File size: 7,454 Bytes
0b32ad6 |
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 |
"""
Speech2Text with CTC loss
Authors
* Heng-Jui Chang 2022
"""
import logging
from pathlib import Path
from typing import List, Union
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from s3prl.dataio.encoder.tokenizer import Tokenizer
from s3prl.metric import cer, per, wer
from s3prl.metric.slot_filling import (
slot_edit_f1_full,
slot_edit_f1_part,
slot_type_f1,
slot_value_cer,
slot_value_wer,
)
from s3prl.nn import BeamDecoder
from . import Task
logger = logging.getLogger(__name__)
__all__ = [
"Speech2TextCTCExample",
"Speech2TextCTCTask",
]
class Speech2TextCTCExample(nn.Module):
"""An example speech-to-text task with CTC objective
Args:
input_size (int, optional): Input size. Defaults to 3.
output_size (int, optional): Output size. Defaults to 4.
"""
def __init__(self, input_size=3, output_size=4):
super().__init__()
self._input_size = input_size
self._output_size = output_size
@property
def input_size(self):
return self._input_size
@property
def output_size(self):
return self._output_size
def forward(self, x, x_len):
"""
Args:
x (torch.Tensor): (batch_size, timestemps, input_size)
x_len (torch.LongTensor): (batch_size, )
Return:
y (torch.Tensor): (batch_size, output_size)
y_len (torch.LongTensor): (batch_size)
"""
assert x.size(-1) == self.input_size
output = torch.randn(x.size(0), x.size(1), self.output_size)
assert output, x_len
class Speech2TextCTCTask(Task):
"""Speech-to-text task with CTC objective
Args:
model (Speech2TextCTCExample)
tokenizer (Tokenizer): Text tokenizer.
decoder (Union[BeamDecoder, dict], optional):
Beam decoder or decoder's config. Defaults to None.
log_metrics (List[str], optional):
Metrics to be logged. Defaults to ["cer", "wer"].
"""
def __init__(
self,
model: torch.nn.Module,
tokenizer: Tokenizer,
decoder: Union[BeamDecoder, dict] = None,
log_metrics: List[str] = ["cer", "wer"],
) -> None:
super().__init__()
self.model = model
assert isinstance(tokenizer, Tokenizer)
self.tokenizer = tokenizer
self.log_metrics = log_metrics
if BeamDecoder is None:
decoder = None
if isinstance(decoder, dict):
decoder = BeamDecoder(**decoder)
logger.info("Using flashlight decoder.")
self.decoder = decoder
self.criterion = nn.CTCLoss(
blank=self.tokenizer.pad_idx,
zero_infinity=True,
)
def predict(self, x: torch.Tensor, x_len: torch.LongTensor):
"""
Args:
x (torch.Tensor): (batch_size, timestamps, input_size)
x_len (torch.LongTensor): (batch_size, )
Return:
logits (torch.Tensor): (batch_size, timestamps, output_size)
prediction (list): prediction strings
valid_length (torch.LongTensor): (batch_size, )
"""
logits, x_len = self.model(x, x_len)
predicted_tokens = torch.argmax(logits, dim=2).detach().cpu()
filtered_tokens = [
[
token
for token in pred_token.unique_consecutive().tolist()
if token != self.tokenizer.pad_idx and token != self.tokenizer.eos_idx
]
for pred_token in predicted_tokens
]
predictions = [
self.tokenizer.decode(token_list) for token_list in filtered_tokens
]
return logits, predictions, x_len
def forward(
self,
_mode: str,
x: torch.Tensor,
x_len: torch.LongTensor,
labels: np.ndarray,
class_ids: torch.LongTensor,
unique_name: np.ndarray,
beam_decode: bool = False,
_dump_dir: str = None,
):
"""
Each forward step in the training loop
Args:
mode (str): train / valid / test
x (torch.Tensor):
Input waveform or acoustic features.
(batch_size, timestamps, input_size)
x_len (torch.LongTensor):
Lengths of inputs.
(batch_size, )
labels (np.ndarray):
Ground truth transcriptions (str).
(batch_size, )
class_ids (torch.LongTensor):
Tokenized ground truth transcriptions.
unique_name (np.ndarray):
Unique names for each sample.
"""
logits, prediction, x_len = self.predict(x, x_len)
log_probs = F.log_softmax(logits, dim=2)
y = class_ids
y_len = torch.tensor(
[(ids != self.tokenizer.pad_idx).long().sum() for ids in class_ids],
dtype=torch.long,
device=logits.device,
)
loss = self.criterion(log_probs.transpose(0, 1), y, x_len, y_len)
hyps = None
if beam_decode and self.decoder is not None:
hyps = self.decoder.decode(log_probs.detach())
cacheable = dict(
loss=loss.detach().cpu().item(),
prediction=prediction,
label=labels.tolist(),
unique_name=unique_name.tolist(),
hypotheses=hyps,
)
return loss, cacheable
def reduction(self, _mode: str, cached_results: List[dict], _dump_dir: str = None):
results = self.parse_cached_results(cached_results)
losses = results["loss"]
predictions = results["prediction"]
labels = results["label"]
unique_names = results["unique_name"]
if _dump_dir is not None:
with (Path(_dump_dir) / "hyp").open("w") as f:
f.writelines(
[f"{uid} {p}\n" for p, uid in zip(predictions, unique_names)]
)
with (Path(_dump_dir) / "ref").open("w") as f:
f.writelines([f"{uid} {p}\n" for p, uid in zip(labels, unique_names)])
beam_hyps = None
if results["hypotheses"][0] is not None:
beam_hyps = [" ".join(hyp[0].words) for hyp in results["hypotheses"]]
logs = {}
logs["loss"] = float(np.mean(losses))
if "wer" in self.log_metrics:
logs["wer"] = wer(predictions, labels)
if "cer" in self.log_metrics:
logs["cer"] = cer(predictions, labels)
if "per" in self.log_metrics:
logs["per"] = per(predictions, labels)
if "slot_type_f1" in self.log_metrics:
logs["slot_type_f1"] = slot_type_f1(predictions, labels)
if "slot_value_cer" in self.log_metrics:
logs["slot_value_cer"] = slot_value_cer(predictions, labels)
if "slot_value_wer" in self.log_metrics:
logs["slot_value_wer"] = slot_value_wer(predictions, labels)
if "slot_edit_f1_full" in self.log_metrics:
logs["slot_edit_f1_full"] = slot_edit_f1_full(predictions, labels)
if "slot_edit_f1_part" in self.log_metrics:
logs["slot_edit_f1_part"] = slot_edit_f1_part(predictions, labels)
if beam_hyps is not None:
logs["wer_beam"] = wer(beam_hyps, labels)
logs["char_beam"] = cer(beam_hyps, labels)
return logs
|