File size: 12,995 Bytes
a7e4030 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 |
import torch
import torch.nn as nn
from transformers import PreTrainedModel
from .configuration_reborn import RebornUASRConfig
from typing import Optional, Tuple, Union, List
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
def get_item(tensor):
# tpu-comment: making this a no-op for xla devices.
if torch.is_tensor(tensor) and tensor.device.type == "xla":
return tensor.detach()
if hasattr(tensor, "item"):
return tensor.item()
if hasattr(tensor, "__getitem__"):
return tensor[0]
return tensor
def post_process(sentence: str, symbol: str):
if symbol == "sentencepiece":
sentence = sentence.replace(" ", "").replace("\u2581", " ").strip()
elif symbol == "wordpiece":
sentence = sentence.replace(" ", "").replace("_", " ").strip()
elif symbol == "letter":
sentence = sentence.replace(" ", "").replace("|", " ").strip()
elif symbol == "silence":
import re
sentence = sentence.replace("<SIL>", "")
sentence = re.sub(' +', ' ', sentence).strip()
elif symbol == "_EOW":
sentence = sentence.replace(" ", "").replace("_EOW", " ").strip()
elif symbol in {"subword_nmt", "@@ ", "@@"}:
if symbol == "subword_nmt":
symbol = "@@ "
sentence = (sentence + " ").replace(symbol, "").rstrip()
elif symbol == "none":
pass
elif symbol is not None:
raise NotImplementedError(f"Unknown post_process option: {symbol}")
return sentence
class SimpleTokenizer(object):
def __init__(self,
phones: List[str],
bos="<s>",
pad="<pad>",
eos="</s>",
unk="<unk>",
extra_special_symbols=None,
) -> None:
self.bos_word, self.unk_word, self.pad_word, self.eos_word = bos, unk, pad, eos
self.symbols = []
self.count = []
self.indices = {}
self.bos_index = self.add_symbol(bos)
self.pad_index = self.add_symbol(pad)
self.eos_index = self.add_symbol(eos)
self.unk_index = self.add_symbol(unk)
if extra_special_symbols:
for s in extra_special_symbols:
self.add_symbol(s)
self.nspecial = len(self.symbols)
for phone in phones:
self.add_symbol(phone)
self.postprocess_code = "silence"
def add_symbol(self, word, n=1, overwrite=False):
"""Adds a word to the dictionary"""
if word in self.indices and not overwrite:
idx = self.indices[word]
self.count[idx] = self.count[idx] + n
return idx
else:
idx = len(self.symbols)
self.indices[word] = idx
self.symbols.append(word)
self.count.append(n)
return idx
def __eq__(self, other):
return self.indices == other.indices
def __getitem__(self, idx):
if idx < len(self.symbols):
return self.symbols[idx]
return self.unk_word
def get_count(self, idx):
return self.count[idx]
def __len__(self):
"""Returns the number of symbols in the dictionary"""
return len(self.symbols)
def __contains__(self, sym):
return sym in self.indices
def index(self, sym):
"""Returns the index of the specified symbol"""
assert isinstance(sym, str)
if sym in self.indices:
return self.indices[sym]
return self.unk_index
def string(
self,
tensor,
bpe_symbol=None,
escape_unk=False,
extra_symbols_to_ignore=None,
unk_string=None,
include_eos=False,
separator=" ",
):
"""Helper for converting a tensor of token indices to a string.
Can optionally remove BPE symbols or escape <unk> words.
"""
if torch.is_tensor(tensor) and tensor.dim() == 2:
return "\n".join(
self.string(
t,
bpe_symbol,
escape_unk,
extra_symbols_to_ignore,
include_eos=include_eos,
)
for t in tensor
)
extra_symbols_to_ignore = set(extra_symbols_to_ignore or [])
if not include_eos:
extra_symbols_to_ignore.add(self.eos())
def token_string(i):
if i == self.unk():
if unk_string is not None:
return unk_string
else:
return self.unk_string(escape_unk)
else:
return self[i]
if hasattr(self, "bos_index"):
extra_symbols_to_ignore.add(self.bos())
sent = separator.join(
token_string(i)
for i in tensor
if get_item(i) not in extra_symbols_to_ignore
)
return post_process(sent, bpe_symbol)
def unk_string(self, escape=False):
"""Return unknown string, optionally escaped as: <<unk>>"""
if escape:
return "<{}>".format(self.unk_word)
else:
return self.unk_word
def bos(self):
"""Helper to get index of beginning-of-sentence symbol"""
return self.bos_index
def pad(self):
"""Helper to get index of pad symbol"""
return self.pad_index
def eos(self):
"""Helper to get index of end-of-sentence symbol"""
return self.eos_index
def unk(self):
"""Helper to get index of unk symbol"""
return self.unk_index
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)
self.tokenizer = None
if len(config.phones) > 0:
self.tokenizer = SimpleTokenizer(config.phones)
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
}
def generate(self, x, padding_mask, merge_consecutive=True, remove_silence=True):
res = self.forward(x, padding_mask)
y_raw_logits = res['x_generated']['dense_x']
y_raw_padding = res['x_generated']['dense_padding_mask']
y_raw_logits[y_raw_padding][..., self.tokenizer.pad_index] = float('inf')
preds = y_raw_logits.argmax(-1)
hyps = []
postprocess_code = "silence" if remove_silence else "none"
for pred in preds:
if merge_consecutive:
# merge consecutive predictions
pred = torch.unique_consecutive(pred)
hyp = self.tokenizer.string(pred, bpe_symbol=postprocess_code)
hyps.append(hyp)
return hyps
def main():
model_config = RebornUASRConfig.from_pretrained("/home/andybi7676/Desktop/uasr-rl/reborn_uasr/config.json")
print(model_config)
model = RebornUASRModel(model_config)
print(model.tokenizer.indices)
if __name__ == "__main__":
main()
|