-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_test.py
219 lines (199 loc) · 10.1 KB
/
train_test.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
import torch
from utils import *
import wandb
from configs import dataset_config,model_config
import math
from plot import *
import matplotlib.pyplot as plt
def train_eval(model,dataloader,fold,epoch,args,optimizer,scheduler,logger,train=False):
assert not model or dataloader or optimizer or scheduler!= None
# if epoch==6:
# print(epoch)
if args.dataset=='Causalogue':
if dataset_config.large==False:
test_sample=math.ceil(200/args.batch_size)
valid_sample=math.ceil(100/args.batch_size)
else:
test_sample=math.ceil(600/args.batch_size)
valid_sample=math.ceil(300/args.batch_size)
if args.dataset=='Causaction':
test_sample=math.ceil(218/args.batch_size)
valid_sample=math.ceil(100/args.batch_size)
if train:
model.train()
logger.info('########################Training######################')
# dataloader = tqdm(dataloader)
else:
model.eval()
logger.info('########################Evaling#######################')
trainstep=0
evalstep=0
teststep=0
s_AUROC_all_train=0
s_AUROC_all_eval=0
s_AUROC_all_test=0
g_AUROC_all_train=0
g_AUROC_all_eval=0
g_AUROC_all_test=0
in_AUROC_all_train=0
in_AUROC_all_eval=0
in_AUROC_all_test=0
s_MSE_all_train=0
s_MSE_all_eval=0
s_MSE_all_test=0
g_MSE_all_train=0
g_MSE_all_eval=0
g_MSE_all_test=0
in_MSE_all_train=0
in_MSE_all_eval=0
in_MSE_all_test=0
distance_all_train=0
distance_all_eval=0
distance_all_test=0
for train_step, batch in enumerate(dataloader, 1):
if args.dataset=='Causalogue':
batch_ids,batch_doc_len,batch_doc_speaker,batch_label,batch_label_mask,batch_utterances, batch_utterances_mask,batch_adj_mask,\
bert_token_b, bert_segment_b, bert_masks_b, bert_clause_b=batch
if args.dataset=='Causaction':
batch_ids,batch_doc_len,batch_label,batch_label_mask,batch_cls,batch_seg_id,batch_action,batch_adj_mask=batch
if train and len(batch_ids)!=args.batch_size:
continue
if train:
trainstep+=1
else:
#print(len(dataloader))
if len(dataloader)==test_sample:
teststep+=1
else:
evalstep+=1
if train:
if args.dataset=='Causalogue':
X_hat,X,A,e,s,rank,pred_results,pred_results_input,confounding=model(batch_doc_len,batch_adj_mask,bert_token_b,bert_masks_b,bert_clause_b)
if args.dataset=='Causaction':
X_hat,X,A,e,s,rank,pred_results,pred_results_input,confounding=model(batch_doc_len,batch_adj_mask,batch_cls,None,None)
else:
with torch.no_grad():
if args.dataset=='Causalogue':
X_hat,X,A,e,s,rank,pred_results,pred_results_input,confounding=model(batch_doc_len,batch_adj_mask,bert_token_b,bert_masks_b,bert_clause_b)
if args.dataset=='Causaction':
X_hat,X,A,e,s,rank,pred_results,pred_results_input,confounding=model(batch_doc_len,batch_adj_mask,batch_cls,None,None)
# if epoch==40:
# for i in range(len(batch_doc_len)):
# #if batch_doc_len[i]==7 or batch_doc_len[i]==8:
# img,im=plot_cka_matrix(causal_graph[i],batch_doc_len[i])
# texts = annotate_heatmap(im, valfmt="{x:.2f}")
# img.savefig('savefig/sslmodel_{}/{}sslmodel_{}.jpg'.format(args.high_level_loss,batch_ids[i],args.high_level_loss))
# plt.show()
# loc=torch.where(causal_strengh!=causal_strengh)
# causal_strengh[loc]=0
loss_KL=model.loss_KL(e,s)
loss_reconsctruction=model.loss_reconstruction(X_hat,X,confounding,rank,batch_label_mask)
loss_high_level=model.loss_hl(pred_results,A,batch_label,batch_label_mask)
#loss_ss=model.loss_ss(H_do,correlation_label,batch_label_mask)
#loss_KL=1
#loss_re=1
# loss_high_level=1
# loss_KL=model.loss_KL(e,s)
# loss_re=model.loss_reconstruction(X_rec,X)
loss = loss_high_level + loss_KL + loss_reconsctruction
if train:
logger.info('TRAIN# fold: {}, epoch: {}, iter: {}, loss_high_level: {}, loss_KL: {}, loss_re:{}'. \
format(fold, epoch, trainstep, loss_high_level, loss_KL, loss_reconsctruction))
wandb.log({'epoch': epoch, 'trainstep':trainstep+len(dataloader)*epoch,'loss_all_train':loss,'loss_high_level_train':loss_high_level,'loss_KL_train':loss_KL,'loss_re_train':loss_reconsctruction})
loss = loss / args.gradient_accumulation_steps
loss.backward()
if args.dataset=='Causalogue':
torch.nn.utils.clip_grad_norm_(model.parameters(),max_norm=10,norm_type=2)
if args.dataset=='Causaction':
torch.nn.utils.clip_grad_norm_(model.parameters(),max_norm=1,norm_type=2)
if train_step % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step()
model.zero_grad()
s_AUROC,g_AUROC,input_AUROC=auroc(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_MSE,g_MSE,input_MSE=mse(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_AUROC_all_train+=s_AUROC
g_AUROC_all_train+=g_AUROC
in_AUROC_all_train+=input_AUROC
s_AUROC=0
g_AUROC=0
input_AUROC=0
s_MSE_all_train+=s_MSE
g_MSE_all_train+=g_MSE
in_MSE_all_train+=input_MSE
s_MSE=0
g_MSE=0
input_MSE=0
# distance=consistency(X_rec,causal_graph,batch_label_mask)
# distance_all_train+=distance
# distance=0
else:
if len(dataloader)==test_sample:
logger.info('TEST# fold: {}, epoch: {}, iter: {}, loss_high_level: {}, loss_KL: {}, loss_re:{}'. \
format(fold, epoch, train_step, loss_high_level, loss_KL, loss_reconsctruction))
wandb.log({'epoch': epoch, 'teststep':teststep+len(dataloader)*epoch,'loss_all_test':loss,'loss_high_level_test':loss_high_level,'loss_KL_test':loss_KL,'loss_re_test':loss_reconsctruction})
s_AUROC,g_AUROC,input_AUROC=auroc(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_MSE,g_MSE,input_MSE=mse(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_AUROC_all_test+=s_AUROC
g_AUROC_all_test+=g_AUROC
in_AUROC_all_test+=input_AUROC
s_AUROC=0
g_AUROC=0
input_AUROC=0
s_MSE_all_test+=s_MSE
g_MSE_all_test+=g_MSE
in_MSE_all_test+=input_MSE
s_MSE=0
g_MSE=0
input_MSE=0
# distance=consistency(X_rec,causal_graph,batch_label_mask)
# distance_all_test+=distance
# distance=0
else:
logger.info('VALID# fold: {}, epoch: {}, iter: {}, loss_high_level: {}, loss_KL: {}, loss_re:{}'. \
format(fold, epoch, train_step, loss_high_level, loss_KL, loss_reconsctruction))
wandb.log({'epoch': epoch, 'evalstep':evalstep+len(dataloader)*epoch,'loss_all_valid':loss,'loss_high_level_valid':loss_high_level,'loss_KL_valid':loss_KL,'loss_re_valid':loss_reconsctruction})
s_AUROC,g_AUROC,input_AUROC=auroc(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_MSE,g_MSE,input_MSE=mse(pred_results,A,pred_results_input,batch_label,batch_label_mask)
s_AUROC_all_eval+=s_AUROC
g_AUROC_all_eval+=g_AUROC
in_AUROC_all_eval+=input_AUROC
s_AUROC=0
g_AUROC=0
input_AUROC=0
s_MSE_all_eval+=s_MSE
g_MSE_all_eval+=g_MSE
in_MSE_all_eval+=input_MSE
s_MSE=0
g_MSE=0
input_MSE=0
# distance=consistency(X_rec,causal_graph,batch_label_mask)
# distance_all_eval+=distance
# distance=0
if train:
s_AUROC_all_train=s_AUROC_all_train/trainstep
g_AUROC_all_train=g_AUROC_all_train/trainstep
in_AUROC_all_train=in_AUROC_all_train/trainstep
s_MSE_all_train=s_MSE_all_train/trainstep
g_MSE_all_train=g_MSE_all_train/trainstep
in_MSE_all_train=in_MSE_all_train/trainstep
distance_all_train=distance_all_train/trainstep
return s_AUROC_all_train,g_AUROC_all_train,in_AUROC_all_train,s_MSE_all_train,g_MSE_all_train,in_MSE_all_train,distance_all_train
elif len(dataloader)==valid_sample:
s_AUROC_all_eval=s_AUROC_all_eval/evalstep
g_AUROC_all_eval=g_AUROC_all_eval/evalstep
in_AUROC_all_eval=in_AUROC_all_eval/evalstep
s_MSE_all_eval=s_MSE_all_eval/evalstep
g_MSE_all_eval=g_MSE_all_eval/evalstep
in_MSE_all_eval=in_MSE_all_eval/evalstep
distance_all_eval=distance_all_eval/evalstep
return s_AUROC_all_eval,g_AUROC_all_eval,in_AUROC_all_eval,s_MSE_all_eval,g_MSE_all_eval,in_MSE_all_eval,distance_all_eval
else:
s_AUROC_all_test=s_AUROC_all_test/teststep
g_AUROC_all_test=g_AUROC_all_test/teststep
in_AUROC_all_test=in_AUROC_all_test/teststep
s_MSE_all_test=s_MSE_all_test/teststep
g_MSE_all_test=g_MSE_all_test/teststep
in_MSE_all_test=in_MSE_all_test/teststep
distance_all_test=distance_all_test/teststep
return s_AUROC_all_test,g_AUROC_all_test,in_AUROC_all_test,s_MSE_all_test,g_MSE_all_test,in_MSE_all_test,distance_all_test