File size: 5,452 Bytes
c14c4fe | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 | from typing import List
from typing import Optional
from typing import Sequence
from typing import Tuple
from typing import Union
import logging
import torch
import torch.nn as nn
from torch.nn import functional as F
import numpy as np
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from funasr.models.transformer.layer_norm import LayerNorm
from funasr.models.encoder.abs_encoder import AbsEncoder
import math
from funasr.models.transformer.utils.repeat import repeat
class EncoderLayer(nn.Module):
def __init__(
self,
input_units,
num_units,
kernel_size=3,
activation="tanh",
stride=1,
include_batch_norm=False,
residual=False,
):
super().__init__()
left_padding = math.ceil((kernel_size - stride) / 2)
right_padding = kernel_size - stride - left_padding
self.conv_padding = nn.ConstantPad1d((left_padding, right_padding), 0.0)
self.conv1d = nn.Conv1d(
input_units,
num_units,
kernel_size,
stride,
)
self.activation = self.get_activation(activation)
if include_batch_norm:
self.bn = nn.BatchNorm1d(num_units, momentum=0.99, eps=1e-3)
self.residual = residual
self.include_batch_norm = include_batch_norm
self.input_units = input_units
self.num_units = num_units
self.stride = stride
@staticmethod
def get_activation(activation):
if activation == "tanh":
return nn.Tanh()
else:
return nn.ReLU()
def forward(self, xs_pad, ilens=None):
outputs = self.conv1d(self.conv_padding(xs_pad))
if self.residual and self.stride == 1 and self.input_units == self.num_units:
outputs = outputs + xs_pad
if self.include_batch_norm:
outputs = self.bn(outputs)
# add parenthesis for repeat module
return self.activation(outputs), ilens
class ConvEncoder(AbsEncoder):
"""
Author: Speech Lab of DAMO Academy, Alibaba Group
Convolution encoder in OpenNMT framework
"""
def __init__(
self,
num_layers,
input_units,
num_units,
kernel_size=3,
dropout_rate=0.3,
position_encoder=None,
activation="tanh",
auxiliary_states=True,
out_units=None,
out_norm=False,
out_residual=False,
include_batchnorm=False,
regularization_weight=0.0,
stride=1,
tf2torch_tensor_name_prefix_torch: str = "speaker_encoder",
tf2torch_tensor_name_prefix_tf: str = "EAND/speaker_encoder",
):
super().__init__()
self._output_size = num_units
self.num_layers = num_layers
self.input_units = input_units
self.num_units = num_units
self.kernel_size = kernel_size
self.dropout_rate = dropout_rate
self.position_encoder = position_encoder
self.out_units = out_units
self.auxiliary_states = auxiliary_states
self.out_norm = out_norm
self.activation = activation
self.out_residual = out_residual
self.include_batch_norm = include_batchnorm
self.regularization_weight = regularization_weight
self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
if isinstance(stride, int):
self.stride = [stride] * self.num_layers
else:
self.stride = stride
self.downsample_rate = 1
for s in self.stride:
self.downsample_rate *= s
self.dropout = nn.Dropout(dropout_rate)
self.cnn_a = repeat(
self.num_layers,
lambda lnum: EncoderLayer(
input_units if lnum == 0 else num_units,
num_units,
kernel_size,
activation,
self.stride[lnum],
include_batchnorm,
residual=True if lnum > 0 else False,
),
)
if self.out_units is not None:
left_padding = math.ceil((kernel_size - stride) / 2)
right_padding = kernel_size - stride - left_padding
self.out_padding = nn.ConstantPad1d((left_padding, right_padding), 0.0)
self.conv_out = nn.Conv1d(
num_units,
out_units,
kernel_size,
)
if self.out_norm:
self.after_norm = LayerNorm(out_units)
def output_size(self) -> int:
return self.num_units
def forward(
self,
xs_pad: torch.Tensor,
ilens: torch.Tensor,
prev_states: torch.Tensor = None,
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
inputs = xs_pad
if self.position_encoder is not None:
inputs = self.position_encoder(inputs)
if self.dropout_rate > 0:
inputs = self.dropout(inputs)
outputs, _ = self.cnn_a(inputs.transpose(1, 2), ilens)
if self.out_units is not None:
outputs = self.conv_out(self.out_padding(outputs))
outputs = outputs.transpose(1, 2)
if self.out_norm:
outputs = self.after_norm(outputs)
if self.out_residual:
outputs = outputs + inputs
return outputs, ilens, None
|