added mixer model
Browse files- results/non_semi_final_stac/ctc_lin.py +756 -0
- results/non_semi_final_stac/hyperparams.yaml +144 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/CKPT.yaml +4 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/brain.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/counter.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/dataloader-TRAIN.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/model.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/modelopt.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/scheduler_encoder.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/scheduler_model.ckpt +3 -0
- results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/tokenizer.ckpt +3 -0
- results/non_semi_final_stac/save/label_encoder.txt +80 -0
results/non_semi_final_stac/ctc_lin.py
ADDED
@@ -0,0 +1,756 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
|
3 |
+
import os
|
4 |
+
import sys
|
5 |
+
import torch
|
6 |
+
import logging
|
7 |
+
import speechbrain as sb
|
8 |
+
from speechbrain.utils.distributed import run_on_main
|
9 |
+
from hyperpyyaml import load_hyperpyyaml
|
10 |
+
from pathlib import Path
|
11 |
+
import torchaudio.transforms as T
|
12 |
+
from cv_train import ASRCV
|
13 |
+
import torchaudio
|
14 |
+
import numpy as np
|
15 |
+
import kenlm
|
16 |
+
from pyctcdecode import build_ctcdecoder
|
17 |
+
import re
|
18 |
+
|
19 |
+
# Commented out IPython magic to ensure Python compatibility.
|
20 |
+
# %cd /content/drive/MyDrive/tunisian_corpora/tunisian_without_wavlm
|
21 |
+
#hparams_file, run_opts, overrides = sb.parse_arguments(["/gpfsstore/rech/nou/uzn19yk/switched_code_tunisian/train/tunisian_asr/hparams/train_semi.yaml"])
|
22 |
+
hparams_file, run_opts, overrides = sb.parse_arguments(["semi_supervised_test_tunisian.yaml"])
|
23 |
+
|
24 |
+
# If distributed_launch=True then
|
25 |
+
# create ddp_group with the right communication protocol
|
26 |
+
sb.utils.distributed.ddp_init_group(run_opts)
|
27 |
+
|
28 |
+
with open(hparams_file) as fin:
|
29 |
+
hparams = load_hyperpyyaml(fin, overrides)
|
30 |
+
|
31 |
+
# Create experiment directory
|
32 |
+
sb.create_experiment_directory(
|
33 |
+
experiment_directory=hparams["output_folder"],
|
34 |
+
hyperparams_to_save=hparams_file,
|
35 |
+
overrides=overrides,
|
36 |
+
)
|
37 |
+
# Dataset prep (parsing Librispeech)
|
38 |
+
|
39 |
+
def dataio_prepare(hparams):
|
40 |
+
"""This function prepares the datasets to be used in the brain class.
|
41 |
+
It also defines the data processing pipeline through user-defined functions."""
|
42 |
+
|
43 |
+
# 1. Define datasets
|
44 |
+
data_folder = hparams["data_folder"]
|
45 |
+
|
46 |
+
train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
|
47 |
+
csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
|
48 |
+
)
|
49 |
+
|
50 |
+
if hparams["sorting"] == "ascending":
|
51 |
+
# we sort training data to speed up training and get better results.
|
52 |
+
train_data = train_data.filtered_sorted(
|
53 |
+
sort_key="duration",
|
54 |
+
key_max_value={"duration": hparams["avoid_if_longer_than"]},
|
55 |
+
)
|
56 |
+
# when sorting do not shuffle in dataloader ! otherwise is pointless
|
57 |
+
hparams["dataloader_options"]["shuffle"] = False
|
58 |
+
|
59 |
+
elif hparams["sorting"] == "descending":
|
60 |
+
train_data = train_data.filtered_sorted(
|
61 |
+
sort_key="duration",
|
62 |
+
reverse=True,
|
63 |
+
key_max_value={"duration": hparams["avoid_if_longer_than"]},
|
64 |
+
)
|
65 |
+
# when sorting do not shuffle in dataloader ! otherwise is pointless
|
66 |
+
hparams["dataloader_options"]["shuffle"] = False
|
67 |
+
|
68 |
+
elif hparams["sorting"] == "random":
|
69 |
+
pass
|
70 |
+
|
71 |
+
else:
|
72 |
+
raise NotImplementedError(
|
73 |
+
"sorting must be random, ascending or descending"
|
74 |
+
)
|
75 |
+
|
76 |
+
valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
|
77 |
+
csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
|
78 |
+
)
|
79 |
+
# We also sort the validation data so it is faster to validate
|
80 |
+
valid_data = valid_data.filtered_sorted(sort_key="duration")
|
81 |
+
test_datasets = {}
|
82 |
+
for csv_file in hparams["test_csv"]:
|
83 |
+
name = Path(csv_file).stem
|
84 |
+
test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
|
85 |
+
csv_path=csv_file, replacements={"data_root": data_folder}
|
86 |
+
)
|
87 |
+
test_datasets[name] = test_datasets[name].filtered_sorted(
|
88 |
+
sort_key="duration"
|
89 |
+
)
|
90 |
+
|
91 |
+
datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
|
92 |
+
|
93 |
+
|
94 |
+
# 2. Define audio pipeline:
|
95 |
+
@sb.utils.data_pipeline.takes("wav")
|
96 |
+
@sb.utils.data_pipeline.provides("sig")
|
97 |
+
def audio_pipeline(wav):
|
98 |
+
info = torchaudio.info(wav)
|
99 |
+
sig = sb.dataio.dataio.read_audio(wav)
|
100 |
+
if len(sig.shape)>1 :
|
101 |
+
sig = torch.mean(sig, dim=1)
|
102 |
+
resampled = torchaudio.transforms.Resample(
|
103 |
+
info.sample_rate, hparams["sample_rate"],
|
104 |
+
)(sig)
|
105 |
+
return resampled
|
106 |
+
|
107 |
+
sb.dataio.dataset.add_dynamic_item(datasets, audio_pipeline)
|
108 |
+
label_encoder = sb.dataio.encoder.CTCTextEncoder()
|
109 |
+
|
110 |
+
# 3. Define text pipeline:
|
111 |
+
@sb.utils.data_pipeline.takes("wrd")
|
112 |
+
@sb.utils.data_pipeline.provides(
|
113 |
+
"wrd", "char_list", "tokens_list", "tokens"
|
114 |
+
)
|
115 |
+
def text_pipeline(wrd):
|
116 |
+
yield wrd
|
117 |
+
char_list = list(wrd)
|
118 |
+
yield char_list
|
119 |
+
tokens_list = label_encoder.encode_sequence(char_list)
|
120 |
+
yield tokens_list
|
121 |
+
tokens = torch.LongTensor(tokens_list)
|
122 |
+
yield tokens
|
123 |
+
|
124 |
+
sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)
|
125 |
+
lab_enc_file = os.path.join(hparams["save_folder"], "label_encoder.txt")
|
126 |
+
special_labels = {
|
127 |
+
"blank_label": hparams["blank_index"],
|
128 |
+
"unk_label": hparams["unk_index"]
|
129 |
+
}
|
130 |
+
label_encoder.load_or_create(
|
131 |
+
path=lab_enc_file,
|
132 |
+
from_didatasets=[train_data],
|
133 |
+
output_key="char_list",
|
134 |
+
special_labels=special_labels,
|
135 |
+
sequence_input=True,
|
136 |
+
)
|
137 |
+
|
138 |
+
# 4. Set output:
|
139 |
+
sb.dataio.dataset.set_output_keys(
|
140 |
+
datasets, ["id", "sig", "wrd", "char_list", "tokens"],
|
141 |
+
)
|
142 |
+
return train_data, valid_data,test_datasets, label_encoder
|
143 |
+
|
144 |
+
class ASR(sb.core.Brain):
|
145 |
+
def compute_forward(self, batch, stage):
|
146 |
+
"""Forward computations from the waveform batches to the output probabilities."""
|
147 |
+
|
148 |
+
batch = batch.to(self.device)
|
149 |
+
wavs, wav_lens = batch.sig
|
150 |
+
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
|
151 |
+
|
152 |
+
if stage == sb.Stage.TRAIN:
|
153 |
+
if hasattr(self.hparams, "augmentation"):
|
154 |
+
wavs = self.hparams.augmentation(wavs, wav_lens)
|
155 |
+
|
156 |
+
# Forward pass
|
157 |
+
feats = self.modules.wav2vec2(wavs, wav_lens)
|
158 |
+
x = self.modules.enc(feats)
|
159 |
+
logits = self.modules.ctc_lin(x)
|
160 |
+
p_ctc = self.hparams.log_softmax(logits)
|
161 |
+
|
162 |
+
return p_ctc, wav_lens
|
163 |
+
|
164 |
+
def custom_encode(self,wavs,wav_lens) :
|
165 |
+
wavs = wavs.to(self.device)
|
166 |
+
if(wav_lens is not None): wav_lens.to(self.device)
|
167 |
+
|
168 |
+
feats = self.modules.wav2vec2(wavs, wav_lens)
|
169 |
+
x = self.modules.enc(feats)
|
170 |
+
logits = self.modules.ctc_lin(x)
|
171 |
+
p_ctc = self.hparams.log_softmax(logits)
|
172 |
+
|
173 |
+
return feats,p_ctc
|
174 |
+
|
175 |
+
|
176 |
+
|
177 |
+
def compute_objectives(self, predictions, batch, stage):
|
178 |
+
"""Computes the loss (CTC) given predictions and targets."""
|
179 |
+
|
180 |
+
p_ctc, wav_lens = predictions
|
181 |
+
|
182 |
+
ids = batch.id
|
183 |
+
tokens, tokens_lens = batch.tokens
|
184 |
+
|
185 |
+
loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
|
186 |
+
|
187 |
+
if stage != sb.Stage.TRAIN:
|
188 |
+
predicted_tokens = sb.decoders.ctc_greedy_decode(
|
189 |
+
p_ctc, wav_lens, blank_id=self.hparams.blank_index
|
190 |
+
)
|
191 |
+
# Decode token terms to words
|
192 |
+
if self.hparams.use_language_modelling:
|
193 |
+
predicted_words = []
|
194 |
+
for logs in p_ctc:
|
195 |
+
text = decoder.decode(logs.detach().cpu().numpy())
|
196 |
+
predicted_words.append(text.split(" "))
|
197 |
+
else:
|
198 |
+
predicted_words = [
|
199 |
+
"".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
|
200 |
+
for utt_seq in predicted_tokens
|
201 |
+
]
|
202 |
+
# Convert indices to words
|
203 |
+
target_words = [wrd.split(" ") for wrd in batch.wrd]
|
204 |
+
|
205 |
+
self.wer_metric.append(ids, predicted_words, target_words)
|
206 |
+
self.cer_metric.append(ids, predicted_words, target_words)
|
207 |
+
|
208 |
+
return loss
|
209 |
+
|
210 |
+
def fit_batch(self, batch):
|
211 |
+
"""Train the parameters given a single batch in input"""
|
212 |
+
should_step = self.step % self.grad_accumulation_factor == 0
|
213 |
+
# Managing automatic mixed precision
|
214 |
+
# TOFIX: CTC fine-tuning currently is unstable
|
215 |
+
# This is certainly due to CTC being done in fp16 instead of fp32
|
216 |
+
if self.auto_mix_prec:
|
217 |
+
with torch.cuda.amp.autocast():
|
218 |
+
with self.no_sync():
|
219 |
+
outputs = self.compute_forward(batch, sb.Stage.TRAIN)
|
220 |
+
loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
|
221 |
+
with self.no_sync(not should_step):
|
222 |
+
self.scaler.scale(
|
223 |
+
loss / self.grad_accumulation_factor
|
224 |
+
).backward()
|
225 |
+
if should_step:
|
226 |
+
|
227 |
+
if not self.hparams.wav2vec2.freeze:
|
228 |
+
self.scaler.unscale_(self.wav2vec_optimizer)
|
229 |
+
self.scaler.unscale_(self.model_optimizer)
|
230 |
+
if self.check_gradients(loss):
|
231 |
+
if not self.hparams.wav2vec2.freeze:
|
232 |
+
if self.optimizer_step >= self.hparams.warmup_steps:
|
233 |
+
self.scaler.step(self.wav2vec_optimizer)
|
234 |
+
self.scaler.step(self.model_optimizer)
|
235 |
+
self.scaler.update()
|
236 |
+
self.zero_grad()
|
237 |
+
self.optimizer_step += 1
|
238 |
+
else:
|
239 |
+
# This is mandatory because HF models have a weird behavior with DDP
|
240 |
+
# on the forward pass
|
241 |
+
with self.no_sync():
|
242 |
+
outputs = self.compute_forward(batch, sb.Stage.TRAIN)
|
243 |
+
|
244 |
+
loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
|
245 |
+
|
246 |
+
with self.no_sync(not should_step):
|
247 |
+
(loss / self.grad_accumulation_factor).backward()
|
248 |
+
if should_step:
|
249 |
+
if self.check_gradients(loss):
|
250 |
+
if not self.hparams.wav2vec2.freeze:
|
251 |
+
if self.optimizer_step >= self.hparams.warmup_steps:
|
252 |
+
self.wav2vec_optimizer.step()
|
253 |
+
self.model_optimizer.step()
|
254 |
+
self.zero_grad()
|
255 |
+
self.optimizer_step += 1
|
256 |
+
|
257 |
+
self.on_fit_batch_end(batch, outputs, loss, should_step)
|
258 |
+
return loss.detach().cpu()
|
259 |
+
|
260 |
+
def evaluate_batch(self, batch, stage):
|
261 |
+
"""Computations needed for validation/test batches"""
|
262 |
+
predictions = self.compute_forward(batch, stage=stage)
|
263 |
+
with torch.no_grad():
|
264 |
+
loss = self.compute_objectives(predictions, batch, stage=stage)
|
265 |
+
return loss.detach()
|
266 |
+
|
267 |
+
def on_stage_start(self, stage, epoch):
|
268 |
+
"""Gets called at the beginning of each epoch"""
|
269 |
+
if stage != sb.Stage.TRAIN:
|
270 |
+
self.cer_metric = self.hparams.cer_computer()
|
271 |
+
self.wer_metric = self.hparams.error_rate_computer()
|
272 |
+
|
273 |
+
def on_stage_end(self, stage, stage_loss, epoch):
|
274 |
+
"""Gets called at the end of an epoch."""
|
275 |
+
# Compute/store important stats
|
276 |
+
stage_stats = {"loss": stage_loss}
|
277 |
+
if stage == sb.Stage.TRAIN:
|
278 |
+
self.train_stats = stage_stats
|
279 |
+
else:
|
280 |
+
stage_stats["CER"] = self.cer_metric.summarize("error_rate")
|
281 |
+
stage_stats["WER"] = self.wer_metric.summarize("error_rate")
|
282 |
+
|
283 |
+
# Perform end-of-iteration things, like annealing, logging, etc.
|
284 |
+
if stage == sb.Stage.VALID:
|
285 |
+
old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
|
286 |
+
stage_stats["loss"]
|
287 |
+
)
|
288 |
+
old_lr_wav2vec, new_lr_wav2vec = self.hparams.lr_annealing_wav2vec(
|
289 |
+
stage_stats["loss"]
|
290 |
+
)
|
291 |
+
sb.nnet.schedulers.update_learning_rate(
|
292 |
+
self.model_optimizer, new_lr_model
|
293 |
+
)
|
294 |
+
if not self.hparams.wav2vec2.freeze:
|
295 |
+
sb.nnet.schedulers.update_learning_rate(
|
296 |
+
self.wav2vec_optimizer, new_lr_wav2vec
|
297 |
+
)
|
298 |
+
self.hparams.train_logger.log_stats(
|
299 |
+
stats_meta={
|
300 |
+
"epoch": epoch,
|
301 |
+
"lr_model": old_lr_model,
|
302 |
+
"lr_wav2vec": old_lr_wav2vec,
|
303 |
+
},
|
304 |
+
train_stats=self.train_stats,
|
305 |
+
valid_stats=stage_stats,
|
306 |
+
)
|
307 |
+
self.checkpointer.save_and_keep_only(
|
308 |
+
meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
|
309 |
+
)
|
310 |
+
elif stage == sb.Stage.TEST:
|
311 |
+
self.hparams.train_logger.log_stats(
|
312 |
+
stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
|
313 |
+
test_stats=stage_stats,
|
314 |
+
)
|
315 |
+
with open(self.hparams.wer_file, "w") as w:
|
316 |
+
self.wer_metric.write_stats(w)
|
317 |
+
|
318 |
+
def init_optimizers(self):
|
319 |
+
"Initializes the wav2vec2 optimizer and model optimizer"
|
320 |
+
|
321 |
+
# If the wav2vec encoder is unfrozen, we create the optimizer
|
322 |
+
if not self.hparams.wav2vec2.freeze:
|
323 |
+
self.wav2vec_optimizer = self.hparams.wav2vec_opt_class(
|
324 |
+
self.modules.wav2vec2.parameters()
|
325 |
+
)
|
326 |
+
if self.checkpointer is not None:
|
327 |
+
self.checkpointer.add_recoverable(
|
328 |
+
"wav2vec_opt", self.wav2vec_optimizer
|
329 |
+
)
|
330 |
+
|
331 |
+
self.model_optimizer = self.hparams.model_opt_class(
|
332 |
+
self.hparams.model.parameters()
|
333 |
+
)
|
334 |
+
|
335 |
+
if self.checkpointer is not None:
|
336 |
+
self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
|
337 |
+
|
338 |
+
def zero_grad(self, set_to_none=False):
|
339 |
+
if not self.hparams.wav2vec2.freeze:
|
340 |
+
self.wav2vec_optimizer.zero_grad(set_to_none)
|
341 |
+
self.model_optimizer.zero_grad(set_to_none)
|
342 |
+
|
343 |
+
|
344 |
+
"""
|
345 |
+
label_encoder = sb.dataio.encoder.CTCTextEncoder()
|
346 |
+
|
347 |
+
train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
|
348 |
+
hparams
|
349 |
+
)
|
350 |
+
|
351 |
+
|
352 |
+
# We dynamicaly add the tokenizer to our brain class.
|
353 |
+
# NB: This tokenizer corresponds to the one used for the LM!!
|
354 |
+
"""
|
355 |
+
from speechbrain.pretrained import EncoderASR,EncoderDecoderASR
|
356 |
+
french_asr_model = EncoderASR.from_hparams(source="speechbrain/asr-wav2vec2-commonvoice-fr", savedir="pretrained_models/asr-wav2vec2-commonvoice-fr").cuda()
|
357 |
+
#french_asr_model = "r"
|
358 |
+
|
359 |
+
cvhparams_file, cvrun_opts, cvoverrides = sb.parse_arguments(["en_cv.yaml"])
|
360 |
+
with open(cvhparams_file) as cvfin:
|
361 |
+
cvhparams = load_hyperpyyaml(cvfin, cvoverrides)
|
362 |
+
english_asr_model = ASRCV(
|
363 |
+
modules=cvhparams["modules"],
|
364 |
+
hparams=cvhparams,
|
365 |
+
run_opts=cvrun_opts,
|
366 |
+
checkpointer=cvhparams["checkpointer"],
|
367 |
+
)
|
368 |
+
english_asr_model.checkpointer.recover_if_possible()
|
369 |
+
asr_brain = ASR(
|
370 |
+
modules=hparams["modules"],
|
371 |
+
hparams=hparams,
|
372 |
+
run_opts=run_opts,
|
373 |
+
checkpointer=hparams["checkpointer"],
|
374 |
+
)
|
375 |
+
asr_brain.checkpointer.recover_if_possible()
|
376 |
+
asr_brain.modules.eval()
|
377 |
+
english_asr_model.modules.eval()
|
378 |
+
french_asr_model.mods.eval()
|
379 |
+
"""
|
380 |
+
asr_brain.tokenizer = label_encoder
|
381 |
+
|
382 |
+
# Testing
|
383 |
+
real = True
|
384 |
+
if real :
|
385 |
+
for k in test_datasets.keys(): # keys are test_clean, test_other etc
|
386 |
+
asr_brain.hparams.wer_file = os.path.join(
|
387 |
+
hparams["output_folder"], "wer_{}.txt".format(k)
|
388 |
+
)
|
389 |
+
asr_brain.evaluate(
|
390 |
+
test_datasets[k], test_loader_kwargs=hparams["dataloader_options"]
|
391 |
+
)
|
392 |
+
"""
|
393 |
+
|
394 |
+
"""
|
395 |
+
from torch.nn.utils.rnn import pad_sequence
|
396 |
+
def load_paths(wavs_path):
|
397 |
+
waveforms = []
|
398 |
+
for path in wavs_path :
|
399 |
+
waveform, _ = torchaudio.load(path)
|
400 |
+
waveforms.append(waveform.squeeze(0))
|
401 |
+
# normalize array length to the bigger arrays by pading with 0's
|
402 |
+
padded_arrays = pad_sequence(waveforms, batch_first=True)
|
403 |
+
return torch.tensor(padded_arrays)
|
404 |
+
|
405 |
+
waveform = load_paths(["/content/drive/MyDrive/tunisian_corpora/tunisian_without_wavlm/samples/Salah10.wav","/content/drive/MyDrive/tunisian_corpora/tunisian_without_wavlm/samples/Salah10.wav"])
|
406 |
+
embeddings, posteriogram = asr_brain.custom_encode(waveform,None)
|
407 |
+
print(embeddings.shape)
|
408 |
+
print(posteriogram.shape)
|
409 |
+
"""
|
410 |
+
|
411 |
+
from speechbrain.pretrained import EncoderASR,EncoderDecoderASR
|
412 |
+
import torchaudio
|
413 |
+
import speechbrain as sb
|
414 |
+
import torch
|
415 |
+
from torch.nn.utils.rnn import pad_sequence
|
416 |
+
import torch
|
417 |
+
import speechbrain as sb
|
418 |
+
import numpy as np
|
419 |
+
import torch.optim as optim
|
420 |
+
import torch.nn as nn
|
421 |
+
|
422 |
+
# Commented out IPython magic to ensure Python compatibility.
|
423 |
+
# %ls
|
424 |
+
|
425 |
+
#UTILS FUNCTIOJNS
|
426 |
+
def get_size_dimensions(arr):
|
427 |
+
size_dimensions = []
|
428 |
+
while isinstance(arr, list):
|
429 |
+
size_dimensions.append(len(arr))
|
430 |
+
arr = arr[0]
|
431 |
+
return size_dimensions
|
432 |
+
|
433 |
+
def scale_array(batch,n):
|
434 |
+
scaled_batch = []
|
435 |
+
|
436 |
+
for array in batch:
|
437 |
+
if(n < len(array)): raise ValueError("Cannot scale Array down")
|
438 |
+
|
439 |
+
repeat = round(n/len(array))+1
|
440 |
+
scaled_length_array= []
|
441 |
+
|
442 |
+
for i in array:
|
443 |
+
for j in range(repeat) :
|
444 |
+
if(len(scaled_length_array) == n): break
|
445 |
+
scaled_length_array.append(i)
|
446 |
+
|
447 |
+
scaled_batch.append(scaled_length_array)
|
448 |
+
|
449 |
+
return torch.tensor(scaled_batch)
|
450 |
+
|
451 |
+
|
452 |
+
def load_paths(wavs_path):
|
453 |
+
waveforms = []
|
454 |
+
for path in wavs_path :
|
455 |
+
waveform, _ = torchaudio.load(path)
|
456 |
+
waveforms.append(waveform.squeeze(0))
|
457 |
+
# normalize array length to the bigger arrays by pading with 0's
|
458 |
+
padded_arrays = pad_sequence(waveforms, batch_first=True)
|
459 |
+
return torch.tensor(padded_arrays)
|
460 |
+
|
461 |
+
|
462 |
+
|
463 |
+
def word_to_vec(input_string):
|
464 |
+
mapping= {'a': 1, 'b': 2, 'c': 3, 'd': 4, 'e': 5, 'f': 6, 'g': 7, 'h': 8, 'i': 9, 'j': 10, 'k': 11, 'l': 12, 'm': 13, 'n': 14, 'o': 15, 'p': 16, 'q': 17, 'r': 18, 's': 19, 't': 20, 'u': 21, 'v': 22, 'w': 23, 'x': 24, 'y': 25, 'z': 26, 'ا': 27, 'ب': 28, 'ت': 29, 'ث': 30, 'ج': 31, 'ح': 32, 'خ': 33, 'د': 34, 'ذ': 35, 'ر': 36, 'ز': 37, 'س': 38, 'ش': 39, 'ص': 40, 'ض': 41, 'ط': 42, 'ظ': 43, 'ع': 44, 'غ': 45, 'ف': 46, 'ق': 47, 'ك': 48, 'ل': 49, 'م': 50, 'ن': 51, 'ه': 52, 'و': 53, 'ي': 54,' ':55}
|
465 |
+
|
466 |
+
numbers = [mapping[word] for word in input_string if word in mapping]
|
467 |
+
return numbers
|
468 |
+
|
469 |
+
device = 'cuda'
|
470 |
+
verbose = 0
|
471 |
+
#FLOW LEVEL FUNCTIONS
|
472 |
+
def merge_strategy(embeddings1, embeddings2, embeddings3,post1, post2,post3):
|
473 |
+
|
474 |
+
|
475 |
+
post1 = post1.to(device)
|
476 |
+
post2 = post2.to(device)
|
477 |
+
post3 = post3.to(device)
|
478 |
+
embeddings1 = embeddings1.to(device)
|
479 |
+
embeddings2 = embeddings2.to(device)
|
480 |
+
embeddings3 = embeddings3.to(device)
|
481 |
+
|
482 |
+
posteriograms_merged = torch.cat((post1,post2,post3),dim=2)
|
483 |
+
embeddings_merged = torch.cat((embeddings1,embeddings2,embeddings3),dim=2)
|
484 |
+
|
485 |
+
if(verbose !=0):
|
486 |
+
print('MERGED POST ',posteriograms_merged.shape)
|
487 |
+
print('MERGED emb ',embeddings_merged.shape)
|
488 |
+
|
489 |
+
return torch.cat((posteriograms_merged,embeddings_merged),dim=2).to(device)
|
490 |
+
|
491 |
+
def decode(model,wavs,wav_lens):
|
492 |
+
|
493 |
+
with torch.no_grad():
|
494 |
+
wav_lens = wav_lens.to(model.device)
|
495 |
+
encoder_out = model.encode_batch(wavs, wav_lens)
|
496 |
+
predictions = model.decoding_function(encoder_out, wav_lens)
|
497 |
+
return predictions
|
498 |
+
|
499 |
+
def middle_layer(batch, lens):
|
500 |
+
|
501 |
+
tn_embeddings, tn_posteriogram = asr_brain.custom_encode(batch,None)
|
502 |
+
|
503 |
+
fr_embeddings = french_asr_model.mods.encoder.wav2vec2(batch)
|
504 |
+
fr_posteriogram =french_asr_model.encode_batch(batch,lens)
|
505 |
+
en_embeddings = english_asr_model.modules.wav2vec2(batch, lens)
|
506 |
+
x = english_asr_model.modules.enc(en_embeddings)
|
507 |
+
en_posteriogram = english_asr_model.modules.ctc_lin(x)
|
508 |
+
#scores, en_posteriogram = english_asr_model.mods.decoder(en_embeddings ,lens)
|
509 |
+
if(verbose !=0):
|
510 |
+
print('[EMBEDDINGS] FR:',fr_embeddings.shape, "EN:",en_embeddings.shape, "TN:", tn_embeddings.shape)
|
511 |
+
print('[POSTERIOGRAM] FR:',fr_posteriogram.shape, "EN:",en_posteriogram.shape,"TN:",tn_posteriogram.shape)
|
512 |
+
|
513 |
+
|
514 |
+
bilangual_sample = merge_strategy(fr_embeddings,en_embeddings,tn_embeddings,fr_posteriogram,en_posteriogram,tn_posteriogram)
|
515 |
+
return bilangual_sample
|
516 |
+
|
517 |
+
class Mixer(sb.core.Brain):
|
518 |
+
|
519 |
+
def compute_forward(self, batch, stage):
|
520 |
+
"""Forward computations from the waveform batches to the output probabilities."""
|
521 |
+
wavs, wav_lens = batch.sig
|
522 |
+
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
|
523 |
+
|
524 |
+
if stage == sb.Stage.TRAIN:
|
525 |
+
if hasattr(self.hparams, "augmentation"):
|
526 |
+
wavs = self.hparams.augmentation(wavs, wav_lens)
|
527 |
+
|
528 |
+
multi_langual_feats = middle_layer(wavs, wav_lens)
|
529 |
+
multi_langual_feats= multi_langual_feats.to(device)
|
530 |
+
feats, _ = self.modules.enc(multi_langual_feats)
|
531 |
+
logits = self.modules.ctc_lin(feats)
|
532 |
+
p_ctc = self.hparams.log_softmax(logits)
|
533 |
+
|
534 |
+
if stage!= sb.Stage.TRAIN:
|
535 |
+
p_tokens = sb.decoders.ctc_greedy_decode(
|
536 |
+
p_ctc, wav_lens, blank_id=self.hparams.blank_index
|
537 |
+
)
|
538 |
+
else :
|
539 |
+
p_tokens = None
|
540 |
+
return p_ctc, wav_lens, p_tokens
|
541 |
+
|
542 |
+
def compute_objectives(self, predictions, batch, stage):
|
543 |
+
"""Computes the loss (CTC) given predictions and targets."""
|
544 |
+
|
545 |
+
p_ctc, wav_lens , predicted_tokens= predictions
|
546 |
+
|
547 |
+
ids = batch.id
|
548 |
+
tokens, tokens_lens = batch.tokens
|
549 |
+
|
550 |
+
loss = self.hparams.ctc_cost(p_ctc, tokens, wav_lens, tokens_lens)
|
551 |
+
|
552 |
+
|
553 |
+
if stage == sb.Stage.VALID:
|
554 |
+
predicted_words = [
|
555 |
+
"".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
|
556 |
+
for utt_seq in predicted_tokens
|
557 |
+
]
|
558 |
+
target_words = [wrd.split(" ") for wrd in batch.wrd]
|
559 |
+
self.wer_metric.append(ids, predicted_words, target_words)
|
560 |
+
self.cer_metric.append(ids, predicted_words, target_words)
|
561 |
+
if stage ==sb.Stage.TEST :
|
562 |
+
if self.hparams.language_modelling:
|
563 |
+
predicted_words = []
|
564 |
+
for logs in p_ctc:
|
565 |
+
text = decoder.decode(logs.detach().cpu().numpy())
|
566 |
+
predicted_words.append(text.split(" "))
|
567 |
+
else :
|
568 |
+
predicted_words = [
|
569 |
+
"".join(self.tokenizer.decode_ndim(utt_seq)).split(" ")
|
570 |
+
for utt_seq in predicted_tokens
|
571 |
+
]
|
572 |
+
|
573 |
+
target_words = [wrd.split(" ") for wrd in batch.wrd]
|
574 |
+
self.wer_metric.append(ids, predicted_words, target_words)
|
575 |
+
self.cer_metric.append(ids, predicted_words, target_words)
|
576 |
+
|
577 |
+
return loss
|
578 |
+
|
579 |
+
def fit_batch(self, batch):
|
580 |
+
"""Train the parameters given a single batch in input"""
|
581 |
+
should_step = self.step % self.grad_accumulation_factor == 0
|
582 |
+
# Managing automatic mixed precision
|
583 |
+
# TOFIX: CTC fine-tuning currently is unstable
|
584 |
+
# This is certainly due to CTC being done in fp16 instead of fp32
|
585 |
+
if self.auto_mix_prec:
|
586 |
+
with torch.cuda.amp.autocast():
|
587 |
+
with self.no_sync():
|
588 |
+
outputs = self.compute_forward(batch, sb.Stage.TRAIN)
|
589 |
+
loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
|
590 |
+
with self.no_sync(not should_step):
|
591 |
+
self.scaler.scale(
|
592 |
+
loss / self.grad_accumulation_factor
|
593 |
+
).backward()
|
594 |
+
if should_step:
|
595 |
+
|
596 |
+
|
597 |
+
self.scaler.unscale_(self.model_optimizer)
|
598 |
+
if self.check_gradients(loss):
|
599 |
+
self.scaler.step(self.model_optimizer)
|
600 |
+
self.scaler.update()
|
601 |
+
self.zero_grad()
|
602 |
+
self.optimizer_step += 1
|
603 |
+
else:
|
604 |
+
# This is mandatory because HF models have a weird behavior with DDP
|
605 |
+
# on the forward pass
|
606 |
+
with self.no_sync():
|
607 |
+
outputs = self.compute_forward(batch, sb.Stage.TRAIN)
|
608 |
+
|
609 |
+
loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
|
610 |
+
|
611 |
+
with self.no_sync(not should_step):
|
612 |
+
(loss / self.grad_accumulation_factor).backward()
|
613 |
+
if should_step:
|
614 |
+
if self.check_gradients(loss):
|
615 |
+
self.model_optimizer.step()
|
616 |
+
self.zero_grad()
|
617 |
+
self.optimizer_step += 1
|
618 |
+
|
619 |
+
self.on_fit_batch_end(batch, outputs, loss, should_step)
|
620 |
+
return loss.detach().cpu()
|
621 |
+
|
622 |
+
def evaluate_batch(self, batch, stage):
|
623 |
+
"""Computations needed for validation/test batches"""
|
624 |
+
predictions = self.compute_forward(batch, stage=stage)
|
625 |
+
with torch.no_grad():
|
626 |
+
loss = self.compute_objectives(predictions, batch, stage=stage)
|
627 |
+
return loss.detach()
|
628 |
+
|
629 |
+
def on_stage_start(self, stage, epoch):
|
630 |
+
"""Gets called at the beginning of each epoch"""
|
631 |
+
if stage != sb.Stage.TRAIN:
|
632 |
+
self.cer_metric = self.hparams.cer_computer()
|
633 |
+
self.wer_metric = self.hparams.error_rate_computer()
|
634 |
+
|
635 |
+
def on_stage_end(self, stage, stage_loss, epoch):
|
636 |
+
"""Gets called at the end of an epoch."""
|
637 |
+
# Compute/store important stats
|
638 |
+
stage_stats = {"loss": stage_loss}
|
639 |
+
if stage == sb.Stage.TRAIN:
|
640 |
+
self.train_stats = stage_stats
|
641 |
+
else:
|
642 |
+
stage_stats["CER"] = self.cer_metric.summarize("error_rate")
|
643 |
+
stage_stats["WER"] = self.wer_metric.summarize("error_rate")
|
644 |
+
|
645 |
+
# Perform end-of-iteration things, like annealing, logging, etc.
|
646 |
+
if stage == sb.Stage.VALID:
|
647 |
+
old_lr_model, new_lr_model = self.hparams.lr_annealing_model(
|
648 |
+
stage_stats["loss"]
|
649 |
+
)
|
650 |
+
sb.nnet.schedulers.update_learning_rate(
|
651 |
+
self.model_optimizer, new_lr_model
|
652 |
+
)
|
653 |
+
self.hparams.train_logger.log_stats(
|
654 |
+
stats_meta={
|
655 |
+
"epoch": epoch,
|
656 |
+
"lr_model": old_lr_model,
|
657 |
+
},
|
658 |
+
train_stats=self.train_stats,
|
659 |
+
valid_stats=stage_stats,
|
660 |
+
)
|
661 |
+
self.checkpointer.save_and_keep_only(
|
662 |
+
meta={"WER": stage_stats["WER"]}, min_keys=["WER"],
|
663 |
+
)
|
664 |
+
elif stage == sb.Stage.TEST:
|
665 |
+
self.hparams.train_logger.log_stats(
|
666 |
+
stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
|
667 |
+
test_stats=stage_stats,
|
668 |
+
)
|
669 |
+
with open(self.hparams.wer_file, "w") as w:
|
670 |
+
self.wer_metric.write_stats(w)
|
671 |
+
|
672 |
+
def init_optimizers(self):
|
673 |
+
|
674 |
+
self.model_optimizer = self.hparams.model_opt_class(
|
675 |
+
self.hparams.model.parameters()
|
676 |
+
)
|
677 |
+
|
678 |
+
if self.checkpointer is not None:
|
679 |
+
self.checkpointer.add_recoverable("modelopt", self.model_optimizer)
|
680 |
+
|
681 |
+
def zero_grad(self, set_to_none=False):
|
682 |
+
|
683 |
+
self.model_optimizer.zero_grad(set_to_none)
|
684 |
+
|
685 |
+
|
686 |
+
hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
|
687 |
+
|
688 |
+
# If distributed_launch=True then
|
689 |
+
# create ddp_group with the right communication protocol
|
690 |
+
sb.utils.distributed.ddp_init_group(run_opts)
|
691 |
+
|
692 |
+
with open(hparams_file) as fin:
|
693 |
+
hparams = load_hyperpyyaml(fin, overrides)
|
694 |
+
|
695 |
+
# Create experiment directory
|
696 |
+
sb.create_experiment_directory(
|
697 |
+
experiment_directory=hparams["output_folder"],
|
698 |
+
hyperparams_to_save=hparams_file,
|
699 |
+
overrides=overrides,
|
700 |
+
)
|
701 |
+
def read_labels_file(labels_file):
|
702 |
+
with open(labels_file, "r",encoding="utf-8") as lf:
|
703 |
+
lines = lf.read().splitlines()
|
704 |
+
division = "==="
|
705 |
+
numbers = {}
|
706 |
+
for line in lines :
|
707 |
+
if division in line :
|
708 |
+
break
|
709 |
+
string, number = line.split("=>")
|
710 |
+
number = int(number)
|
711 |
+
string = string[1:-2]
|
712 |
+
numbers[number] = string
|
713 |
+
return [numbers[x] for x in range(len(numbers))]
|
714 |
+
train_data, valid_data, test_datasets, label_encoder = dataio_prepare(
|
715 |
+
hparams
|
716 |
+
)
|
717 |
+
|
718 |
+
|
719 |
+
labels = read_labels_file(os.path.join(hparams["save_folder"], "label_encoder.txt"))
|
720 |
+
labels = [""] + labels[1:-1] + ["1"]
|
721 |
+
if hparams["language_modelling"]:
|
722 |
+
decoder = build_ctcdecoder(
|
723 |
+
labels,
|
724 |
+
kenlm_model_path=hparams["ngram_lm_path"], # either .arpa or .bin file
|
725 |
+
alpha=0.5, # tuned on a val set
|
726 |
+
beta=1, # tuned on a val set
|
727 |
+
)
|
728 |
+
|
729 |
+
|
730 |
+
|
731 |
+
|
732 |
+
mixer = Mixer(
|
733 |
+
modules=hparams["modules"],
|
734 |
+
hparams=hparams,
|
735 |
+
run_opts=run_opts,
|
736 |
+
checkpointer=hparams["checkpointer"],
|
737 |
+
)
|
738 |
+
mixer.tokenizer = label_encoder
|
739 |
+
|
740 |
+
|
741 |
+
mixer.fit(
|
742 |
+
mixer.hparams.epoch_counter,
|
743 |
+
train_data,
|
744 |
+
valid_data,
|
745 |
+
train_loader_kwargs=hparams["dataloader_options"],
|
746 |
+
valid_loader_kwargs=hparams["test_dataloader_options"],
|
747 |
+
)
|
748 |
+
print(test_datasets.keys())
|
749 |
+
for k in test_datasets.keys(): # keys are test_clean, test_other etc
|
750 |
+
mixer.hparams.wer_file = os.path.join(
|
751 |
+
hparams["output_folder"], "wer_{}.txt".format(k)
|
752 |
+
)
|
753 |
+
mixer.evaluate(
|
754 |
+
test_datasets[k], test_loader_kwargs=hparams["test_dataloader_options"]
|
755 |
+
)
|
756 |
+
|
results/non_semi_final_stac/hyperparams.yaml
ADDED
@@ -0,0 +1,144 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Generated 2023-09-08 from:
|
2 |
+
# /gpfsssd/scratch/rech/nou/uzn19yk/switched_data/stac.yaml
|
3 |
+
# yamllint disable
|
4 |
+
# Generated 2023-08-03 from:
|
5 |
+
# /home/salah/new_tunisian_model/hparams/train_tunisian_withwavlm.yaml
|
6 |
+
# yamllint disable
|
7 |
+
# ################################
|
8 |
+
# Model: wav2vec2 + DNN + CTC
|
9 |
+
# Augmentation: SpecAugment
|
10 |
+
# Authors: Titouan Parcollet 2021
|
11 |
+
# ################################
|
12 |
+
|
13 |
+
seed: 1994
|
14 |
+
__set_seed: !!python/object/apply:torch.manual_seed [1234]
|
15 |
+
output_folder: results/non_semi_final_stac
|
16 |
+
wer_file: results/non_semi_final_stac/wer.txt
|
17 |
+
save_folder: results/non_semi_final_stac/save
|
18 |
+
train_log: results/non_semi_final_stac/train_log.txt
|
19 |
+
|
20 |
+
|
21 |
+
|
22 |
+
# Data files
|
23 |
+
data_folder: junk # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
|
24 |
+
train_tsv_file: junk/train.tsv # Standard CommonVoice .tsv files
|
25 |
+
dev_tsv_file: junk/dev.tsv # Standard CommonVoice .tsv files
|
26 |
+
test_tsv_file: junk/test.tsv # Standard CommonVoice .tsv files
|
27 |
+
accented_letters: true
|
28 |
+
|
29 |
+
csv_folder: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean/
|
30 |
+
train_csv: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean//train.csv
|
31 |
+
valid_csv: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean//dev.csv
|
32 |
+
test_csv:
|
33 |
+
- all_tests/cs_test.csv
|
34 |
+
- all_tests/stac_test.csv
|
35 |
+
|
36 |
+
# We remove utterance slonger than 10s in the train/dev/test sets as
|
37 |
+
# longer sentences certainly correspond to "open microphones".
|
38 |
+
avoid_if_longer_than: 13.0
|
39 |
+
avoid_if_shorter_than: 0.5
|
40 |
+
|
41 |
+
# Training parameters
|
42 |
+
number_of_epochs: 20
|
43 |
+
lr: 0.0002
|
44 |
+
lr_weights: 0.01
|
45 |
+
sorting: ascending
|
46 |
+
auto_mix_prec: false
|
47 |
+
sample_rate: 16000
|
48 |
+
language_modelling: true
|
49 |
+
ngram_lm_path:
|
50 |
+
/gpfsstore/rech/nou/uzn19yk/switched_code_tunisian/train/tunisian_asr/arpas/pluslanguages_everything.arpa
|
51 |
+
|
52 |
+
# With data_parallel batch_size is split into N jobs
|
53 |
+
# With DDP batch_size is multiplied by N jobs
|
54 |
+
# Must be 3 per GPU to fit 32GB of VRAM
|
55 |
+
batch_size: 3
|
56 |
+
test_batch_size: 4
|
57 |
+
|
58 |
+
# Dataloader options
|
59 |
+
dataloader_options:
|
60 |
+
batch_size: 3
|
61 |
+
num_workers: 6
|
62 |
+
|
63 |
+
test_dataloader_options:
|
64 |
+
batch_size: 4
|
65 |
+
num_workers: 6
|
66 |
+
|
67 |
+
# Model parameters
|
68 |
+
activation: !name:torch.nn.Sigmoid
|
69 |
+
dnn_layers: 1
|
70 |
+
dnn_neurons: 768
|
71 |
+
freeze_encoder: true
|
72 |
+
|
73 |
+
# Outputs
|
74 |
+
output_neurons: 76 # BPE size, index(blank/eos/bos) = 0
|
75 |
+
|
76 |
+
# Functions and classes
|
77 |
+
#
|
78 |
+
epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
|
79 |
+
limit: 20
|
80 |
+
|
81 |
+
encoder_dim: 3217
|
82 |
+
enc: &id001 !new:speechbrain.nnet.RNN.LSTM
|
83 |
+
input_shape: [null, null, 3217]
|
84 |
+
num_layers: 2
|
85 |
+
bidirectional: true
|
86 |
+
dropout: 0.2
|
87 |
+
hidden_size: 1024
|
88 |
+
|
89 |
+
ctc_lin: &id002 !new:speechbrain.nnet.linear.Linear
|
90 |
+
|
91 |
+
input_size: 2048
|
92 |
+
n_neurons: 76
|
93 |
+
|
94 |
+
log_softmax: !new:speechbrain.nnet.activations.Softmax
|
95 |
+
apply_log: true
|
96 |
+
|
97 |
+
ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
|
98 |
+
blank_index: 0
|
99 |
+
|
100 |
+
modules:
|
101 |
+
enc: *id001
|
102 |
+
ctc_lin: *id002
|
103 |
+
model: &id003 !new:torch.nn.ModuleList
|
104 |
+
- [*id001, *id002]
|
105 |
+
model_opt_class: !name:torch.optim.Adam
|
106 |
+
lr: 0.0002
|
107 |
+
|
108 |
+
weights_opt_class: !name:torch.optim.Adam
|
109 |
+
lr: 0.01
|
110 |
+
|
111 |
+
lr_annealing_model: &id004 !new:speechbrain.nnet.schedulers.NewBobScheduler
|
112 |
+
initial_value: 0.0002
|
113 |
+
improvement_threshold: 0.0025
|
114 |
+
annealing_factor: 0.8
|
115 |
+
patient: 0
|
116 |
+
|
117 |
+
lr_annealing_weights: &id005 !new:speechbrain.nnet.schedulers.NewBobScheduler
|
118 |
+
initial_value: 0.01
|
119 |
+
improvement_threshold: 0.0025
|
120 |
+
annealing_factor: 0.9
|
121 |
+
patient: 0
|
122 |
+
|
123 |
+
label_encoder: &id007 !new:speechbrain.dataio.encoder.CTCTextEncoder
|
124 |
+
|
125 |
+
|
126 |
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
|
127 |
+
checkpoints_dir: results/non_semi_final_stac/save
|
128 |
+
recoverables:
|
129 |
+
model: *id003
|
130 |
+
scheduler_model: *id004
|
131 |
+
scheduler_encoder: *id005
|
132 |
+
counter: *id006
|
133 |
+
tokenizer: *id007
|
134 |
+
blank_index: 0
|
135 |
+
unk_index: 1
|
136 |
+
|
137 |
+
|
138 |
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
|
139 |
+
save_file: results/non_semi_final_stac/train_log.txt
|
140 |
+
|
141 |
+
error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
|
142 |
+
|
143 |
+
cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
|
144 |
+
split_tokens: true
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/CKPT.yaml
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# yamllint disable
|
2 |
+
WER: 51.292116454039906
|
3 |
+
end-of-epoch: true
|
4 |
+
unixtime: 1694130018.9642384
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/brain.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e5c026fe6fa51700406bd476e131950c797b0b3bacb3daae0854e85689bb4cf9
|
3 |
+
size 50
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/counter.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f5ca38f748a1d6eaf726b8a42fb575c3c71f1864a8143301782de13da2d9202b
|
3 |
+
size 2
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/dataloader-TRAIN.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d7e1edcac43af8cea1439d222314af06354ae31da6a3d90b8cc6bcebc5c8e397
|
3 |
+
size 4
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/model.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:da683a8efa5709a06af9b258452c243da841780a0a7942c196c472a3e21e5010
|
3 |
+
size 240389017
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/modelopt.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:416feb314443cf839f4425fc382e555dec90e3dea26fa52b75e4ac1b702c5078
|
3 |
+
size 480787579
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/scheduler_encoder.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:2e2efd50f0cf28a080e2625fdd8a1852c669841537cdc0a57fce60bc6c1eec11
|
3 |
+
size 515
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/scheduler_model.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cec54cc9236fa7aa965b397675d24299b973675cc0c6345de038fc70e51629ab
|
3 |
+
size 703
|
results/non_semi_final_stac/save/CKPT+2023-09-08+01-40-18+00/tokenizer.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:21080a140faeb4f39fad188aaf081914ec782be9c4320d6415e8822709e18017
|
3 |
+
size 39
|
results/non_semi_final_stac/save/label_encoder.txt
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
'و' => 74
|
2 |
+
'ي' => 1
|
3 |
+
'ن' => 2
|
4 |
+
' ' => 3
|
5 |
+
'م' => 4
|
6 |
+
'ش' => 5
|
7 |
+
'ل' => 6
|
8 |
+
'س' => 7
|
9 |
+
'ت' => 8
|
10 |
+
'ا' => 9
|
11 |
+
'د' => 10
|
12 |
+
'ر' => 11
|
13 |
+
'ى' => 12
|
14 |
+
'ب' => 13
|
15 |
+
'ح' => 14
|
16 |
+
'ط' => 15
|
17 |
+
'ع' => 16
|
18 |
+
'ك' => 17
|
19 |
+
'ف' => 18
|
20 |
+
'ق' => 19
|
21 |
+
'ذ' => 20
|
22 |
+
'ث' => 21
|
23 |
+
'ج' => 22
|
24 |
+
'ة' => 23
|
25 |
+
'غ' => 24
|
26 |
+
'o' => 25
|
27 |
+
'k' => 26
|
28 |
+
'b' => 27
|
29 |
+
'n' => 28
|
30 |
+
'خ' => 29
|
31 |
+
'ه' => 30
|
32 |
+
'v' => 31
|
33 |
+
'i' => 32
|
34 |
+
'l' => 33
|
35 |
+
'à' => 34
|
36 |
+
'ص' => 35
|
37 |
+
'ض' => 36
|
38 |
+
'a' => 37
|
39 |
+
'u' => 38
|
40 |
+
't' => 39
|
41 |
+
'm' => 40
|
42 |
+
'q' => 41
|
43 |
+
'e' => 42
|
44 |
+
'd' => 43
|
45 |
+
'c' => 44
|
46 |
+
'p' => 45
|
47 |
+
'r' => 46
|
48 |
+
'أ' => 47
|
49 |
+
'إ' => 48
|
50 |
+
's' => 49
|
51 |
+
'j' => 50
|
52 |
+
'ز' => 51
|
53 |
+
'ء' => 52
|
54 |
+
'h' => 53
|
55 |
+
'f' => 54
|
56 |
+
'آ' => 55
|
57 |
+
'ئ' => 56
|
58 |
+
'ؤ' => 57
|
59 |
+
'ظ' => 58
|
60 |
+
'y' => 59
|
61 |
+
'é' => 60
|
62 |
+
"'" => 61
|
63 |
+
'z' => 62
|
64 |
+
'x' => 63
|
65 |
+
'w' => 64
|
66 |
+
'g' => 65
|
67 |
+
'è' => 66
|
68 |
+
'û' => 67
|
69 |
+
'ç' => 68
|
70 |
+
'ê' => 69
|
71 |
+
'ô' => 70
|
72 |
+
'ù' => 71
|
73 |
+
'î' => 72
|
74 |
+
'â' => 73
|
75 |
+
'<blank>' => 0
|
76 |
+
1 => 75
|
77 |
+
================
|
78 |
+
'starting_index' => 0
|
79 |
+
'unk_label' => 1
|
80 |
+
'blank_label' => '<blank>'
|