# Copyright 2019 Shigeki Karita
# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""Encoder definition."""
from typing import Optional, Tuple
import torch
from typeguard import typechecked
from espnet2.asr.postencoder.abs_postencoder import AbsPostEncoder
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet.nets.pytorch_backend.transformer.attention import MultiHeadedAttention
from espnet.nets.pytorch_backend.transformer.embedding import PositionalEncoding
from espnet.nets.pytorch_backend.transformer.encoder_layer import EncoderLayer
from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
from espnet.nets.pytorch_backend.transformer.multi_layer_conv import (
Conv1dLinear,
MultiLayeredConv1d,
)
from espnet.nets.pytorch_backend.transformer.positionwise_feed_forward import (
PositionwiseFeedForward,
)
from espnet.nets.pytorch_backend.transformer.repeat import repeat
[docs]class TransformerPostEncoder(AbsPostEncoder):
"""Transformer encoder module.
Args:
input_size: input dim
output_size: dimension of attention
attention_heads: the number of heads of multi head attention
linear_units: the number of units of position-wise feed forward
num_blocks: the number of decoder blocks
dropout_rate: dropout rate
attention_dropout_rate: dropout rate in attention
positional_dropout_rate: dropout rate after adding positional encoding
input_layer: input layer type
pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
normalize_before: whether to use layer_norm before the first block
concat_after: 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: linear of conv1d
positionwise_conv_kernel_size: kernel size of positionwise conv1d layer
padding_idx: padding_idx for input_layer=embed
"""
@typechecked
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: Optional[str] = "linear",
pos_enc_class=PositionalEncoding,
normalize_before: bool = True,
concat_after: bool = False,
positionwise_layer_type: str = "linear",
positionwise_conv_kernel_size: int = 1,
padding_idx: int = -1,
):
super().__init__()
self._output_size = output_size
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),
torch.nn.ReLU(),
pos_enc_class(output_size, positional_dropout_rate),
)
elif input_layer == "None":
self.embed = torch.nn.Sequential(
torch.nn.Linear(input_size, output_size),
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,
)
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.")
self.encoders = repeat(
num_blocks,
lambda lnum: EncoderLayer(
output_size,
MultiHeadedAttention(
attention_heads, output_size, attention_dropout_rate
),
positionwise_layer(*positionwise_layer_args),
dropout_rate,
normalize_before,
concat_after,
),
)
if self.normalize_before:
self.after_norm = LayerNorm(output_size)
[docs] def output_size(self) -> int:
return self._output_size
[docs] def forward(
self,
xs_pad: torch.Tensor,
ilens: torch.Tensor,
prev_states: torch.Tensor = None,
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
"""Embed positions in tensor.
Args:
xs_pad: input tensor (B, L, D)
ilens: input length (B)
prev_states: Not to be used now.
Returns:
position embedded tensor and mask
"""
masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
xs_pad = self.embed(xs_pad)
xs_pad, masks = self.encoders(xs_pad, masks)
if self.normalize_before:
xs_pad = self.after_norm(xs_pad)
olens = masks.squeeze(1).sum(1)
return xs_pad, olens