File size: 6,545 Bytes
eef5961 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 |
import torch
import torch.nn as nn
from transformers import PreTrainedModel
from .configuration_reborn import RebornUASRConfig
from typing import Optional, Tuple, Union
class RebornSegmenter(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.conv1 = nn.Conv1d(config.segmenter_input_dim, config.segmenter_hidden_dim, config.segmenter_kernel_size, padding=config.segmenter_kernel_size//2)
self.conv2 = nn.Conv1d(config.segmenter_hidden_dim, config.segmenter_hidden_dim, 3, padding=1)
self.conv3 = nn.Conv1d(config.segmenter_hidden_dim, 2, 1)
self.dropout = nn.Dropout(config.segmenter_dropout)
self.relu = nn.ReLU()
def forward(self, x):
"""
Input:
x: (B, T, C)
padding_mask: (B, T) # 0: not padding; 1: padding
Output:
boundary: (B, T, 2) # 0: not boundary; 1: boundary
"""
x = x.transpose(1, 2)
x = self.dropout(self.relu(self.conv1(x)))
x = self.dropout(self.relu(self.conv2(x)))
x = self.conv3(x)
x = x.transpose(1, 2)
return x
def boundary_predict(self, x, padding_mask, deterministic=False):
"""
Input:
x: (B, T, C)
padding_mask: (B, T)
Output:
boundary: (B, T) # 0: not boundary; 1: boundary
boundary_logits: (B, T, 2) # 0: not boundary; 1: boundary
"""
boundary_logits = self.forward(x)
if deterministic:
boundary = boundary_logits.argmax(-1)
boundary[padding_mask] = -1
else:
boundary = torch.distributions.Categorical(logits=boundary_logits).sample()
boundary[padding_mask] = -1
return boundary, boundary_logits
def pre_segment(self, logits, padding_mask, return_boundary=False, deterministic=True):
"""
Input:
logits: (B, T, C)
padding_mask: (B, T)
Output:
new_logits: (B, T', C)
new_padding_mask: (B, T')
"""
bsz, tsz, csz = logits.size()
boundary, boundary_logits = self.boundary_predict(logits, padding_mask, deterministic=deterministic)
# max boundary number
# print("boundary", boundary)
# print(torch.sum(boundary==1, dim=1))
new_tsz = int(torch.max(torch.sum(boundary==1, dim=1)).item())+1 # add <bos>
new_logits = logits.new_zeros(bsz, new_tsz, csz)
new_pad = padding_mask.new_zeros(bsz, new_tsz)
for b in range(bsz):
# merge consecutive segments when meeting a boundary (mean_pool_join)
new_idx = 0
count = 0
for t in range(tsz):
if padding_mask[b, t] == 1:
break
if boundary[b, t] == 1:
new_logits[b, new_idx] /= count
new_idx += 1
count = 0
new_logits[b, new_idx] += logits[b, t]
count += 1
if count > 0:
# last segment
new_logits[b, new_idx] /= count
new_idx += 1
count = 0
if new_idx < new_tsz:
pad = new_tsz - new_idx
new_logits[b, -pad:] = 0
new_pad[b, -pad:] = True
if return_boundary:
return new_logits, new_pad, boundary, boundary_logits
return new_logits, new_pad
class RebornGenerator(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.output_dim = config.generator_output_dim
self.stride = config.generator_stride
self.dropout = nn.Dropout(config.generator_dropout)
cnn_input_dim = config.generator_input_dim
cnn_output_dim = config.generator_output_dim
padding = config.generator_kernel // 2
self.proj = nn.Sequential(
nn.Conv1d(
cnn_input_dim,
cnn_output_dim,
kernel_size=config.generator_kernel,
stride=config.generator_stride,
dilation=config.generator_dilation,
padding=padding,
bias=config.generator_bias,
),
)
def forward(self, dense_x, tokens, dense_padding_mask):
dense_x = self.dropout(dense_x)
# (B, T, C) -> (B, C, T)
dense_x = dense_x.transpose(-2, -1)
dense_x = self.proj(dense_x)
# (B, C, T) -> (B, T, C)
dense_x = dense_x.transpose(-2, -1)
if self.stride > 1:
dense_padding_mask = dense_padding_mask[:, :: self.stride]
if dense_padding_mask.size(1) != dense_x.size(1):
new_padding = dense_padding_mask.new_zeros(dense_x.shape[:-1])
diff = new_padding.size(1) - dense_padding_mask.size(1)
assert (
diff > 0
), f"{new_padding.shape}, {dense_padding_mask.shape}, {dense_x.shape}, {diff}"
if diff > 0:
new_padding[:, diff:] = dense_padding_mask
else:
assert diff < 0
new_padding = dense_padding_mask[:, :diff]
dense_padding_mask = new_padding
result = {}
token_x = None
if tokens is not None:
token_x = dense_x.new_zeros(tokens.numel(), self.output_dim)
token_x.scatter_(1, tokens.view(-1, 1).long(), 1)
token_x = token_x.view(tokens.shape + (self.output_dim,))
result["dense_x"] = dense_x
result["token_x"] = token_x
result["dense_padding_mask"] = dense_padding_mask
return result
class RebornUASRModel(PreTrainedModel):
config_class = RebornUASRConfig
def __init__(self, config):
super().__init__(config)
self.pca = nn.Linear(1024, 512)
self.segmenter = RebornSegmenter(config)
self.generator = RebornGenerator(config)
def forward(
self,
x: Optional[torch.Tensor], # (B, T, C)
padding_mask: Optional[torch.Tensor], # (B, T)
):
x_reduced = self.pca(x)
x_segmented, segmented_padding_mask = self.segmenter.pre_segment(x_reduced, padding_mask, deterministic=True)
x_generated = self.generator(x_segmented, None, segmented_padding_mask)
return {
'x_reduced': x_reduced,
'x_segmented': x_segmented,
'x_generated': x_generated
}
|