-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
executable file
·314 lines (258 loc) · 10.4 KB
/
main.py
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
import os
import argparse
import torch
import numpy as np
import pandas as pd
from torch import nn
import neptune.new as neptune
from omegaconf import OmegaConf
from datasets import load_from_disk
from evotorch.algorithms import SNES
from evotorch.logging import NeptuneLogger, StdOutLogger
from evotorch.neuroevolution.net import count_parameters
from transformers.trainer_utils import get_last_checkpoint
from transformers import (
AutoConfig,
AutoFeatureExtractor,
Wav2Vec2Model
)
from dotenv import load_dotenv
from sklearn.metrics import (
accuracy_score,
f1_score,
precision_score,
recall_score
)
from utils.utils import (
DataColletor,
CollateWav2vec2,
DatasetEVO,
CustomSupervisedNE,
preprocess_data,
preprocess_metadata,
get_label_id,
predict
)
from utils.model import FullyConnectedEvo
load_dotenv()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument(
'-c',
'--config_path',
default='config/default.yaml',
type=str,
help="YAML file with configurations"
)
parser.add_argument(
'--continue_train',
default=False,
action='store_true',
help='If True, continues training using the checkpoint_path parameter'
)
parser.add_argument(
'--train',
default=False,
action='store_true',
help='If True, runs ga'
)
parser.add_argument(
'--test',
default=False,
action='store_true'
)
parser.add_argument(
'--pre_processing',
default=False,
action='store_true'
)
args = parser.parse_args()
cfg = OmegaConf.load(args.config_path)
if args.train:
accs_val = []
recalls_val = []
precisions_val = []
f1_scores_val = []
accs_test = []
recalls_test = []
precisions_test = []
f1_scores_test = []
# Load preloaded pre-processed audio signals
train_dataset = load_from_disk(cfg.metadata.preloaded_train)
# Load Network to check how many parameters it has
network = FullyConnectedEvo(
input_size=cfg.fc_model.input_size,
num_classes=cfg.fc_model.num_classes
)
print(f'Network has {count_parameters(network)} parameters')
# Runs N experiments
for experiment in range(cfg.train.num_experiments):
print(f"Experiment: {experiment+1}/{cfg.train.num_experiments}")
train_dataset_final = DatasetEVO(batch=train_dataset)
ser_problem = CustomSupervisedNE(
train_dataset_final, # Using the dataset specified earlier
FullyConnectedEvo, # Training the MNIST30K module designed earlier
nn.CrossEntropyLoss(), # Minimizing CrossEntropyLoss
network_args={
"input_size": cfg.fc_model.input_size,
"num_classes": cfg.fc_model.num_classes
},
minibatch_size=cfg.genetic_algorithm.minibatch_size, # With a minibatch size of 256
common_minibatch=True, # Always using the same minibatch across all solutions on an actor
num_actors=1, # The total number of CPUs used
device=device
)
searcher = SNES(
ser_problem,
stdev_init=cfg.genetic_algorithm.stdev_init,
popsize=cfg.genetic_algorithm.population_size
)
print("Searcher keys: ")
print([k for k in searcher.iter_status_keys()])
run = neptune.init(
project=os.getenv('NEPTUNE_PROJECT'),
api_token=os.getenv('NEPTUNE_API_TOKEN')
)
run["parameters"] = {
**cfg.metadata,
**cfg.fc_model,
**cfg.genetic_algorithm,
**cfg.data,
**cfg.train,
**cfg.test,
**cfg.logging
}
StdOutLogger(searcher, interval=1)
NeptuneLogger(
searcher,
interval = 1,
run=run
)
searcher.run(cfg.genetic_algorithm.generations)
run.stop()
print(searcher.status['center'])
print(searcher.status['pop_best_eval'])
net = ser_problem.parameterize_net(searcher.status['center']).cpu()
net.eval()
val_dataset = load_from_disk(cfg.metadata.preloaded_val)
test_dataset = load_from_disk(cfg.metadata.preloaded_test)
val_dataset_final = DatasetEVO(batch=val_dataset)
test_dataset_final = DatasetEVO(batch=test_dataset)
val_loader_final = torch.utils.data.DataLoader(
val_dataset_final,
batch_size=8,
shuffle=True,
drop_last=False,
num_workers=10
)
test_loader_final = torch.utils.data.DataLoader(
test_dataset_final,
batch_size=8,
shuffle=True,
drop_last=False,
num_workers=10
)
preds_val, labels_ids_val, test_loss_val = predict(test_dataloader=val_loader_final, model=net)
preds_test, labels_ids_test, test_loss_test = predict(test_dataloader=test_loader_final, model=net)
new_labels_ids_val = []
new_labels_ids_test = []
for labels_id in labels_ids_val:
new_labels_ids_val.append(int(labels_id.cpu().detach().numpy()))
for labels_id in labels_ids_test:
new_labels_ids_test.append(int(labels_id.cpu().detach().numpy()))
accs_val.append(accuracy_score(new_labels_ids_val, preds_val))
recalls_val.append(precision_score(new_labels_ids_val, preds_val, average='macro'))
precisions_val.append(recall_score(new_labels_ids_val, preds_val, average='macro'))
f1_scores_val.append(f1_score(new_labels_ids_val, preds_val, average='macro'))
accs_test.append(accuracy_score(new_labels_ids_test, preds_test))
recalls_test.append(precision_score(new_labels_ids_test, preds_test, average='macro'))
precisions_test.append(recall_score(new_labels_ids_test, preds_test, average='macro'))
f1_scores_test.append(f1_score(new_labels_ids_test, preds_test, average='macro'))
os.makedirs("scores", exist_ok=True)
df_val = pd.DataFrame(list(zip(accs_val, recalls_val, precisions_val, f1_scores_val)), columns =['Acc', 'Recall', 'Precision', 'F1_Score'])
df_test = pd.DataFrame(list(zip(accs_test, recalls_test, precisions_test, f1_scores_test)), columns =['Acc', 'Recall', 'Precision', 'F1_Score'])
df_val.to_csv(
f"scores/Val-Pop{cfg.genetic_algorithm.population_size}-batch_{cfg.genetic_algorithm.minibatch_size}-gen_{cfg.genetic_algorithm.generations}.csv",
index=False
)
df_test.to_csv(
f"scores/Test-Pop{cfg.genetic_algorithm.population_size}-batch_{cfg.genetic_algorithm.minibatch_size}-gen_{cfg.genetic_algorithm.generations}.csv",
index=False
)
if args.pre_processing:
if os.path.isdir(cfg.train.model_checkpoint):
last_checkpoint = get_last_checkpoint(cfg.train.model_checkpoint)
print("> Resuming Train with checkpoint: ", last_checkpoint)
else:
last_checkpoint = None
train_df = pd.read_csv(cfg.metadata.train_path)
val_df = pd.read_csv(cfg.metadata.dev_path)
test_df = pd.read_csv(cfg.metadata.test_path)
train_dataset = preprocess_metadata(df=train_df)
val_dataset = preprocess_metadata(df=val_df)
test_dataset = preprocess_metadata(df=test_df)
label2id, id2label, num_labels = get_label_id(dataset=train_dataset, label_column=cfg.metadata.label_column)
feature_extractor = AutoFeatureExtractor.from_pretrained(cfg.train.model_checkpoint)
config = AutoConfig.from_pretrained(
pretrained_model_name_or_path=last_checkpoint if last_checkpoint else cfg.train.model_checkpoint,
num_labels=num_labels,
label2id=label2id,
id2label=id2label,
)
print(label2id)
model = Wav2Vec2Model(config).to(device)
model.freeze_feature_encoder()
train_data_collator = DataColletor(
batch=train_dataset,
processor=feature_extractor,
sampling_rate=cfg.data.target_sampling_rate,
apply_dbfs_norm=cfg.data.apply_dbfs_norm,
target_dbfs=cfg.data.target_dbfs,
label2id=label2id
)
val_data_collator = DataColletor(
batch=val_dataset,
processor=feature_extractor,
sampling_rate=cfg.data.target_sampling_rate,
apply_dbfs_norm=cfg.data.apply_dbfs_norm,
target_dbfs=cfg.data.target_dbfs,
label2id=label2id
)
test_data_collator = DataColletor(
batch=test_dataset,
processor=feature_extractor,
sampling_rate=cfg.data.target_sampling_rate,
apply_dbfs_norm=cfg.data.apply_dbfs_norm,
target_dbfs=cfg.data.target_dbfs,
label2id=label2id
)
train_loader = torch.utils.data.DataLoader(
train_data_collator,
batch_size=1,
shuffle=True,
drop_last=False,
num_workers=10,
collate_fn=CollateWav2vec2(feature_extractor)
)
val_loader = torch.utils.data.DataLoader(
val_data_collator,
batch_size=1,
shuffle=True,
drop_last=False,
num_workers=10,
collate_fn=CollateWav2vec2(feature_extractor)
)
test_loader = torch.utils.data.DataLoader(
test_data_collator,
batch_size=1,
shuffle=True,
drop_last=False,
num_workers=10,
collate_fn=CollateWav2vec2(feature_extractor)
)
preprocess_data(data_loader=train_loader, model=model, out_path=cfg.metadata.preloaded_train)
preprocess_data(data_loader=val_loader, model=model, out_path=cfg.metadata.preloaded_val)
preprocess_data(data_loader=test_loader, model=model, out_path=cfg.metadata.preloaded_test)
if __name__ == '__main__':
main()