with_cache=False):
        """Init decoder state"""
        state = TransformerDecoderState(src)
        if with_cache:
            state._init_cache(memory_bank, self.num_layers)
        return state


class PositionalEncoding(nn.Module):
    def __init__(self, dropout, dim, max_len=5000):
        pe = torch.zeros(max_len, dim)
        position = torch.arange(0, max_len).unsqueeze(1)
        div_term = torch.exp((torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim)))
        pe[:, 0::2] = torch.sin(position.float() * div_term)
        pe[:, 1::2] = torch.cos(position.float() * div_term)
        pe = pe.unsqueeze(0)
        super().__init__()
        self.register_buffer("pe", pe)
        self.dropout = nn.Dropout(p=dropout)
        self.dim = dim

    def forward(self, emb, step=None):
        emb = emb * math.sqrt(self.dim)
        if step:
            emb = emb + self.pe[:, step][:, None, :]

        else:
            emb = emb + self.pe[:, : emb.size(1)]
        emb = self.dropout(emb)
        return emb

    def get_emb(self, emb):
        return self.pe[:, : emb.size(1)]


class TransformerDecoderLayer(nn.Module):
    """
    Args:
      d_model (int): the dimension of keys/values/queries in
                       MultiHeadedAttention, also the input size of
                       the first-layer of the PositionwiseFeedForward.
      heads (int): the number of heads for MultiHeadedAttention.
      d_ff (int): the second-layer of the PositionwiseFeedForward.
      dropout (float): dropout probability(0-1.0).
      self_attn_type (string): type of self-attention scaled-dot, average
    """

    def __init__(self, d_model, heads, d_ff, dropout):
        super().__init__()

        self.self_attn = MultiHeadedAttention(heads, d_model, dropout=dropout)

        self.context_attn = MultiHeadedAttention(heads, d_model, dropout=dropout)
        self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
        self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
        self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
        self.drop = nn.Dropout(dropout)
        mask = self._get_attn_subsequent_mask(MAX_SIZE)
        # Register self.mask as a saved_state in TransformerDecoderLayer, so
        # it gets TransformerDecoderLayer's cuda behavior automatically.
        self.register_buffer("mask", mask)

    def forward(
        self,
        inputs,
        memory_bank,
        src_pad_mask,
        tgt_pad_mask,
        previous_input=None,
        layer_cache=None,
        step=None,
    ):
        """
        Args:
            inputs (`FloatTensor`): `[batch_size x 1 x model_dim]`
            memory_bank (`FloatTensor`): `[batch_size x src_len x model_dim]`
            src_pad_mask (`LongTensor`): `[batch_size x 1 x src_len]`
            tgt_pad_mask (`LongTensor`): `[batch_size x 1 x 1]`

        Returns:
            (`FloatTensor`, `FloatTensor`, `FloatTensor`):

            * output `[batch_size x 1 x model_dim]`
            * attn `[batch_size x 1 x src_len]`
            * all_input `[batch_size x current_step x model_dim]`

        """
        dec_mask = torch.gt(tgt_pad_mask + self.mask[:, : tgt_pad_mask.size(1), : tgt_pad_mask.size(1)], 0)
        input_norm = self.layer_norm_1(inputs)
        all_input = input_norm
        if previous_input is not None:
            all_input = torch.cat((previous_input, input_norm), dim=1)
            dec_mask = None

        query = self.self_attn(
            all_input,
            all_input,
            input_norm,
            mask=dec_mask,
            layer_cache=layer_cache,
            type="self",
        )

        query = self.drop(query) + inputs

        query_norm = self.layer_norm_2(query)
        mid = self.context_attn(
            memory_bank,
            memory_bank,
            query_norm,
            mask=src_pad_mask,
            layer_cache=layer_cache,
            type="context",
        )
        output = self.feed_forward(self.drop(mid) + query)

        return output, all_input
        # return output

    def _get_attn_subsequent_mask(self, size):
        """
        Get an attention mask to avoid using the subsequent info.

        Args:
            size: int

        Returns:
            (`LongTensor`):

            * subsequent_mask `[1 x size x size]`
        """
        attn_shape = (1, size, size)
        subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype("uint8")
        subsequent_mask = torch.from_numpy(subsequent_mask)
        return subsequent_mask


class MultiHeadedAttention(nn.Module):
    """
    Multi-Head Attention module from
    "Attention is All You Need"
    :cite:`DBLP:journals/corr/VaswaniSPUJGKP17`.

    Similar to standard `dot` attention but uses
    multiple attention distributions simulataneously
    to select relevant items.

    .. mermaid::

       graph BT
          A[key]
          B[value]
          C[query]
          O[output]
          subgraph Attn
            D[Attn 1]
            E[Attn 2]
            F[Attn N]
          end
          A --> D
          C --> D
          A --> E
          C --> E
          A --> F
          C --> F
          D --> O
          E --> O
          F --> O
          B --> O

    Also includes several additional tricks.

    Args:
       head_count (int): number of parallel heads
       model_dim (int): the dimension of keys/values/queries,
           must be divisible by head_count
       dropout (float): dropout parameter
    """

    def __init__(self, head_count, model_dim, dropout=0.1, use_final_linear=True):
        assert model_dim % head_count == 0
        self.dim_per_head = model_dim // head_count
        self.model_dim = model_dim

        super().__init__()
        self.head_count = head_count

        self.linear_keys = nn.Linear(model_dim, head_count * self.dim_per_head)
        self.linear_values = nn.Linear(model_dim, head_count * self.dim_per_head)
        self.linear_query = nn.Linear(model_dim, head_count * self.dim_per_head)
        self.softmax = nn.Softmax(dim=-1)
        self.dropout = nn.Dropout(dropout)
        self.use_final_linear = use_final_linear
        if self.use_final_linear:
            self.final_linear = nn.Linear(model_dim, model_dim)

    def forward(
        self,
        key,
        value,
        query,
        mask=None,
        layer_cache=None,
        type=None,
        predefined_graph_1=None,
    ):
        """
        Compute the context vector and the attention vectors.

        Args:
           key (`FloatTensor`): set of `key_len`
                key vectors `[batch, key_len, dim]`
           value (`FloatTensor`): set of `key_len`
                value vectors `[batch, key_len, dim]`
           query (`FloatTensor`): set of `query_len`
                 query vectors  `[batch, query_len, dim]`
           mask: binary mask indicating which keys have
                 non-zero attention `[batch, query_len, key_len]`
        Returns:
           (`FloatTensor`, `FloatTensor`) :

           * output context vectors `[batch, query_len, dim]`
           * one of the attention vectors `[batch, query_len, key_len]`
        """
        batch_size = key.size(0)
        dim_per_head = self.dim_per_head
        head_count = self.head_count

        def shape(x):
            """projection"""
            return x.view(batch_size, -1, head_count, dim_per_head).transpose(1, 2)

        def unshape(x):
            """compute context"""
            return x.transpose(1, 2).contiguous().view(batch_size, -1, head_count * dim_per_head)

        # 1) Project key, value, and query.
        if layer_cache is not None:
            if type == "self":
                query, key, value = (
                    self.linear_query(query),
                    self.linear_keys(query),
                    self.linear_values(query),
                )

                key = shape(key)
                value = shape(value)

                if layer_cache is not None:
                    device = key.device
                    if layer_cache["self_keys"] is not None:
                        key = torch.cat((layer_cache["self_keys"].to(device), key), dim=2)
                    if layer_cache["self_values"] is not None:
                        value = torch.cat((layer_cache["self_values"].to(device), value), dim=2)
                    layer_cache["self_keys"] = key
                    layer_cache["self_values"] = value
            elif type == "context":
                query = self.linear_query(query)
                if layer_cache is not None:
                    if layer_cache["memory_keys"] is None:
                        key, value = self.linear_keys(key), self.linear_values(value)
                        key = shape(key)
                        value = shape(value)
                    else:
                        key, value = (
                            layer_cache["memory_keys"],
                            layer_cache["memory_values"],
                        )
                    layer_cache["memory_keys"] = key
                    layer_cache["memory_values"] = value
                else:
                    key, value = self.linear_keys(key), self.linear_values(value)
                    key = shape(key)
                    value = shape(value)
        else:
            key = self.linear_keys(key)
            value = self.linear_values(value)
            query = self.linear_query(query)
            key = shape(key)
            value = shape(value)

        query = shape(query)

        # 2) Calculate and scale scores.
        query = query / math.sqrt(dim_per_head)
        scores = torch.matmul(query, key.transpose(2, 3))

        if mask is not None:
            mask = mask.unsqueeze(1).expand_as(scores)
            scores = scores.masked_fill(mask, -1e18)

        # 3) Apply attention dropout and compute context vectors.

        attn = self.softmax(scores)

        if predefined_graph_1 is not None:
            attn_masked = attn[:, -1] * predefined_graph_1
            attn_masked = attn_masked / (torch.sum(attn_masked, 2).unsqueeze(2) + 1e-9)

            attn = torch.cat([attn[:, :-1], attn_masked.unsqueeze(1)], 1)

        drop_attn = self.dropout(attn)
        if self.use_final_linear:
            context = unshape(torch.matmul(drop_attn, value))
            output = self.final_linear(context)
            return output
        else:
            context = torch.matmul(drop_attn, value)
            return context


class DecoderState(object):
    """Interface for grouping together the current state of a recurrent
    decoder. In the simplest case just represents the hidden state of
    the model.  But can also be used for implementing various forms of
    input_feeding and non-recurrent models.

    Modules need to implement this to utilize beam search decoding.
    """

    def detach(self):
        """Need to document this"""
        self.hidden = tuple([_.detach() for _ in self.hidden])
        self.input_feed = self.input_feed.detach()

    def beam_update(self, idx, positions, beam_size):
        """Need to document this"""
        for e in self._all:
            sizes = e.size()
            br = sizes[1]
            if len(sizes) == 3:
                sent_states = e.view(sizes[0], beam_size, br // beam_size, sizes[2])[:, :, idx]
            else:
                sent_states = e.view(sizes[0], beam_size, br // beam_size, sizes[2], sizes[3])[:, :, idx]

            sent_states.data.copy_(sent_states.data.index_select(1, positions))

    def map_batch_fn(self, fn):
        raise NotImplementedError()


class TransformerDecoderState(DecoderState):
    """Transformer Decoder state base class"""

    def __init__(self, src):
        """
        Args:
            src (FloatTensor): a sequence of source words tensors
                    with optional feature tensors, of size (len x batch).
        """
        self.src = src
        self.previous_input = None
        self.previous_layer_inputs = None
        self.cache = None

    @property
    def _all(self):
        """
        Contains attributes that need to be updated in self.beam_update().
        """
        if self.previous_input is not None and self.previous_layer_inputs is not None:
            return (self.previous_input, self.previous_layer_inputs, self.src)
        else:
            return (self.src,)

    def detach(self):
        if self.previous_input is not None:
            self.previous_input = self.previous_input.detach()
        if self.previous_layer_inputs is not None:
            self.previous_layer_inputs = self.previous_layer_inputs.detach()
        self.src = self.src.detach()

    def update_state(self, new_input, previous_layer_inputs):
        state = TransformerDecoderState(self.src)
        state.previous_input = new_input
        state.previous_layer_inputs = previous_layer_inputs
        return state

    def _init_cache(self, memory_bank, num_layers):
        self.cache = {}

        for l in range(num_layers):
            layer_cache = {"memory_keys": None, "memory_values": None}
            layer_cache["self_keys"] = None
            layer_cache["self_values"] = None
            self.cache["layer_{}".format(l)] = layer_cache

    def repeat_beam_size_times(self, beam_size):
        """Repeat beam_size times along batch dimension."""
        self.src = self.src.data.repeat(1, beam_size, 1)

    def map_batch_fn(self, fn):
        def _recursive_map(struct, batch_dim=0):
            for k, v in struct.items():
                if v is not None:
                    if isinstance(v, dict):
                        _recursive_map(v)
                    else:
                        struct[k] = fn(v, batch_dim)

        self.src = fn(self.src, 0)
        if self.cache is not None:
            _recursive_map(self.cache)


def gelu(x):
    return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))


class PositionwiseFeedForward(nn.Module):
    """A two-layer Feed-Forward-Network with residual layer norm.

    Args:
        d_model (int): the size of input for the first-layer of the FFN.
        d_ff (int): the hidden layer size of the second-layer
            of the FNN.
        dropout (float): dropout probability in :math:`[0, 1)`.
    """

    def __init__(self, d_model, d_ff, dropout=0.1):
        super().__init__()
        self.w_1 = nn.Linear(d_model, d_ff)
        self.w_2 = nn.Linear(d_ff, d_model)
        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
        self.actv = gelu
        self.dropout_1 = nn.Dropout(dropout)
        self.dropout_2 = nn.Dropout(dropout)

    def forward(self, x):
        inter = self.dropout_1(self.actv(self.w_1(self.layer_norm(x))))
        output = self.dropout_2(self.w_2(inter))
        return output + x


#
# TRANSLATOR
# The following code is used to generate summaries using the
# pre-trained weights and beam search.
#


def build_predictor(args, tokenizer, symbols, model, logger=None):
    # we should be able to refactor the global scorer a lot
    scorer = GNMTGlobalScorer(args.alpha, length_penalty="wu")
    translator = Translator(args, model, tokenizer, symbols, global_scorer=scorer, logger=logger)
    return translator


class GNMTGlobalScorer(object):
    """
    NMT re-ranking score from
    "Google's Neural Machine Translation System" :cite:`wu2016google`

    Args:
       alpha (float): length parameter
       beta (float):  coverage parameter
    """

    def __init__(self, alpha, length_penalty):
        self.alpha = alpha
        penalty_builder = PenaltyBuilder(length_penalty)
        self.length_penalty = penalty_builder.length_penalty()

    def score(self, beam, logprobs):
        """
        Rescores a prediction based on penalty functions
        """
        normalized_probs = self.length_penalty(beam, logprobs, self.alpha)
        return normalized_probs


class PenaltyBuilder(object):
    """
    Returns the Length and Coverage Penalty function for Beam Search.

    Args:
        length_pen (str): option name of length pen
        cov_pen (str): option name of cov pen
    """

    def __init__(self, length_pen):
        self.length_pen = length_pen

    def length_penalty(self):
        if self.length_pen == "wu":
            return self.length_wu
        elif self.length_pen == "avg":
            return self.length_average
        else:
            return self.length_none

    """
    Below are all the different penalty terms implemented so far
    """

    def length_wu(self, beam, logprobs, alpha=0.0):
        """
        NMT length re-ranking score from
        "Google's Neural Machine Translation System" :cite:`wu2016google`.
        """

        modifier = ((5 + len(beam.next_ys)) ** alpha) / ((5 + 1) ** alpha)
        return logprobs / modifier

    def length_average(self, beam, logprobs, alpha=0.0):
        """
        Returns the average probability of tokens in a sequence.
        """
        return logprobs / len(beam.next_ys)

    def length_none(self, beam, logprobs, alpha=0.0, beta=0.0):
        """
        Returns unmodified scores.
        """
        return logprobs


class Translator(object):
    """
    Uses a model to translate a batch of sentences.

    Args:
       model (:obj:`onmt.modules.NMTModel`):
          NMT model to use for translation
       fields (dict of Fields): data fields
       beam_size (int): size of beam to use
       n_best (int): number of translations produced
       max_length (int): maximum length output to produce
       global_scores (:obj:`GlobalScorer`):
         object to rescore final translations
       copy_attn (bool): use copy attention during translation
       beam_trace (bool): trace beam search for debugging
       logger(logging.Logger): logger.
    """

    def __init__(self, args, model, vocab, symbols, global_scorer=None, logger=None):
        self.logger = logger

        self.args = args
        self.model = model
        self.generator = self.model.generator
        self.vocab = vocab
        self.symbols = symbols
        self.start_token = symbols["BOS"]
        self.end_token = symbols["EOS"]

        self.global_scorer = global_scorer
        self.beam_size = args.beam_size
        self.min_length = args.min_length
        self.max_length = args.max_length

    def translate(self, batch, step, attn_debug=False):
        """Generates summaries from one batch of data."""
        self.model.eval()
        with torch.no_grad():
            batch_data = self.translate_batch(batch)
            translations = self.from_batch(batch_data)
        return translations

    def translate_batch(self, batch, fast=False):
        """
        Translate a batch of sentences.

        Mostly a wrapper around :obj:`Beam`.

        Args:
           batch (:obj:`Batch`): a batch from a dataset object
           fast (bool): enables fast beam search (may not support all features)
        """
        with torch.no_grad():
            return self._fast_translate_batch(batch, self.max_length, min_length=self.min_length)

    # Where the beam search lives
    # I have no idea why it is being called from the method above
    def _fast_translate_batch(self, batch, max_length, min_length=0):
        """Beam Search using the encoder inputs contained in `batch`."""

        # The batch object is funny
        # Instead of just looking at the size of the arguments we encapsulate
        # a size argument.
        # Where is it defined?
        beam_size = self.beam_size
        batch_size = batch.batch_size
        src = batch.src
        segs = batch.segs
        mask_src = batch.mask_src

        src_features = self.model.bert(src, segs, mask_src)
        dec_states = self.model.decoder.init_decoder_state(src, src_features, with_cache=True)
        device = src_features.device

        # Tile states and memory beam_size times.
        dec_states.map_batch_fn(lambda state, dim: tile(state, beam_size, dim=dim))
        src_features = tile(src_features, beam_size, dim=0)
        batch_offset = torch.arange(batch_size, dtype=torch.long, device=device)
        beam_offset = torch.arange(0, batch_size * beam_size, step=beam_size, dtype=torch.long, device=device)
        alive_seq = torch.full([batch_size * beam_size, 1], self.start_token, dtype=torch.long, device=device)

        # Give full probability to the first beam on the first step.
        topk_log_probs = torch.tensor([0.0] + [float("-inf")] * (beam_size - 1), device=device).repeat(batch_size)

        # Structure that holds finished hypotheses.
        hypotheses = [[] for _ in range(batch_size)]  # noqa: F812

        results = {}
        results["predictions"] = [[] for _ in range(batch_size)]  # noqa: F812
        results["scores"] = [[] for _ in range(batch_size)]  # noqa: F812
        results["gold_score"] = [0] * batch_size
        results["batch"] = batch

        for step in range(max_length):
            decoder_input = alive_seq[:, -1].view(1, -1)

            # Decoder forward.
            decoder_input = decoder_input.transpose(0, 1)

            dec_out, dec_states = self.model.decoder(decoder_input, src_features, dec_states, step=step)

            # Generator forward.
            log_probs = self.generator(dec_out.transpose(0, 1).squeeze(0))
            vocab_size = log_probs.size(-1)

            if step < min_length:
                log_probs[:, self.end_token] = -1e20

            # Multiply probs by the beam probability.
            log_probs += topk_log_probs.view(-1).unsqueeze(1)

            alpha = self.global_scorer.alpha
            length_penalty = ((5.0 + (step + 1)) / 6.0) ** alpha

            # Flatten probs into a list of possibilities.
            curr_scores = log_probs / length_penalty

            if self.args.block_trigram:
                cur_len = alive_seq.size(1)
                if cur_len > 3:
                    for i in range(alive_seq.size(0)):
                        fail = False
                        words = [int(w) for w in alive_seq[i]]
                        words = [self.vocab.ids_to_tokens[w] for w in words]
                        words = " ".join(words).replace(" ##", "").split()
                        if len(words) <= 3:
                            continue
                        trigrams = [(words[i - 1], words[i], words[i + 1]) for i in range(1, len(words) - 1)]
                        trigram = tuple(trigrams[-1])
                        if trigram in trigrams[:-1]:
                            fail = True
                        if fail:
                            curr_scores[i] = -10e20

            curr_scores = curr_scores.reshape(-1, beam_size * vocab_size)
            topk_scores, topk_ids = curr_scores.topk(beam_size, dim=-1)

            # Recover log probs.
            topk_log_probs = topk_scores * length_penalty

            # Resolve beam origin and true word ids.
            topk_beam_index = topk_ids.div(vocab_size)
            topk_ids = topk_ids.fmod(vocab_size)

            # Map beam_index to batch_index in the flat representation.
            batch_index = topk_beam_index + beam_offset[: topk_beam_index.size(0)].unsqueeze(1)
            select_indices = batch_index.view(-1)

            # Append last prediction.
            alive_seq = torch.cat([alive_seq.index_select(0, select_indices), topk_ids.view(-1, 1)], -1)

            is_finished = topk_ids.eq(self.end_token)
            if step + 1 == max_length:
                is_finished.fill_(1)
            # End condition is top beam is finished.
            end_condition = is_finished[:, 0].eq(1)
            # Save finished hypotheses.
            if is_finished.any():
                predictions = alive_seq.view(-1, beam_size, alive_seq.size(-1))
                for i in range(is_finished.size(0)):
                    b = batch_offset[i]
                    if end_condition[i]:
                        is_finished[i].fill_(1)
                    finished_hyp = is_finished[i].nonzero().view(-1)
                    # Store finished hypotheses for this batch.
                    for j in finished_hyp:
                        hypotheses[b].append((topk_scores[i, j], predictions[i, j, 1:]))
                    # If the batch reached the end, save the n_best hypotheses.
                    if end_condition[i]:
                        best_hyp = sorted(hypotheses[b], key=lambda x: x[0], reverse=True)
                        score, pred = best_hyp[0]

                        results["scores"][b].append(score)
                        results["predictions"][b].append(pred)
                non_finished = end_condition.eq(0).nonzero().view(-1)
                # If all sentences are translated, no need to go further.
                if len(non_finished) == 0:
                    break
                # Remove finished batches for the next step.
                topk_log_probs = topk_log_probs.index_select(0, non_finished)
                batch_index = batch_index.index_select(0, non_finished)
                batch_offset = batch_offset.index_select(0, non_finished)
                alive_seq = predictions.index_select(0, non_finished).view(-1, alive_seq.size(-1))
            # Reorder states.
            select_indices = batch_index.view(-1)
            src_features = src_features.index_select(0, select_indices)
            dec_states.map_batch_fn(lambda state, dim: state.index_select(dim, select_indices))

        return results

    def from_batch(self, translation_batch):
        batch = translation_batch["batch"]
        assert len(translation_batch["gold_score"]) == len(translation_batch["predictions"])
        batch_size = batch.batch_size

        preds, _, _, tgt_str, src = (
            translation_batch["predictions"],
            translation_batch["scores"],
            translation_batch["gold_score"],
            batch.tgt_str,
            batch.src,
        )

        translations = []
        for b in range(batch_size):
            pred_sents = self.vocab.convert_ids_to_tokens([int(n) for n in preds[b][0]])
            pred_sents = " ".join(pred_sents).replace(" ##", "")
            gold_sent = " ".join(tgt_str[b].split())
            raw_src = [self.vocab.ids_to_tokens[int(t)] for t in src[b]][:500]
            raw_src = " ".join(raw_src)
            translation = (pred_sents, gold_sent, raw_src)
            translations.append(translation)

        return translations


def tile(x, count, dim=0):
    """
    Tiles x on dimension dim count times.
    """
    perm = list(range(len(x.size())))
    if dim != 0:
        perm[0], perm[dim] = perm[dim], perm[0]
        x = x.permute(perm).contiguous()
    out_size = list(x.size())
    out_size[0] *= count
    batch = x.size(0)
    x = x.view(batch, -1).transpose(0, 1).repeat(count, 1).transpose(0, 1).contiguous().view(*out_size)
    if dim != 0:
        x = x.permute(perm).contiguous()
    return x


#
# Optimizer for training. We keep this here in case we want to add
# a finetuning script.
#


class BertSumOptimizer(object):
    """Specific optimizer for BertSum.

    As described in [1], the authors fine-tune BertSum for abstractive
    summarization using two Adam Optimizers with different warm-up steps and
    learning rate. They also use a custom learning rate scheduler.

    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
        arXiv preprint arXiv:1908.08345 (2019).
    """

    def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
        self.encoder = model.encoder
        self.decoder = model.decoder
        self.lr = lr
        self.warmup_steps = warmup_steps

        self.optimizers = {
            "encoder": torch.optim.Adam(
                model.encoder.parameters(),
                lr=lr["encoder"],
                betas=(beta_1, beta_2),
                eps=eps,
            ),
            "decoder": torch.optim.Adam(
                model.decoder.parameters(),
                lr=lr["decoder"],
                betas=(beta_1, beta_2),
                eps=eps,
            ),
        }

        self._step = 0
        self.current_learning_rates = {}

    def _update_rate(self, stack):
        return self.lr[stack] * min(self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-1.5))

    def zero_grad(self):
        self.optimizer_decoder.zero_grad()
        self.optimizer_encoder.zero_grad()

    def step(self):
        self._step += 1
        for stack, optimizer in self.optimizers.items():
            new_rate = self._update_rate(stack)
            for param_group in optimizer.param_groups:
                param_group["lr"] = new_rate
            optimizer.step()
            self.current_learning_rates[stack] = new_rate