model.py

# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.

import torch
import torch.nn as nn
import torch
from torch.autograd import Variable
import copy


class Seq2Seq(nn.Module):
    """
    Build Seqence-to-Sequence.

    Parameters:

    * `encoder`- encoder of seq2seq model. e.g. roberta
    * `decoder`- decoder of seq2seq model. e.g. transformer
    * `config`- configuration of encoder model.
    * `beam_size`- beam size for beam search.
    * `max_length`- max length of target for beam search.
    * `sos_id`- start of symbol ids in target for beam search.
    * `eos_id`- end of symbol ids in target for beam search.
    """

    def __init__(
        self,
        encoder,
        decoder,
        config,
        beam_size=None,
        max_length=None,
        sos_id=None,
        eos_id=None,
    ):
        super(Seq2Seq, self).__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.config = config
        self.register_buffer("bias", torch.tril(torch.ones(2048, 2048)))
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.lsm = nn.LogSoftmax(dim=-1)
        self.tie_weights()

        self.beam_size = beam_size
        self.max_length = max_length
        self.sos_id = sos_id
        self.eos_id = eos_id

    def _tie_or_clone_weights(self, first_module, second_module):
        """Tie or clone module weights depending of weither we are using TorchScript or not"""
        if self.config.torchscript:
            first_module.weight = nn.Parameter(second_module.weight.clone())
        else:
            first_module.weight = second_module.weight

    def tie_weights(self):
        """Make sure we are sharing the input and output embeddings.
        Export to TorchScript can't handle parameter sharing so we are cloning them instead.
        """
        self._tie_or_clone_weights(
            self.lm_head, self.encoder.embeddings.word_embeddings
        )

    def forward(
        self,
        source_ids=None,
        source_mask=None,
        target_ids=None,
        target_mask=None,
        args=None,
    ):
        outputs = self.encoder(source_ids, attention_mask=source_mask)
        encoder_output = outputs[0].permute([1, 0, 2]).contiguous()
        if target_ids is not None:
            attn_mask = -1e4 * (
                1 - self.bias[: target_ids.shape[1], : target_ids.shape[1]]
            )
            tgt_embeddings = (
                self.encoder.embeddings(target_ids).permute([1, 0, 2]).contiguous()
            )
            out = self.decoder(
                tgt_embeddings,
                encoder_output,
                tgt_mask=attn_mask,
                memory_key_padding_mask=(1 - source_mask).bool(),
            )
            hidden_states = torch.tanh(self.dense(out)).permute([1, 0, 2]).contiguous()
            lm_logits = self.lm_head(hidden_states)
            # Shift so that tokens < n predict n
            active_loss = target_mask[..., 1:].ne(0).view(-1) == 1
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = target_ids[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1))[active_loss],
                shift_labels.view(-1)[active_loss],
            )

            outputs = loss, loss * active_loss.sum(), active_loss.sum()
            return outputs
        else:
            # Predict
            preds = []
            try:
                zero = torch.cuda.LongTensor(1).fill_(0)
            except Exception as e:
                zero = torch.LongTensor(1).fill_(0)
            for i in range(source_ids.shape[0]):
                context = encoder_output[:, i : i + 1]
                context_mask = source_mask[i : i + 1, :]
                beam = Beam(self.beam_size, self.sos_id, self.eos_id)
                input_ids = beam.getCurrentState()
                context = context.repeat(1, self.beam_size, 1)
                context_mask = context_mask.repeat(self.beam_size, 1)
                for _ in range(self.max_length):
                    if beam.done():
                        break
                    attn_mask = -1e4 * (
                        1 - self.bias[: input_ids.shape[1], : input_ids.shape[1]]
                    )
                    tgt_embeddings = (
                        self.encoder.embeddings(input_ids)
                        .permute([1, 0, 2])
                        .contiguous()
                    )
                    out = self.decoder(
                        tgt_embeddings,
                        context,
                        tgt_mask=attn_mask,
                        memory_key_padding_mask=(1 - context_mask).bool(),
                    )
                    out = torch.tanh(self.dense(out))
                    hidden_states = out.permute([1, 0, 2]).contiguous()[:, -1, :]
                    out = self.lsm(self.lm_head(hidden_states)).data
                    beam.advance(out)
                    input_ids.data.copy_(
                        input_ids.data.index_select(0, beam.getCurrentOrigin())
                    )
                    input_ids = torch.cat((input_ids, beam.getCurrentState()), -1)
                hyp = beam.getHyp(beam.getFinal())
                pred = beam.buildTargetTokens(hyp)[: self.beam_size]
                pred = [
                    torch.cat(
                        [x.view(-1) for x in p] + [zero] * (self.max_length - len(p))
                    ).view(1, -1)
                    for p in pred
                ]
                preds.append(torch.cat(pred, 0).unsqueeze(0))

            preds = torch.cat(preds, 0)
            return preds


class Beam(object):
    def __init__(self, size, sos, eos):
        self.size = size
        if torch.cuda.is_available():
            self.tt = torch.cuda
        else:
            self.tt = torch
        # The score for each translation on the beam.
        self.scores = self.tt.FloatTensor(size).zero_()
        # The backpointers at each time-step.
        self.prevKs = []
        # The outputs at each time-step.
        self.nextYs = [self.tt.LongTensor(size).fill_(0)]
        self.nextYs[0][0] = sos
        # Has EOS topped the beam yet.
        self._eos = eos
        self.eosTop = False
        # Time and k pair for finished.
        self.finished = []

    def getCurrentState(self):
        "Get the outputs for the current timestep."
        batch = self.tt.LongTensor(self.nextYs[-1]).view(-1, 1)
        return batch

    def getCurrentOrigin(self):
        "Get the backpointers for the current timestep."
        return self.prevKs[-1]

    def advance(self, wordLk):
        """
        Given prob over words for every last beam `wordLk` and attention
        `attnOut`: Compute and update the beam search.

        Parameters:

        * `wordLk`- probs of advancing from the last step (K x words)
        * `attnOut`- attention at the last step

        Returns: True if beam search is complete.
        """
        numWords = wordLk.size(1)

        # Sum the previous scores.
        if len(self.prevKs) > 0:
            beamLk = wordLk + self.scores.unsqueeze(1).expand_as(wordLk)

            # Don't let EOS have children.
            for i in range(self.nextYs[-1].size(0)):
                if self.nextYs[-1][i] == self._eos:
                    beamLk[i] = -1e20
        else:
            beamLk = wordLk[0]
        flatBeamLk = beamLk.view(-1)
        bestScores, bestScoresId = flatBeamLk.topk(self.size, 0, True, True)

        self.scores = bestScores

        # bestScoresId is flattened beam x word array, so calculate which
        # word and beam each score came from
        prevK = bestScoresId // numWords
        self.prevKs.append(prevK)
        self.nextYs.append((bestScoresId - prevK * numWords))

        for i in range(self.nextYs[-1].size(0)):
            if self.nextYs[-1][i] == self._eos:
                s = self.scores[i]
                self.finished.append((s, len(self.nextYs) - 1, i))

        # End condition is when top-of-beam is EOS and no global score.
        if self.nextYs[-1][0] == self._eos:
            self.eosTop = True

    def done(self):
        return self.eosTop and len(self.finished) >= self.size

    def getFinal(self):
        if len(self.finished) == 0:
            self.finished.append((self.scores[0], len(self.nextYs) - 1, 0))
        self.finished.sort(key=lambda a: -a[0])
        if len(self.finished) != self.size:
            unfinished = []
            for i in range(self.nextYs[-1].size(0)):
                if self.nextYs[-1][i] != self._eos:
                    s = self.scores[i]
                    unfinished.append((s, len(self.nextYs) - 1, i))
            unfinished.sort(key=lambda a: -a[0])
            self.finished += unfinished[: self.size - len(self.finished)]
        return self.finished[: self.size]

    def getHyp(self, beam_res):
        """
        Walk back to construct the full hypothesis.
        """
        hyps = []
        for _, timestep, k in beam_res:
            hyp = []
            for j in range(len(self.prevKs[:timestep]) - 1, -1, -1):
                hyp.append(self.nextYs[j + 1][k])
                k = self.prevKs[j][k]
            hyps.append(hyp[::-1])
        return hyps

    def buildTargetTokens(self, preds):
        sentence = []
        for pred in preds:
            tokens = []
            for tok in pred:
                if tok == self._eos:
                    break
                tokens.append(tok)
            sentence.append(tokens)
        return sentence