-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
470 lines (393 loc) · 20.6 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
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
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
import numpy as np
import os, time, subprocess
import matplotlib.pyplot as plt
import seaborn as sns
import scipy
import librosa
from helpers import *
from MIMII import *
import torch
import config
import shutil
from torchsummary.torchsummary import summary
from torchTools import *
from torch.utils.data.sampler import SubsetRandomSampler, SequentialSampler
from models import *
from sklearn.metrics import roc_auc_score, roc_curve
import matplotlib as mpl
import matplotlib
from sklearn.svm import OneClassSVM
def train_neural_embeddings(f_train, f_val, f_test, epochs=100, center_loss=True,
rotate_train=False, rotate_test=False, length_s=2., channels=None, embedding_dimension=25,
batch_size=128, lr=1e-3, lr_center=0.01, min_delta=1e-3):
snrs = [6, 0]
# channels = [0]
dataset_n = dataset(length_s=length_s, labels=['normal'], target=['type', 'id'], snrs=snrs, channels=channels)
idx_all = dataset_n.get_all_ids()
idx_train, idx_val, idx_test = [], [], []
f_all = dataset_n.get_filepaths(idx_all)
for i, f in tqdm(enumerate(f_all), desc='Identifying audio samples'):
if f in f_train:
idx_train.append(i)
elif f in f_val:
idx_val.append(i)
elif f in f_test:
idx_test.append(i)
else:
raise ValueError('ID does not belong anywhere. What a pitty, a pitty, a pitty...')
#assert not any([i in idx_val for i in idx_train])
#assert not any([i in idx_train for i in idx_val])
#assert not any([i in idx_test for i in idx_train])
#assert not any([i in idx_train for i in idx_test])
trainloader = DataLoader(dataset_n, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=True,
sampler=SubsetRandomSampler(idx_train))
valloader = DataLoader(dataset_n, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=False,
sampler=SequentialSampler(idx_val))
if not os.path.exists('./images/'): create_folder('images')
input_shape, output_shape = (len(dataset_n.channels), int(length_s*fs)), len(dataset_n.get_classes())
model = RawdNet(input_shape, output_shape, embedding_dimension)
if torch.cuda.is_available():
device = torch.device('cuda')
model.to(device)
summary(model, torch.rand((1, *input_shape)))
else:
raise ValueError('No cuda device!')
if center_loss:
nllloss = nn.NLLLoss().cuda() # CrossEntropyLoss = log_softmax + NLLLoss
centerloss = CenterLoss(len(dataset_n.get_classes()), embedding_dimension, 1.0).cuda()
criterion = [nllloss, centerloss]
optimizer4nn = torch.optim.Adam(model.parameters(), lr=lr)
optimzer4center = torch.optim.SGD(centerloss.parameters(), lr=lr_center)
optimizer = [optimizer4nn, optimzer4center]
else:
criterion = torch.nn.NLLLoss().cuda() # torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
train_losses, train_accs = [], []
valid_losses, valid_accs = [], []
early_stopping = EarlyStopping(patience=7, delta=min_delta, path=os.path.join(FOLDER, 'model_checkpoint.pt'), verbose=False)
print('Training...')
for epoch in range(1, epochs+1):
t0 = time.time()
model.train()
# shuffle dataset and organize the indexes into batches
running_loss, running_acc = 0., 0.
ip1s, idxs = [], []
for i, (x_train_batch, y_train_batch) in tqdm(enumerate(trainloader), desc="Epoch %d" % epoch):
# rotate microphone array
if rotate_train:
x_train_batch = np.roll(x_train_batch, np.random.randint(x_train_batch.shape[1]), axis=1)
x_train_batch = torch.Tensor(x_train_batch).type(torch.FloatTensor).cuda()
y_train_batch = torch.Tensor(y_train_batch).type(torch.long).cuda()
if center_loss:
optimizer[0].zero_grad()
optimizer[1].zero_grad()
else:
optimizer.zero_grad()
ip1, y_hat = model(x_train_batch)
if center_loss:
loss = criterion[0](y_hat, torch.argmax(y_train_batch, 1)) + criterion[1](torch.argmax(y_train_batch, 1), ip1)
else:
loss = criterion(y_hat, torch.argmax(y_train_batch, 1))
ip1s.append(ip1)
idxs.append((torch.argmax(y_train_batch, 1)))
running_acc += torch_accuracy(y_hat, y_train_batch)
running_loss += loss.item()
loss.backward()
if center_loss:
optimizer[0].step()
optimizer[1].step()
else:
optimizer.step()
train_loss = running_loss / (i + 1)
train_acc = running_acc / (i + 1)
train_losses.append(train_loss)
train_accs.append(train_acc)
if embedding_dimension == 2:
feat = torch.cat(ip1s, 0)
labels = torch.cat(idxs, 0)
visualize(feat.data.cpu().numpy(), labels.data.cpu().numpy(), dataset_n.get_classes(), epoch)
# validation
with torch.no_grad():
model.eval()
running_loss, running_acc = 0., 0.
for i, (x_val_batch, y_val_batch) in enumerate(valloader):
if rotate_test:
x_val_batch = np.roll(x_val_batch, np.random.randint(x_val_batch.shape[1]), axis=1)
x_val_batch = torch.Tensor(x_val_batch).type(torch.FloatTensor).cuda()
y_val_batch = torch.Tensor(y_val_batch).type(torch.long).cuda()
#print(i, x_val_batch.shape, y_val_batch.shape)
ip1, y_hat = model(x_val_batch)
if center_loss:
loss = criterion[0](y_hat, torch.argmax(y_val_batch, 1)) + criterion[1](torch.argmax(y_val_batch, 1), ip1)
else:
loss = criterion(y_hat, torch.argmax(y_val_batch, 1))
running_acc += torch_accuracy(y_hat, y_val_batch)
running_loss += loss.item()
valid_loss = running_loss / (i + 1)
valid_acc = running_acc / (i + 1)
valid_losses.append(valid_loss)
valid_accs.append(valid_acc)
tab = ' ' if epoch < 10 else ''
print('%ds - epoch: %s%d/%d - loss: %.4f - acc: %.3f - val_loss: %.4f - val_acc: %.3f' % (int(time.time() - t0), tab, epoch, epochs,
train_loss, train_acc, valid_loss, valid_acc)),
early_stopping(valid_loss, model)
if early_stopping.early_stop:
print("Early stopping!")
break
model = torch.load(FOLDER + '\\model_checkpoint.pt')
torch.save(model, FOLDER + '\\model.pt')
if True:
from matplotlib.ticker import MaxNLocator
ax = plt.figure().gca()
plt.plot(range(1,len(train_losses)+1), train_losses, 'b', label='Training loss')
plt.plot(range(1,len(valid_losses)+1), valid_losses, 'g', label='Validation loss')
plt.axvline(early_stopping.best_epoch, linestyle='--', color='r', label='Early Stopping Checkpoint')
plt.xlim(0, len(train_losses) + 1) # consistent scale
plt.grid(True)
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel = str(criterion)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
plt.legend()
plt.tight_layout()
plt.savefig(FOLDER + '/train_history.png', dpi=600) # , bbox_inches='tight')
plt.clf()
plt.cla()
plt.close()
subprocess.Popen('C:\\Users\\VRYSIS\\Dropbox\\condition_monitoring\\' + FOLDER + '/train_history.png', shell=True,
creationflags=subprocess.CREATE_NEW_PROCESS_GROUP)
testloader = DataLoader(dataset_n, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=False,
sampler=SequentialSampler(idx_val))
### evaluate
model.eval()
running_loss, running_acc = 0., 0.
with torch.no_grad():
for i, (x_val_batch, y_val_batch) in enumerate(testloader):
x_val_batch = torch.Tensor(x_val_batch).type(torch.FloatTensor).cuda()
y_val_batch = torch.Tensor(y_val_batch).type(torch.long).cuda()
ip1, y_hat = model(x_val_batch)
if center_loss:
loss = criterion[0](y_hat, torch.argmax(y_val_batch, 1)) + criterion[1](torch.argmax(y_val_batch, 1), ip1)
else:
loss = criterion(y_hat, torch.argmax(y_val_batch, 1))
running_acc += torch_accuracy(y_hat, y_val_batch)
running_loss += loss.item()
test_loss = running_loss / (i + 1)
test_acc = running_acc / (i + 1)
print('test_loss: %.3f - test_acc: %.3f ' % (test_loss, test_acc))
return model
def predictNeuralEmbeddings(model, batch_size, f_test, length_s=1., type='fan', id=4, snr=None, channels=None, epoch=None):
assert snr in [6, 0, -6]
if epoch == None: epoch = ''
# get loaders per id
dataset_n = dataset(length_s=length_s, labels=['normal'], ids=[id], types=[type], snrs=[snr], target=['label'], channels=channels)
dataset_ab = dataset(length_s=length_s, labels=['abnormal'], ids=[id], types=[type], snrs=[snr], target=['label'], channels=channels)
# remove test set sample indices and clear
idx_all = dataset_n.get_all_ids()
idx_test, idx_train = [], []
f_all = dataset_n.get_filepaths(idx_all)
for i, f in enumerate(f_all):
if f not in f_test:
idx_train.append(i)
else:
idx_test.append(i)
print('------------------------------------')
assert not any([i in idx_test for i in idx_train])
assert not any([i in idx_train for i in idx_test])
# construct loaders
trainloader = DataLoader(dataset_n, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=False,
sampler=SequentialSampler(idx_train))
testloader = DataLoader(dataset_n, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=False,
sampler=SequentialSampler(idx_test))
abtestloader = DataLoader(dataset_ab, batch_size=batch_size, pin_memory=True, num_workers=8, drop_last=False,shuffle=False)
model.eval()
x_train, y_train = [], []
with torch.no_grad():
for i, (x_train_batch, y_train_batch) in enumerate(trainloader):
x_train_batch = torch.Tensor(x_train_batch).type(torch.FloatTensor).cuda()
y_train_batch = torch.Tensor(y_train_batch).type(torch.long).cuda()
ip1, y_hat = model(x_train_batch)
x_train.extend(ip1.detach().cpu().numpy().tolist())
y_train.extend(y_train_batch.detach().cpu().numpy().tolist())
x_train = np.array(x_train)
y_train = np.array(y_train)
x_test_n, y_test_n = [], []
for i, (x_test_batch, y_test_batch) in enumerate(testloader):
x_test_batch = torch.Tensor(x_test_batch).type(torch.FloatTensor).cuda()
y_test_batch = torch.Tensor(y_test_batch).type(torch.long).cuda()
ip1, _ = model(x_test_batch)
x_test_n.extend(ip1.detach().cpu().numpy().tolist())
y_test_n.extend(y_test_batch.detach().cpu().numpy().tolist())
x_test_n = np.array(x_test_n)
y_test_n = np.array(y_test_n)
x_test_ab, y_test_ab = [], []
for i, (x_test_batch, y_test_batch) in enumerate(abtestloader):
x_test_batch = torch.Tensor(x_test_batch).type(torch.FloatTensor).cuda()
y_test_batch = torch.Tensor(y_test_batch).type(torch.long).cuda()
ip1, _ = model(x_test_batch)
x_test_ab.extend(ip1.detach().cpu().numpy().tolist())
y_test_ab.extend(y_test_batch.detach().cpu().numpy().tolist())
x_test_ab = np.array(x_test_ab)
y_test_ab = np.array(y_test_ab)
embedding_dimension = x_train.shape[1]
if embedding_dimension == 2:
plt.scatter(x_train[:,0], x_train[:,1], label='x_train')
plt.scatter(x_test_n[:,0], x_test_n[:,1], label='normal')
plt.scatter(x_test_ab[:,0], x_test_ab[:,1], label='abnormal')
plt.legend()
plt.savefig('./images/emb_%s%d%s.pdf' % (type, id, str(epoch))) # , bbox_inches='tight')
plt.clf()
plt.cla()
plt.close()
# bag of frames
k = int(10 // length_s)
x_train = np.reshape(x_train, (k, x_train.shape[0]//k, embedding_dimension)).swapaxes(0, 1).reshape(x_train.shape[0]//k, embedding_dimension*k)
x_test_n = np.reshape(x_test_n, (k, x_test_n.shape[0]//k, embedding_dimension)).swapaxes(0, 1).reshape(x_test_n.shape[0]//k, embedding_dimension*k)
x_test_ab = np.reshape(x_test_ab, (k, x_test_ab.shape[0]//k, embedding_dimension)).swapaxes(0, 1).reshape(x_test_ab.shape[0]//k, embedding_dimension*k)
if embedding_dimension == 2:
for i in range(x_train.shape[1]):
for j in range(x_train.shape[1]):
if not i > j:
plt.scatter(x_test_ab[:, i], x_test_ab[:,j],label='test_abnormal')
plt.scatter(x_test_n[:, i], x_test_n[:,j], label='test_normal')
plt.scatter(x_train[:, i], x_train[:,j], label='train')
plt.legend()
plt.savefig('./images/emb_%s%d%s_%d%d.pdf' % (type, id, str(epoch),i,j),dpi=600) # , bbox_inches='tight')
plt.clf()
plt.cla()
plt.close()
return x_train, x_test_n, x_test_ab
def save_neural_embeddings(x_train, x_test_n, x_test_ab, type, id, snr, epoch, folder = './neural_embeddings/'):
if epoch is None: epoch = ''
if not os.path.exists(folder): create_folder(folder)
np.save(folder+'x_train_%s_%d_%ddB_%d' % (type, id, snr, epoch), x_train)
np.save(folder+'x_test_n_%s_%d_%ddB_%d' % (type, id, snr, epoch), x_test_n)
np.save(folder+'x_test_ab_%s_%d_%sdB_%d' % (type, id, snr, epoch), x_test_ab)
if x_train.shape[-1] == 2:
plt.scatter(x_train[:,0], x_train[:,1], label='x_train')
plt.scatter(x_test_n[:,0], x_test_n[:,1], label='normal')
plt.scatter(x_test_ab[:,0], x_test_ab[:,1], label='abnormal')
plt.legend()
plt.savefig('./images/embed_%s%d%d%s.pdf' % (type, id, snr, str(epoch)), dpi=1000) # , bbox_inches='tight')
plt.clf()
plt.cla()
plt.close()
def load_neural_embeddings(type, id, snr, epoch, folder='./neural_embeddings/'):
if epoch is None: epoch = ''
x_train = np.load(folder+'x_train_%s_%d_%ddB_%d.npy' % (type, id, snr, epoch))
x_test_n = np.load(folder+'x_test_n_%s_%d_%ddB_%d.npy' % (type, id, snr, epoch))
x_test_ab = np.load(folder+'x_test_ab_%s_%d_%ddB_%d.npy' % (type, id, snr, epoch))
return x_train, x_test_n, x_test_ab
def deepOneClass(x_train, x_test_n, x_test_ab):
t0 = time.time()
model = DOC(x_train.shape[1]).cuda()
out_dim = model.out_features
x_train = torch.Tensor(x_train).type(torch.FloatTensor).cuda()
x_test_n = torch.Tensor(x_test_n).type(torch.FloatTensor).cuda()
x_test_ab = torch.Tensor(x_test_ab).type(torch.FloatTensor).cuda()
# Training Loop #
batch_size = 128
n_epochs = 200
val_ratio = 0.10
early_stopping = EarlyStopping(patience=10, path=os.path.join(FOLDER, 'ad_model_checkpoint.pt'), verbose=False)
indices = np.arange(x_train.shape[0])
n = np.int(np.floor(val_ratio * indices.shape[0]))
val_indices = indices[:n]
train_indices = indices[n:]
N = train_indices.shape[0]
svdd_loss = SVDD_loss(out_dim, 1.0).cuda()
criterion = svdd_loss
optimizer4nn = torch.optim.Adam(model.parameters(), lr=model.lr)
optimzer4svdd = torch.optim.SGD(svdd_loss.parameters(), lr=model.lr_svdd)
optimizer = [optimizer4nn, optimzer4svdd]
c_per_epoch = []
for epoch in range(1, n_epochs+1):
# Shuffle Training Indices #
np.random.shuffle(train_indices)
running_loss = 0.0
model.train()
for i in range( N // batch_size):
optimizer[0].zero_grad()
optimizer[1].zero_grad()
X = x_train[train_indices[i * batch_size: (i + 1) * batch_size]]
Y = model(X)
var_reg = 0.1 * Y.var(dim=0).mean()
loss = criterion(torch.tensor(0).repeat(Y.shape[0]).cuda(), Y) + var_reg
running_loss += loss.item()
# Update Model Parameters #
loss.backward()
optimizer[0].step()
optimizer[1].step()
c = svdd_loss.centers
c_per_epoch.append(c)
# Compute Validation Loss #
model.eval()
with torch.no_grad():
Y_val = model(x_train[val_indices])
var_reg = 0.1 * Y_val.var(dim=0).mean()
val_loss = criterion(torch.tensor(0).repeat(Y_val.shape[0]).cuda(), Y_val) + var_reg # + c * val_loss_reg
early_stopping(val_loss, model)
if early_stopping.early_stop:
break
# load the last checkpoint with the best model
model = torch.load(FOLDER + '\\ad_model_checkpoint.pt')
model.eval()
c = c_per_epoch[early_stopping.best_epoch-1]
with torch.no_grad():
Y_train = model(x_train)
Y_pred_n = model(x_test_n)
Y_pred_ab = model(x_test_ab)
anomaly_scores = ((Y_train - c) ** 2).sum(dim=1).cpu().numpy()
anomaly_scores_n = ((Y_pred_n - c) ** 2).sum(dim=1).cpu().numpy()
anomaly_scores_ab = ((Y_pred_ab - c) ** 2).sum(dim=1).cpu().numpy()
y_score = anomaly_scores_n.tolist() + anomaly_scores_ab.tolist()
y_true = [0] * len(anomaly_scores_n) + [1] * len(anomaly_scores_ab)
auc = roc_auc_score(y_true, y_score)
pauc = roc_auc_score(y_true, y_score, max_fpr=0.1)
return auc, pauc
def oneClass(x_train, x_test_n, x_test_ab, epoch=None):
if epoch is None: epoch = ''
t0 = time.time()
svm = OneClassSVM(nu=0.01)
svm.fit(x_train)
y_hat_test_n = svm.score_samples(x_test_n)
y_hat_test_ab = svm.score_samples(x_test_ab)
y_score = y_hat_test_n.tolist() + y_hat_test_ab.tolist()
y_true = [1] * len(y_hat_test_n) + [0] * len(y_hat_test_ab)
auc = roc_auc_score(y_true, y_score)
pauc = roc_auc_score(y_true, y_score, max_fpr=0.1)
fpr, tpr, _ = roc_curve(y_true, y_score)
print('AUC: %.3f - pAUC: %.3f (%s)' % (auc, pauc, 'OCSVM'))
return auc
def anomaly_detection(model, f_test, type, id, snr, length_s=2., channels=None,
epoch=None, batch_size=128):
t0 = time.time()
if os.path.exists('./neural_embeddings/x_train_%s_%d_%ddB_%d.npy' % (type, id, snr, epoch)):
# print('loading embeddings...')
x_train, x_test_n, x_test_ab = load_neural_embeddings(type, id, snr, epoch)
else:
x_train, x_test_n, x_test_ab = predictNeuralEmbeddings(model, batch_size, f_test, length_s=length_s, type=type,
id=id, snr=snr, channels=channels, epoch=epoch)
save_neural_embeddings(x_train, x_test_n, x_test_ab, type, id, snr, epoch)
print("%ds - Anomaly detection:" % int(time.time() - t0), type, id, snr, 'dB -', x_train.shape, x_test_n.shape, x_test_ab.shape)
auc, pauc = deepOneClass(x_train, x_test_n, x_test_ab)
auc_svm = oneClass(x_train, x_test_n, x_test_ab)
return auc, auc_svm
fs = 16000
FOLDER = ''
dataset = MIMII
if __name__ == '__main__':
f_train, f_val, f_test = dataset.get_train_val_test_filepaths(test_ratio=0.25, val_ratio=0.2, shuffle=False, roll=False)
FOLDER = create_folder('models/' + time.strftime("%m-%d") + '_CNN1D')
shutil.copy('main.py', FOLDER + '/main.txt')
shutil.copy('models.py', FOLDER + '/models.txt')
sys.stdout = Logger(FOLDER + '/console.txt')
model = train_neural_embeddings(f_train, f_val, f_test)
total_aucs = np.zeros((2, 4, 4, 3))
for t, type in enumerate(dataset.all_types):
for i, id in enumerate(dataset.get_ids):
for s, snr in enumerate(dataset.all_snrs):
total_aucs[:, t, i, s] = anomaly_detection(model, length_s=2.,
f_test=f_test, type=type, id=id, snr=snr, epoch=12)
np.savetxt(FOLDER + "/aucs_deep.csv", total_aucs[0].reshape(16, 3), delimiter=";")
np.savetxt(FOLDER + "/aucs_ocsvm.csv", total_aucs[1].reshape(16, 3), delimiter=";")