-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path_train.py
187 lines (149 loc) · 6.51 KB
/
_train.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
import time
import torch
import numpy as np
from torch import nn
from tqdm import tqdm
from util.train import evaluate as EVALUATE
def process_siam(model, data_loader, optimizer=None, args=None):
model.cuda()
if not optimizer == None:
model.train()
else:
model.eval()
criterion_1 = nn.CosineSimilarity()
criterion_2 = nn.MSELoss(reduction='sum')
criterion_3 = nn.BCELoss(reduction='sum')
total_loss, total_data, total_loss_siam, total_loss_pp, total_loss_fp = 0, 0, 0, 0, 0
start_time = time.time()
result_dict = {}
all_z = []
for batch in data_loader:
l1, l2, s1, s2 = batch['length_1'].long(), batch['length_2'].long(), \
batch['seq_1'].long(), batch['seq_2'].long()
sample = model(s1.cuda(), l1, s2.cuda(), l2)
p1, p2, z1, z2 = sample['p1'], sample['p2'], sample['z1'].detach(), sample['z2'].detach()
if args.use_pp_prediction:
pp1, pp2 = sample['pp1'], sample['pp2']
elif args.use_fp_prediction:
fp1, fp2 = sample['fp1'], sample['fp2']
loss_siam = -(criterion_1(p1, z2).sum() + criterion_1(p2, z1).sum()) * 0.5
loss_siam = loss_siam.cpu()
if args.no_use_siam:
loss_siam = loss_siam - loss_siam
total_loss_siam += loss_siam.item()
if args.use_pp_prediction:
loss_pp = (criterion_2(pp1, batch['pp'].cuda()) + \
criterion_2(pp2, batch['pp'].cuda())) * 0.5
loss_pp = loss_pp.cpu()
loss = loss_siam + loss_pp * args.pp_loss_ratio
total_loss_pp += loss_pp.item()
elif args.use_fp_prediction:
loss_fp = (criterion_3(fp1, batch['fp'].cuda()) + \
criterion_3(fp2, batch['fp'].cuda())) * 0.5
loss_fp = loss_fp.cpu()
loss = loss_siam + loss_fp * args.fp_loss_ratio
total_loss_fp += loss_fp.item()
else:
loss = loss_siam
if not optimizer == None:
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
total_data += z1.size(0)
all_z.append(z1) # add only z1
all_z = torch.cat(all_z, dim=0)
std = calc_latent_std(all_z)
result_dict['std'] = np.mean(std)
result_dict['loss_siam'] = total_loss_siam / total_data
if args.use_pp_prediction:
result_dict['loss_pp'] = total_loss_pp / total_data
else:
result_dict['loss_pp'] = 0
result_dict['time'] = time.time() - start_time
return model, result_dict
def process_clf(model, data_loader, optimizer=None):
model.cuda()
if not optimizer == None:
model.train()
else:
model.eval()
criterion = nn.BCELoss(reduction='sum')
total_loss, total_data = 0, 0
total_pred_p, total_true_label = [], []
start_time = time.time()
result_dict = {}
for batch in data_loader:
l, s = batch['length_1'].long(), batch['seq_1'].long()
true_target = batch['target'].long().cuda()
pred_p= torch.sigmoid(model(s.cuda(), l))
loss = criterion(pred_p, true_target.float())
if not optimizer == None:
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
total_data += pred_p.size(0)
total_pred_p += pred_p.tolist()
total_true_label += true_target.tolist()
total_pred_p = np.array(total_pred_p)
total_true_label = np.array(total_true_label)
total_pred_label = np.round(total_pred_p)
result_dict = {}
result_dict['accuracy'] = EVALUATE.calc_accuracy(total_pred_label, total_true_label)
result_dict['precision'] = EVALUATE.calc_precision(total_pred_label, total_true_label)
result_dict['recall'] = EVALUATE.calc_recall(total_pred_label, total_true_label)
result_dict['auc_roc'] = EVALUATE.calc_roc_auc(total_pred_p, total_true_label)
result_dict['auc_prc'] = EVALUATE.calc_prc_auc(total_pred_p, total_true_label)
result_dict['loss'] = total_loss / total_data
result_dict['time'] = time.time() - start_time
return model, result_dict
def process_clf_validation_smiles_enumerate(model, data_loader, n_trial=16):
# data_loader should not be shuffled !!!
model.cuda()
model.eval()
criterion = nn.BCELoss(reduction='sum')
total_loss, total_data = 0, 0
start_time = time.time()
result_dict = {}
total_pred_p_score = []
for _ in range(n_trial):
enm_pred_p_score, enm_true_label = [], []
for batch in data_loader:
l, s = batch['length_1'].long(), batch['seq_1'].long()
true_target = batch['target'].long().cuda()
pred_p_score = model(s.cuda(), l)
loss = criterion(torch.sigmoid(pred_p_score), true_target.float())
total_loss += loss.item() # float
total_data += pred_p_score.size(0) # int
enm_pred_p_score.append(pred_p_score)
enm_true_label.append(true_target)
enm_pred_p_score = torch.cat(enm_pred_p_score, dim=0)
enm_true_label = torch.cat(enm_true_label, dim=0)
total_pred_p_score.append(enm_pred_p_score)
total_pred_p_score = torch.stack(total_pred_p_score, dim=0) # [n_trail, n_data]
avg_pred_p_score = total_pred_p_score.mean(dim=0) # [n_data]
avg_pred_p = torch.sigmoid(avg_pred_p_score) # apply sigmoid after
total_true_label = enm_true_label
total_true_label = total_true_label.cpu().detach().numpy()
avg_pred_p = avg_pred_p.cpu().detach().numpy()
avg_pred_label = np.round(avg_pred_p)
total_pred_label, total_pred_p, total_true_label = \
avg_pred_label, avg_pred_p, total_true_label
result_dict = {}
result_dict['accuracy'] = EVALUATE.calc_accuracy(total_pred_label, total_true_label)
result_dict['precision'] = EVALUATE.calc_precision(total_pred_label, total_true_label)
result_dict['recall'] = EVALUATE.calc_recall(total_pred_label, total_true_label)
result_dict['auc_roc'] = EVALUATE.calc_roc_auc(total_pred_p, total_true_label)
result_dict['auc_prc'] = EVALUATE.calc_prc_auc(total_pred_p, total_true_label)
result_dict['loss'] = total_loss / total_data
result_dict['time'] = time.time() - start_time
return model, result_dict
def calc_latent_std(x): # x[bs hd]
x = x.cpu().detach().numpy()
std = np.std(x, axis=1)
return std
if __name__ == '__main__':
x = torch.rand(16, 1000)
y = calc_latent_std(x)
print(y)