espnet2/asr/encoder/conformer_encoder.py

# Copyright 2020 Tomoki Hayashi
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

"""Conformer encoder definition."""

from typing import Optional
from typing import Tuple

import logging
import torch

from typeguard import check_argument_types

from espnet.nets.pytorch_backend.conformer.convolution import ConvolutionModule
from espnet.nets.pytorch_backend.conformer.encoder_layer import EncoderLayer
from espnet.nets.pytorch_backend.nets_utils import get_activation
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet.nets.pytorch_backend.transformer.attention import (
    MultiHeadedAttention,  # noqa: H301
    RelPositionMultiHeadedAttention,  # noqa: H301
    LegacyRelPositionMultiHeadedAttention,  # noqa: H301
)
from espnet.nets.pytorch_backend.transformer.embedding import (
    PositionalEncoding,  # noqa: H301
    ScaledPositionalEncoding,  # noqa: H301
    RelPositionalEncoding,  # noqa: H301
    LegacyRelPositionalEncoding,  # noqa: H301
)
from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
from espnet.nets.pytorch_backend.transformer.multi_layer_conv import Conv1dLinear
from espnet.nets.pytorch_backend.transformer.multi_layer_conv import MultiLayeredConv1d
from espnet.nets.pytorch_backend.transformer.positionwise_feed_forward import (
    PositionwiseFeedForward,  # noqa: H301
)
from espnet.nets.pytorch_backend.transformer.repeat import repeat
from espnet.nets.pytorch_backend.transformer.subsampling import check_short_utt
from espnet.nets.pytorch_backend.transformer.subsampling import Conv2dSubsampling
from espnet.nets.pytorch_backend.transformer.subsampling import Conv2dSubsampling6
from espnet.nets.pytorch_backend.transformer.subsampling import Conv2dSubsampling8
from espnet.nets.pytorch_backend.transformer.subsampling import TooShortUttError
from espnet2.asr.encoder.abs_encoder import AbsEncoder


class ConformerEncoder(AbsEncoder):
    """Conformer encoder module.

    Args:
        input_size (int): Input dimension.
        output_size (int): Dimention of attention.
        attention_heads (int): The number of heads of multi head attention.
        linear_units (int): The number of units of position-wise feed forward.
        num_blocks (int): The number of decoder blocks.
        dropout_rate (float): Dropout rate.
        attention_dropout_rate (float): Dropout rate in attention.
        positional_dropout_rate (float): Dropout rate after adding positional encoding.
        input_layer (Union[str, torch.nn.Module]): Input layer type.
        normalize_before (bool): Whether to use layer_norm before the first block.
        concat_after (bool): Whether to concat attention layer's input and output.
            If True, additional linear will be applied.
            i.e. x -> x + linear(concat(x, att(x)))
            If False, no additional linear will be applied. i.e. x -> x + att(x)
        positionwise_layer_type (str): "linear", "conv1d", or "conv1d-linear".
        positionwise_conv_kernel_size (int): Kernel size of positionwise conv1d layer.
        rel_pos_type (str): Whether to use the latest relative positional encoding or
            the legacy one. The legacy relative positional encoding will be deprecated
            in the future. More Details can be found in
            https://github.com/espnet/espnet/pull/2816.
        encoder_pos_enc_layer_type (str): Encoder positional encoding layer type.
        encoder_attn_layer_type (str): Encoder attention layer type.
        activation_type (str): Encoder activation function type.
        macaron_style (bool): Whether to use macaron style for positionwise layer.
        use_cnn_module (bool): Whether to use convolution module.
        zero_triu (bool): Whether to zero the upper triangular part of attention matrix.
        cnn_module_kernel (int): Kernerl size of convolution module.
        padding_idx (int): Padding idx for input_layer=embed.

    """

    def __init__(
        self,
        input_size: int,
        output_size: int = 256,
        attention_heads: int = 4,
        linear_units: int = 2048,
        num_blocks: int = 6,
        dropout_rate: float = 0.1,
        positional_dropout_rate: float = 0.1,
        attention_dropout_rate: float = 0.0,
        input_layer: str = "conv2d",
        normalize_before: bool = True,
        concat_after: bool = False,
        positionwise_layer_type: str = "linear",
        positionwise_conv_kernel_size: int = 3,
        macaron_style: bool = False,
        rel_pos_type: str = "legacy",
        pos_enc_layer_type: str = "rel_pos",
        selfattention_layer_type: str = "rel_selfattn",
        activation_type: str = "swish",
        use_cnn_module: bool = True,
        zero_triu: bool = False,
        cnn_module_kernel: int = 31,
        padding_idx: int = -1,
    ):
        assert check_argument_types()
        super().__init__()
        self._output_size = output_size

        if rel_pos_type == "legacy":
            if pos_enc_layer_type == "rel_pos":
                pos_enc_layer_type = "legacy_rel_pos"
            if selfattention_layer_type == "rel_selfattn":
                selfattention_layer_type = "legacy_rel_selfattn"
        elif rel_pos_type == "latest":
            assert selfattention_layer_type != "legacy_rel_selfattn"
            assert pos_enc_layer_type != "legacy_rel_pos"
        else:
            raise ValueError("unknown rel_pos_type: " + rel_pos_type)

        activation = get_activation(activation_type)
        if pos_enc_layer_type == "abs_pos":
            pos_enc_class = PositionalEncoding
        elif pos_enc_layer_type == "scaled_abs_pos":
            pos_enc_class = ScaledPositionalEncoding
        elif pos_enc_layer_type == "rel_pos":
            assert selfattention_layer_type == "rel_selfattn"
            pos_enc_class = RelPositionalEncoding
        elif pos_enc_layer_type == "legacy_rel_pos":
            assert selfattention_layer_type == "legacy_rel_selfattn"
            pos_enc_class = LegacyRelPositionalEncoding
            logging.warning(
                "Using legacy_rel_pos and it will be deprecated in the future."
            )
        else:
            raise ValueError("unknown pos_enc_layer: " + pos_enc_layer_type)

        if input_layer == "linear":
            self.embed = torch.nn.Sequential(
                torch.nn.Linear(input_size, output_size),
                torch.nn.LayerNorm(output_size),
                torch.nn.Dropout(dropout_rate),
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "conv2d":
            self.embed = Conv2dSubsampling(
                input_size,
                output_size,
                dropout_rate,
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "conv2d6":
            self.embed = Conv2dSubsampling6(
                input_size,
                output_size,
                dropout_rate,
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "conv2d8":
            self.embed = Conv2dSubsampling8(
                input_size,
                output_size,
                dropout_rate,
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer == "embed":
            self.embed = torch.nn.Sequential(
                torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif isinstance(input_layer, torch.nn.Module):
            self.embed = torch.nn.Sequential(
                input_layer,
                pos_enc_class(output_size, positional_dropout_rate),
            )
        elif input_layer is None:
            self.embed = torch.nn.Sequential(
                pos_enc_class(output_size, positional_dropout_rate)
            )
        else:
            raise ValueError("unknown input_layer: " + input_layer)
        self.normalize_before = normalize_before
        if positionwise_layer_type == "linear":
            positionwise_layer = PositionwiseFeedForward
            positionwise_layer_args = (
                output_size,
                linear_units,
                dropout_rate,
                activation,
            )
        elif positionwise_layer_type == "conv1d":
            positionwise_layer = MultiLayeredConv1d
            positionwise_layer_args = (
                output_size,
                linear_units,
                positionwise_conv_kernel_size,
                dropout_rate,
            )
        elif positionwise_layer_type == "conv1d-linear":
            positionwise_layer = Conv1dLinear
            positionwise_layer_args = (
                output_size,
                linear_units,
                positionwise_conv_kernel_size,
                dropout_rate,
            )
        else:
            raise NotImplementedError("Support only linear or conv1d.")

        if selfattention_layer_type == "selfattn":
            encoder_selfattn_layer = MultiHeadedAttention
            encoder_selfattn_layer_args = (
                attention_heads,
                output_size,
                attention_dropout_rate,
            )
        elif selfattention_layer_type == "legacy_rel_selfattn":
            assert pos_enc_layer_type == "legacy_rel_pos"
            encoder_selfattn_layer = LegacyRelPositionMultiHeadedAttention
            encoder_selfattn_layer_args = (
                attention_heads,
                output_size,
                attention_dropout_rate,
            )
            logging.warning(
                "Using legacy_rel_selfattn and it will be deprecated in the future."
            )
        elif selfattention_layer_type == "rel_selfattn":
            assert pos_enc_layer_type == "rel_pos"
            encoder_selfattn_layer = RelPositionMultiHeadedAttention
            encoder_selfattn_layer_args = (
                attention_heads,
                output_size,
                attention_dropout_rate,
                zero_triu,
            )
        else:
            raise ValueError("unknown encoder_attn_layer: " + selfattention_layer_type)

        convolution_layer = ConvolutionModule
        convolution_layer_args = (output_size, cnn_module_kernel, activation)

        self.encoders = repeat(
            num_blocks,
            lambda lnum: EncoderLayer(
                output_size,
                encoder_selfattn_layer(*encoder_selfattn_layer_args),
                positionwise_layer(*positionwise_layer_args),
                positionwise_layer(*positionwise_layer_args) if macaron_style else None,
                convolution_layer(*convolution_layer_args) if use_cnn_module else None,
                dropout_rate,
                normalize_before,
                concat_after,
            ),
        )
        if self.normalize_before:
            self.after_norm = LayerNorm(output_size)

    def output_size(self) -> int:
        return self._output_size

    def forward(
        self,
        xs_pad: torch.Tensor,
        ilens: torch.Tensor,
        prev_states: torch.Tensor = None,
    ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
        """Calculate forward propagation.

        Args:
            xs_pad (torch.Tensor): Input tensor (#batch, L, input_size).
            ilens (torch.Tensor): Input length (#batch).
            prev_states (torch.Tensor): Not to be used now.

        Returns:
            torch.Tensor: Output tensor (#batch, L, output_size).
            torch.Tensor: Output length (#batch).
            torch.Tensor: Not to be used now.

        """
        masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)

        if (
            isinstance(self.embed, Conv2dSubsampling)
            or isinstance(self.embed, Conv2dSubsampling6)
            or isinstance(self.embed, Conv2dSubsampling8)
        ):
            short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
            if short_status:
                raise TooShortUttError(
                    f"has {xs_pad.size(1)} frames and is too short for subsampling "
                    + f"(it needs more than {limit_size} frames), return empty results",
                    xs_pad.size(1),
                    limit_size,
                )
            xs_pad, masks = self.embed(xs_pad, masks)
        else:
            xs_pad = self.embed(xs_pad)
        xs_pad, masks = self.encoders(xs_pad, masks)
        if isinstance(xs_pad, tuple):
            xs_pad = xs_pad[0]
        if self.normalize_before:
            xs_pad = self.after_norm(xs_pad)

        olens = masks.squeeze(1).sum(1)
        return xs_pad, olens, None