espnet2/asr/ctc.py

import logging

import torch
import torch.nn.functional as F
from typeguard import check_argument_types


class CTC(torch.nn.Module):
    """CTC module.

    Args:
        odim: dimension of outputs
        encoder_output_sizse: number of encoder projection units
        dropout_rate: dropout rate (0.0 ~ 1.0)
        ctc_type: builtin or warpctc
        reduce: reduce the CTC loss into a scalar
    """

    def __init__(
        self,
        odim: int,
        encoder_output_sizse: int,
        dropout_rate: float = 0.0,
        ctc_type: str = "builtin",
        reduce: bool = True,
        ignore_nan_grad: bool = False,
    ):
        assert check_argument_types()
        super().__init__()
        eprojs = encoder_output_sizse
        self.dropout_rate = dropout_rate
        self.ctc_lo = torch.nn.Linear(eprojs, odim)
        self.ctc_type = ctc_type
        self.ignore_nan_grad = ignore_nan_grad

        if self.ctc_type == "builtin":
            self.ctc_loss = torch.nn.CTCLoss(reduction="none")
        elif self.ctc_type == "warpctc":
            import warpctc_pytorch as warp_ctc

            if ignore_nan_grad:
                raise NotImplementedError(
                    "ignore_nan_grad option is not supported for warp_ctc"
                )
            self.ctc_loss = warp_ctc.CTCLoss(size_average=True, reduce=reduce)
        else:
            raise ValueError(
                f'ctc_type must be "builtin" or "warpctc": {self.ctc_type}'
            )

        self.reduce = reduce

    def loss_fn(self, th_pred, th_target, th_ilen, th_olen) -> torch.Tensor:
        if self.ctc_type == "builtin":
            th_pred = th_pred.log_softmax(2)
            loss = self.ctc_loss(th_pred, th_target, th_ilen, th_olen)

            if loss.requires_grad and self.ignore_nan_grad:
                # ctc_grad: (L, B, O)
                ctc_grad = loss.grad_fn(torch.ones_like(loss))
                ctc_grad = ctc_grad.sum([0, 2])
                indices = torch.isfinite(ctc_grad)
                size = indices.long().sum()
                if size == 0:
                    # Return as is
                    logging.warning(
                        "All samples in this mini-batch got nan grad."
                        " Returning nan value instead of CTC loss"
                    )
                elif size != th_pred.size(1):
                    logging.warning(
                        f"{th_pred.size(1) - size}/{th_pred.size(1)}"
                        " samples got nan grad."
                        " These were ignored for CTC loss."
                    )

                    # Create mask for target
                    target_mask = torch.full(
                        [th_target.size(0)],
                        1,
                        dtype=torch.bool,
                        device=th_target.device,
                    )
                    s = 0
                    for ind, le in enumerate(th_olen):
                        if not indices[ind]:
                            target_mask[s : s + le] = 0
                        s += le

                    # Calc loss again using maksed data
                    loss = self.ctc_loss(
                        th_pred[:, indices, :],
                        th_target[target_mask],
                        th_ilen[indices],
                        th_olen[indices],
                    )
            else:
                size = th_pred.size(1)

            if self.reduce:
                # Batch-size average
                loss = loss.sum() / size
            else:
                loss = loss / size
            return loss

        elif self.ctc_type == "warpctc":
            # warpctc only supports float32
            th_pred = th_pred.to(dtype=torch.float32)

            th_target = th_target.cpu().int()
            th_ilen = th_ilen.cpu().int()
            th_olen = th_olen.cpu().int()
            loss = self.ctc_loss(th_pred, th_target, th_ilen, th_olen)
            if self.reduce:
                # NOTE: sum() is needed to keep consistency since warpctc
                # return as tensor w/ shape (1,)
                # but builtin return as tensor w/o shape (scalar).
                loss = loss.sum()
            return loss
        else:
            raise NotImplementedError

    def forward(self, hs_pad, hlens, ys_pad, ys_lens):
        """Calculate CTC loss.

        Args:
            hs_pad: batch of padded hidden state sequences (B, Tmax, D)
            hlens: batch of lengths of hidden state sequences (B)
            ys_pad: batch of padded character id sequence tensor (B, Lmax)
            ys_lens: batch of lengths of character sequence (B)
        """
        # hs_pad: (B, L, NProj) -> ys_hat: (B, L, Nvocab)
        ys_hat = self.ctc_lo(F.dropout(hs_pad, p=self.dropout_rate))
        # ys_hat: (B, L, D) -> (L, B, D)
        ys_hat = ys_hat.transpose(0, 1)

        # (B, L) -> (BxL,)
        ys_true = torch.cat([ys_pad[i, :l] for i, l in enumerate(ys_lens)])

        loss = self.loss_fn(ys_hat, ys_true, hlens, ys_lens).to(
            device=hs_pad.device, dtype=hs_pad.dtype
        )

        return loss

    def log_softmax(self, hs_pad):
        """log_softmax of frame activations

        Args:
            Tensor hs_pad: 3d tensor (B, Tmax, eprojs)
        Returns:
            torch.Tensor: log softmax applied 3d tensor (B, Tmax, odim)
        """
        return F.log_softmax(self.ctc_lo(hs_pad), dim=2)

    def argmax(self, hs_pad):
        """argmax of frame activations

        Args:
            torch.Tensor hs_pad: 3d tensor (B, Tmax, eprojs)
        Returns:
            torch.Tensor: argmax applied 2d tensor (B, Tmax)
        """
        return torch.argmax(self.ctc_lo(hs_pad), dim=2)