# Copyright 2023 Jiatong Shi
# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
"""Linear encoder definition."""
from typing import Optional, Tuple
import torch
from typeguard import typechecked
from espnet2.asr.encoder.abs_encoder import AbsEncoder
from espnet.nets.pytorch_backend.nets_utils import make_pad_mask
from espnet.nets.pytorch_backend.transformer.layer_norm import LayerNorm
from espnet.nets.pytorch_backend.transformer.subsampling import (
Conv2dSubsampling,
Conv2dSubsampling2,
Conv2dSubsampling6,
Conv2dSubsampling8,
TooShortUttError,
check_short_utt,
)
[docs]class LinearEncoder(AbsEncoder):
"""Linear encoder module.
Args:
input_size: input dim
output_size: dimension of attention
linear_units: the number of units of position-wise feed forward
dropout_rate: dropout rate
input_layer: input layer type
normalize_before: whether to use layer_norm before the first block
padding_idx: padding_idx for input_layer=embed
"""
@typechecked
def __init__(
self,
input_size: int,
output_size: int = 256,
dropout_rate: float = 0.1,
input_layer: Optional[str] = "conv2d",
normalize_before: bool = True,
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(),
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
elif input_layer == "conv2d2":
self.embed = Conv2dSubsampling2(input_size, output_size, dropout_rate)
elif input_layer == "conv2d6":
self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
elif input_layer == "conv2d8":
self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
elif input_layer == "embed":
self.embed = (
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
)
elif input_layer is None:
if input_size == output_size:
self.embed = None
else:
self.embed = torch.nn.Linear(input_size, output_size)
else:
raise ValueError("unknown input_layer: " + input_layer)
self.normalize_before = normalize_before
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)
if self.embed is None:
xs_pad = xs_pad
elif (
isinstance(self.embed, Conv2dSubsampling)
or isinstance(self.embed, Conv2dSubsampling2)
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)
if self.normalize_before:
xs_pad = self.after_norm(xs_pad)
olens = masks.squeeze(1).sum(1)
return xs_pad, olens, None