-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
249 lines (216 loc) · 13.4 KB
/
trainer.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
import os
import torch
import numpy as np
import torch.nn as nn
from copy import deepcopy
from torch import tensor
from torch.nn import Module
from torch.autograd import Function
from torch.distributions import Categorical
import torch.nn.functional as F
import torchvision
import torchaudio
from scipy import signal
import matplotlib.pyplot as plt
from data_preprocess import data_preprocess_dalia
from data_preprocess import data_preprocess_ptb
from data_preprocess import data_preprocess_wesad
from data_preprocess import data_preprocess_IEEE_small
from data_preprocess import data_preprocess_IEEE_big
from data_preprocess import data_preprocess_bidmc
from data_preprocess import data_preprocess_capno
from data_preprocess import data_preprocess_clemson
from losses import _IPR_SSL, _SNR_SSL, _EMD_SSL, resample_v2_loss, shiftfreq_loss
def setup_data_loader(args):
if args.dataset == 'dalia':
return data_preprocess_dalia.prep_dalia(args)
elif args.dataset == 'ptb':
return data_preprocess_ptb.prep_ptb(args)
elif args.dataset == 'wesad':
return data_preprocess_wesad.prep_wesad(args)
elif args.dataset == 'ieee_small':
return data_preprocess_IEEE_small.prep_ieee_small(args)
elif args.dataset == 'ieee_big':
return data_preprocess_IEEE_big.prep_ieeebig(args)
elif args.dataset == 'bidmc':
return data_preprocess_bidmc.prep_bidmc(args)
elif args.dataset == 'capno' or args.dataset == 'capno_64':
return data_preprocess_capno.prep_capno(args)
elif args.dataset == 'clemson':
return data_preprocess_clemson.prep_clemson(args)
class Trainer:
def __init__(self, args, model):
self.args = args
self.model = model
self.model_name = 'models' + str(args.model) + args.dataset
if self.args.optimizer == 'adam':
self.optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
#self.optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr)
elif self.args.optimizer == 'sgd':
self.optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.9, weight_decay=args.weight_decay)
self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, mode='min', patience=15, factor=0.5, min_lr=1e-7, verbose=False)
#self.scheduler = torch.optim.lr_scheduler.CyclicLR(self.optimizer, base_lr=args.lr, max_lr=args.lr*10, cycle_momentum=False, step_size_up=100, step_size_down=100, mode='triangular2')
def save_random(self, trainx, shiftedx, epoch, freq, labels):
labels = labels.cpu().detach().numpy()
L, loc0, loc1 = self.args.L, self.args.loc_0, self.args.loc_1
bpms = freq.squeeze() * 60 if not self.args.data_type == 'step' else freq.squeeze() *60*(32/60)
error_rate = np.abs(bpms-labels)
indeces = np.argsort(error_rate)
random_index = indeces[-10:]
if not self.args.past_work1:
epoch_folder = os.path.join('figures', 'epoch_' + str(self.args.dataset) + str(self.args.target_domain) + str(epoch))
else: epoch_folder = os.path.join('figures', 'epoch_' + str(self.args.dataset) + str(self.args.target_domain) + 'pw1' + str(epoch))
os.makedirs(epoch_folder, exist_ok=True)
for i in random_index:
trainx_fft = torch.abs(torch.fft.rfft(trainx[i], n=L))
shifted_fft = torch.abs(torch.fft.rfft(shiftedx[i], n=L))
freq = self.args.fs*(torch.fft.rfftfreq(n=L))
plt.plot(freq[loc0:loc1],trainx_fft[0,loc0:loc1].squeeze().cpu().detach().numpy())
plt.plot(freq[loc0:loc1],shifted_fft[0,loc0:loc1].squeeze().cpu().detach().numpy())
plt.vlines(labels[i]/60, 0, 200, colors='k', linestyles='dashed')
plt.xticks(np.linspace(freq[loc0], freq[loc1], 10))
plt.savefig(os.path.join(epoch_folder, str(i) + '.png'))
plt.cla()
return 0
def evaluate_model(self, valid_loader, DEVICE, after_train=False):
if after_train:
self.model.load_state_dict(torch.load('saved_models' + str(self.args.cuda) + '/' + str(self.args.model) + str(self.args.dataset) + str(self.args.data_type) + str(self.args.seed) +'.pt'))
self.model.eval()
avg_val_mse, avg_freq, avg_loss, total_samples, preds, true = 0., 0, 0, 0, [], []
with torch.no_grad():
for i, val_x in enumerate(valid_loader):
filtered_signal, _ = self.model(val_x[0])
if not self.args.past_work1:
filtered_signal = filtered_signal.unsqueeze(1)
loss, freq = self.freq_losses(filtered_signal, val_x[0], val_x[2])
else:
loss, freq = self.freq_losses_2(filtered_signal, val_x[2])
avg_loss += loss.detach().cpu().item()
avg_freq += torch.sum((freq*60).detach().cpu())
if self.args.data_type == 'ecg' or self.args.data_type == 'ppg' or self.args.data_type == 'resp':
avg_val_mse += (torch.nn.L1Loss(reduction='sum')((freq*60).squeeze(), val_x[1]).item())
elif self.args.data_type == 'step':
avg_val_mse += torch.sum((torch.abs((val_x[1] - freq*60*(32/60))/val_x[1])).squeeze()).item()
else:
eval_points = val_x[3] == 1
if eval_points.any():
avg_val_mse += (torch.nn.L1Loss(reduction='sum')(freq[eval_points]*60, val_x[1][eval_points]).to(DEVICE)).item()
total_samples += val_x[0].size(0)
if self.args.data_type == 'step': preds.append(freq*60*(32/60))
else: preds.append(freq*60)
if after_train: true.append(val_x[1].cpu().detach().numpy())
if after_train:
preds, true = torch.cat(preds).cpu().detach().numpy(), np.concatenate(true)
plt.scatter(preds, true)
plt.savefig('figures' + '/' + 'eval_' + str(self.args.seed) + str(self.args.dataset) + str(self.args.target_domain) + '.png')
plt.cla()
if self.args.data_type == 'step':
print(f'MAPE: {np.mean(np.abs((true-preds)/true))}, MAE: {np.mean(np.abs(preds-true))}')
return np.array([np.mean(np.abs((true-preds)/true)), 0, np.mean(np.abs(preds-true))])
print(f'MSE: {np.mean(np.abs(preds-true))}, RMSE: {np.sqrt(np.mean(np.square(preds-true)))}, r2: {np.corrcoef(preds, true)[0,1]}')
return np.array([np.mean(np.abs(preds-true)), np.sqrt(np.mean(np.square(preds-true))), np.corrcoef(preds, true)[0,1]])
return avg_loss/total_samples
def shift_tensors2(self, back_to_time, sample_shift, device_id):
back_to_time = nn.ConstantPad1d(100,0)(back_to_time)
fft_of_current = torch.fft.rfft(back_to_time, n=back_to_time.size(2))
freq = torch.fft.rfftfreq(n=back_to_time.size(2)).cuda(self.args.cuda)
fft_of_shifted = torch.zeros(fft_of_current.size(), requires_grad=True)
ifft_of_shifted = torch.zeros(back_to_time.size(), requires_grad=True)
fft_of_shifted = torch.exp(-1j*2*torch.pi*freq[None,:]*sample_shift[:,None])[:,None,:]*fft_of_current
ifft_of_shifted = torch.fft.irfft(fft_of_shifted, back_to_time.size(2))
ifft_of_shifted -= ifft_of_shifted.min(0, keepdim=True)[0]
ifft_of_shifted /= ifft_of_shifted.max(0, keepdim=True)[0]
ifft_of_shifted = ifft_of_shifted[:,:,100:-100]
return ifft_of_shifted
def freq_losses(self, filtered_signal, orig_signal, lin_ratio=None):
L, loc_0, loc_1 = self.args.L, self.args.loc_0, self.args.loc_1
x_fft = torch.abs(torch.fft.rfft(filtered_signal, n=L, norm='forward'))
org_fft = torch.abs(torch.fft.rfft(orig_signal, n=L, norm='forward'))
freq = self.args.fs*(torch.fft.rfftfreq(n=L).cuda(self.args.cuda))
if self.args.aug_type == 'resample_v2':
return resample_v2_loss(x_fft, org_fft, freq, loc_0, loc_1)
if self.args.aug_type == 'freq_shift':
return shiftfreq_loss(x_fft, org_fft, lin_ratio, loc_0, loc_1, freq, self.args)
else:
l1 = torch.sum((torch.sum(x_fft[:,:,0:loc_0], dim=2).squeeze() + torch.sum(x_fft[:,:,loc_1:], dim=2).squeeze()), dim=0)
# Entropy
freq_interest = x_fft[:,:,loc_0:loc_1]/torch.sum(x_fft[:,:,loc_0:loc_1], axis=2, keepdim=True)
freq_interest_org = org_fft[:,:,loc_0:loc_1]/torch.sum(org_fft[:,:,loc_0:loc_1], axis=2, keepdim=True)
l2 = torch.sum(-torch.sum(freq_interest*torch.log(freq_interest), dim=2), dim=0)
#########
kl_loss = nn.KLDivLoss(reduction='sum')
l3 = kl_loss(torch.log(freq_interest), freq_interest_org)
#########
peak_locs = torch.argmax(x_fft[:,0, loc_0:loc_1], dim=1)
return l1+l2+l3, freq[peak_locs+loc_0]
def train(self, train_loader, valid_loader, split_seed, device_id):
n_epoch, counter = 999, 0
temp_avg_loss = np.inf
for epoch in range(n_epoch):
self.model.train()
avg_loss, avg_freq, avg_mse, total_sample = 0, 0, 0, 0
for i, train_x in enumerate(train_loader):
filtered_signal, _ = self.model(train_x[0])
if not self.args.past_work1:
filtered_signal = filtered_signal.unsqueeze(1)
loss, freq = self.freq_losses(filtered_signal, train_x[0], train_x[2])
else:
loss, freq = self.freq_losses_2(filtered_signal, train_x[2])
filtered_signal = filtered_signal.unsqueeze(1)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
avg_loss += loss.detach().cpu().item()
avg_freq += torch.sum((freq*60).detach().cpu())
if self.args.data_type == 'ecg' or self.args.data_type == 'ppg':
avg_mse += (torch.nn.L1Loss(reduction='sum')((freq*60).squeeze(), train_x[1]).item())
total_sample += train_x[0].size(0)
if epoch % 40 == 0 and self.args.plot: self.save_random(train_x[0], filtered_signal, epoch, freq.detach().cpu(), train_x[1])
avg_val_mse=self.evaluate_model(valid_loader, device_id)
avg_loss, avg_freq, avg_mse = avg_loss/total_sample, avg_freq/total_sample, avg_mse/total_sample
if avg_loss + 0.001 < temp_avg_loss:
counter = 0
temp_avg_loss = avg_loss
torch.save(deepcopy(self.model.state_dict()), 'saved_models' + str(self.args.cuda) + '/'
+ str(self.args.model) + str(self.args.dataset) + str(self.args.data_type) + str(self.args.seed) +'.pt')
else:
counter += 1
if counter > 30:
return avg_loss
self.scheduler.step(avg_loss)
return avg_loss
###############################
def freq_losses_2(self, filtered_signal, lin_ratio=None):
L, loc_0, loc_1 = self.args.L, self.args.loc_0, self.args.loc_1
x_fft = torch.abs(torch.fft.rfft(filtered_signal, n=L, norm='forward'))
freq = self.args.fs*(torch.fft.rfftfreq(n=L).cuda(self.args.cuda))
delta_freq = 0.1
if self.args.aug_type == 'resample_v2':
l1_loss, l2_loss = torch.ones((x_fft.size(0),1)).cuda(self.args.cuda), torch.ones((x_fft.size(0),1)).cuda(self.args.cuda)
l3_loss, bpms = torch.ones((x_fft.size(0),1)).cuda(self.args.cuda), torch.ones((x_fft.size(0),1)).cuda(self.args.cuda)
for i in range(len(filtered_signal)):
loc0, loc1 = (loc_0*lin_ratio[i]).int(), (loc_1*lin_ratio[i]).int()
l1_loss[i] = _IPR_SSL(freq, x_fft[i,:], low_hz=freq[loc0], high_hz=freq[loc1], device=self.args.cuda)
l2_loss[i] = _SNR_SSL(freq, x_fft[i,:], low_hz=freq[loc0], high_hz=freq[loc1], freq_delta=delta_freq, normalized=False, bandpassed=False, device=self.args.cuda)
l3_loss[i] = _EMD_SSL(freq, x_fft[i,:], low_hz=freq[loc0], high_hz=freq[loc1], device=self.args.cuda)
return torch.sum(l1_loss+l2_loss+l3_loss), bpms
else:
l1 = _IPR_SSL(freq, x_fft, low_hz=freq[loc_0], high_hz=freq[loc_1], device=self.args.cuda)
l2 = _SNR_SSL(freq, x_fft, low_hz=freq[loc_0], high_hz=freq[loc_1], freq_delta=delta_freq, normalized=False, bandpassed=False, device=self.args.cuda)
l3 = _EMD_SSL(freq, x_fft, low_hz=freq[loc_0], high_hz=freq[loc_1], device=self.args.cuda)
peak_locs = torch.argmax(x_fft[:, loc_0:loc_1], dim=1)
return l1+l2+l3, freq[peak_locs+loc_0]
###################################
def assign_fft_params(args):
if args.data_type == 'ppg' or args.data_type == 'step': L = 512
elif args.data_type == 'resp': L = 1024
elif args.data_type == 'ecg': L = 2048
else: L = 2048
freq = args.fs*(torch.fft.rfftfreq(n=L).cuda(args.cuda))
if args.data_type == 'resp': loc_0, loc_1 = torch.where((freq > 4/60))[0][0].item(), torch.where((freq > 40/60))[0][0].item() # 4 to 40 respiration
elif args.data_type == 'step': loc_0, loc_1 = torch.where((freq > 40/60))[0][0].item(), torch.where((freq > 140/60))[0][0].item() # 40 to 140 step --> 20 to 60
else: loc_0, loc_1 = torch.where((freq > 30/60))[0][0].item(), torch.where((freq > 210/60))[0][0].item() # 30 to 210 HR
args.L = L
args.loc_0 = loc_0
args.loc_1 = loc_1
return