-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_localization_and_classification.py
287 lines (237 loc) · 13.3 KB
/
train_localization_and_classification.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
import glob
import os
import warnings
warnings.filterwarnings('ignore')
import shutil
import torch
import torch.nn as nn
import torchaudio
import torchaudio.functional as F
import torchaudio.transforms as T
from torch.utils.data import DataLoader
import argparse
import matplotlib
import matplotlib.pyplot as plt
import pandas as pd
import time
from IPython.display import Audio, display
from sklearn.metrics import f1_score, accuracy_score
import numpy as np
from dataset import BGGunDataset, ForenGunDataset
from models import CNNExtractor, RNNExtractor, CRNNExtractor, TransformerExtractor, CTransExtractor, MultitaskClassifer, \
OurClassifer, OurExtractor
def load_model(model, path):
model.load_state_dict(torch.load(path))
if __name__ == '__main__':
print(torch.__version__)
print(torchaudio.__version__)
parser = argparse.ArgumentParser()
parser.add_argument('--backbone', default='CNN', type=str, choices=['CNN', 'RNN', 'CRNN', 'Trans', 'CTrans','Our'])
parser.add_argument('--pretrained', default=None, type=str)
parser.add_argument('--dataset', default='BGG', type=str, choices=['BGG', 'Foren', 'Urban'])
parser.add_argument('--datadir', default='./data/gun_sound_v2', type=str)
parser.add_argument('--input_sec', default=3, type=int)
parser.add_argument('--sr', default=44100, type=int)
parser.add_argument('--lr', default=1e-4, type=float)
parser.add_argument('--batch_size', default=128, type=int)
parser.add_argument('--epochs', default=200, type=int)
parser.add_argument('--hdim', default=64, type=int)
parser.add_argument('--save', default=None, type=str)
parser.add_argument('--save_midi', action='store_true')
parser.add_argument('--debug', action='store_true')
args = parser.parse_args()
print(args)
# load dataset
if args.dataset == 'BGG':
a = pd.read_csv('./data/v3_exp1_train.csv')
b = pd.read_csv('./data/v3_exp2_train.csv')
train_df = pd.concat([a, b], axis=0)
#train_df = b
a = pd.read_csv('./data/v3_exp1_test.csv')
b = pd.read_csv('./data/v3_exp2_test.csv')
#test_df = b
test_df = pd.concat([a, b], axis=0)
train_dataset = BGGunDataset(args.datadir, train_df, args.sr, args.input_sec, 'train')
val_dataset = BGGunDataset(args.datadir, train_dataset.get_val_df(), args.sr, args.input_sec, 'val',
dicts=train_dataset.dicts)
test_dataset = BGGunDataset(args.datadir, test_df, args.sr, args.input_sec, 'test',
dicts=train_dataset.dicts)
elif args.dataset == 'Foren':
train_dataset = ForenGunDataset(args.datadir, './data/GunshotAudioForensicsDataset/train_v2.csv', args.sr, args.input_sec, 'train',
)
val_dataset = ForenGunDataset(args.datadir, './data/GunshotAudioForensicsDataset/val_v2.csv', args.sr, args.input_sec, 'val',
dicts=train_dataset.dicts)
test_dataset = ForenGunDataset(args.datadir, './data/GunshotAudioForensicsDataset/test_v2.csv', args.sr, args.input_sec, 'test',
dicts=train_dataset.dicts)
label_dicts = train_dataset.dicts
print(label_dicts)
print(f'EXP2 Gunshot Localization and Classification Training:{len(train_dataset)}, Val:{len(val_dataset)}, Test:{len(test_dataset)}')
print()
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4)
val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4)
# build model
if args.backbone == 'CNN':
backbone = CNNExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
elif args.backbone == 'RNN':
backbone = RNNExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
elif args.backbone == 'CRNN':
backbone = CRNNExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
elif args.backbone == 'Trans':
backbone = TransformerExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
elif args.backbone == 'CTrans':
backbone = CTransExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
elif args.backbone == 'Our':
backbone = OurExtractor(args.hdim, sample_rate=args.sr, n_fft=512, n_mels=96).cuda()
else:
raise Exception(f'Not Supported Backbone:{args.backbone}')
if args.pretrained:
load_model(backbone, args.pretrained)
print(f'pretrained from {args.pretrained}')
# savepath = 'from-pretrained-'+ savepath
if args.backbone != 'Our':
classifier = MultitaskClassifer(args.hdim, len(label_dicts['cate']),
len(label_dicts['dist']), len(label_dicts['dire'])).cuda()
elif args.backbone == 'Our':
classifier = OurClassifer(args.hdim, len(label_dicts['cate']),
len(label_dicts['dist']), len(label_dicts['dire'])).cuda()
print(backbone)
print(classifier)
params = sum(p.numel() for p in backbone.parameters() if p.requires_grad)
print(f'# of Backbone Parameters:{params}')
ce = nn.CrossEntropyLoss()
opt = torch.optim.Adam(list(backbone.parameters()) + list(classifier.parameters()), lr=args.lr)
best_val_ce = 1000
for epoch in range(args.epochs):
backbone.train()
classifier.train()
ce_loss = 0.0
cate_preds, cate_trues, dist_preds, dist_trues, dire_preds, dire_trues = [], [], [], [], [], []
for step, (wv, cate, dist, dire) in enumerate(train_loader):
wv, cate, dist, dire = wv.cuda(), cate.cuda(), dist.cuda(), dire.cuda()
features = backbone(wv)
cate_out, dist_out, dire_out = classifier(features)
cate_loss = ce(cate_out, cate)
dist_loss = ce(dist_out, dist)
dire_loss = ce(dire_out, dire)
cate_pred = torch.argmax(cate_out, dim=-1)
dist_pred = torch.argmax(dist_out, dim=-1)
dire_pred = torch.argmax(dire_out, dim=-1)
loss = cate_loss + dist_loss + dire_loss
opt.zero_grad()
loss.backward()
opt.step()
ce_loss += loss.item()
cate_preds.append(cate_pred.detach().cpu())
cate_trues.append(cate.detach().cpu())
dist_preds.append(dist_pred.detach().cpu())
dist_trues.append(dist.detach().cpu())
dire_preds.append(dire_pred.detach().cpu())
dire_trues.append(dire.detach().cpu())
cate_preds = torch.cat(cate_preds, dim=0).numpy()
cate_trues = torch.cat(cate_trues, dim=0).numpy()
dist_preds = torch.cat(dist_preds, dim=0).numpy()
dist_trues = torch.cat(dist_trues, dim=0).numpy()
dire_preds = torch.cat(dire_preds, dim=0).numpy()
dire_trues = torch.cat(dire_trues, dim=0).numpy()
train_ce_loss = ce_loss / len(train_loader)
train_cate_acc = accuracy_score(cate_trues, cate_preds)
train_cate_f1 = f1_score(cate_trues, cate_preds, average='macro')
train_dist_acc = accuracy_score(dist_trues, dist_preds)
train_dist_f1 = f1_score(dist_trues, dist_preds, average='macro')
train_dire_acc = accuracy_score(dire_trues, dire_preds)
train_dire_f1 = f1_score(dire_trues, dire_preds, average='macro')
with torch.no_grad():
backbone.eval()
classifier.eval()
ce_loss = 0.0
cate_preds, cate_trues, dist_preds, dist_trues, dire_preds, dire_trues = [], [], [], [], [], []
for step, (wv, cate, dist, dire) in enumerate(val_loader):
wv, cate, dist, dire = wv.cuda(), cate.cuda(), dist.cuda(), dire.cuda()
features = backbone(wv)
cate_out, dist_out, dire_out = classifier(features)
cate_loss = ce(cate_out, cate)
dist_loss = ce(dist_out, dist)
dire_loss = ce(dire_out, dire)
loss = cate_loss + dist_loss + dire_loss
cate_pred = torch.argmax(cate_out, dim=-1)
dist_pred = torch.argmax(dist_out, dim=-1)
dire_pred = torch.argmax(dire_out, dim=-1)
ce_loss += loss.item()
cate_preds.append(cate_pred.cpu())
cate_trues.append(cate.cpu())
dist_preds.append(dist_pred.cpu())
dist_trues.append(dist.cpu())
dire_preds.append(dire_pred.cpu())
dire_trues.append(dire.cpu())
cate_preds = torch.cat(cate_preds, dim=0).numpy()
cate_trues = torch.cat(cate_trues, dim=0).numpy()
dist_preds = torch.cat(dist_preds, dim=0).numpy()
dist_trues = torch.cat(dist_trues, dim=0).numpy()
dire_preds = torch.cat(dire_preds, dim=0).numpy()
dire_trues = torch.cat(dire_trues, dim=0).numpy()
val_ce_loss = ce_loss / len(test_loader)
val_cate_acc = accuracy_score(cate_trues, cate_preds)
val_cate_f1 = f1_score(cate_trues, cate_preds, average='macro')
val_dist_acc = accuracy_score(dist_trues, dist_preds)
val_dist_f1 = f1_score(dist_trues, dist_preds, average='macro')
val_dire_acc = accuracy_score(dire_trues, dire_preds)
val_dire_f1 = f1_score(dire_trues, dire_preds, average='macro')
ce_loss = 0.0
cate_preds, cate_trues, dist_preds, dist_trues, dire_preds, dire_trues = [], [], [], [], [], []
for step, (wv, cate, dist, dire) in enumerate(test_loader):
wv, cate, dist, dire = wv.cuda(), cate.cuda(), dist.cuda(), dire.cuda()
features = backbone(wv)
cate_out, dist_out, dire_out = classifier(features)
cate_loss = ce(cate_out, cate)
dist_loss = ce(dist_out, dist)
dire_loss = ce(dire_out, dire)
loss = cate_loss + dist_loss + dire_loss
cate_pred = torch.argmax(cate_out, dim=-1)
dist_pred = torch.argmax(dist_out, dim=-1)
dire_pred = torch.argmax(dire_out, dim=-1)
ce_loss += loss.item()
cate_preds.append(cate_pred.cpu())
cate_trues.append(cate.cpu())
dist_preds.append(dist_pred.cpu())
dist_trues.append(dist.cpu())
dire_preds.append(dire_pred.cpu())
dire_trues.append(dire.cpu())
cate_preds = torch.cat(cate_preds, dim=0).numpy()
cate_trues = torch.cat(cate_trues, dim=0).numpy()
dist_preds = torch.cat(dist_preds, dim=0).numpy()
dist_trues = torch.cat(dist_trues, dim=0).numpy()
dire_preds = torch.cat(dire_preds, dim=0).numpy()
dire_trues = torch.cat(dire_trues, dim=0).numpy()
test_ce_loss = ce_loss / len(test_loader)
test_cate_acc = accuracy_score(cate_trues, cate_preds)
test_cate_f1 = f1_score(cate_trues, cate_preds, average='macro')
test_dist_acc = accuracy_score(dist_trues, dist_preds)
test_dist_f1 = f1_score(dist_trues, dist_preds, average='macro')
test_dire_acc = accuracy_score(dire_trues, dire_preds)
test_dire_f1 = f1_score(dire_trues, dire_preds, average='macro')
print(epoch, f'train ce:{train_ce_loss:.4f}, '
f'train cate-acc:{train_cate_acc:.4f}, train cate-F1:{train_cate_f1:.4f}, '
f'train dist-acc:{train_dist_acc:.4f}, train dist-F1:{train_dist_f1:.4f}, '
f'train dire-acc:{train_dire_acc:.4f}, train dire-F1:{train_dire_f1:.4f} '
f'val ce:{val_ce_loss:.4f}, '
f'val cate-acc:{val_cate_acc:.4f}, val cate-F1:{val_cate_f1:.4f}, '
f'val dist-acc:{val_dist_acc:.4f}, val dist-F1:{val_dist_f1:.4f}, '
f'val dire-acc:{val_dire_acc:.4f}, val dire-F1:{val_dire_f1:.4f} '
f'test ce:{test_ce_loss:.4f}, '
f'test cate-acc:{test_cate_acc:.4f}, test cate-F1:{test_cate_f1:.4f}, '
f'test dist-acc:{test_dist_acc:.4f}, test dist-F1:{test_dist_f1:.4f}, '
f'test dire-acc:{train_dire_acc:.4f}, test dire-F1:{test_dire_f1:.4f}')
if best_val_ce >= val_ce_loss:
best_val_ce = val_ce_loss
best_results = [str(i) for i in [epoch, train_ce_loss, train_cate_acc, train_cate_f1,
train_dist_acc, train_dist_f1, train_dire_acc, train_dire_f1,
val_ce_loss, val_cate_acc, val_cate_f1,
val_dist_acc, val_dist_f1, val_dire_acc, val_dire_f1,
test_ce_loss, test_cate_acc, test_cate_f1,
test_dist_acc, test_dist_f1, train_dire_acc, test_dire_f1]]
best_results = '\t'.join(best_results)
if args.dataset == 'BGG':
torch.save(backbone.state_dict(), f'demo_{args.backbone}_backbone.pt')
torch.save(classifier.state_dict(), f'demo_{args.backbone}_classifier.pt')
print(best_results)