-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
271 lines (232 loc) · 11.8 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
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
import os
import torch
import torch.nn as nn
from itertools import chain
from torch.autograd import Variable
from torch.utils.data import ConcatDataset, DataLoader, RandomSampler
from torchvision import datasets, transforms
from wilds.common.data_loaders import get_train_loader,get_eval_loader
from models.hydranet import HydraNet
from dummy_datasets import get_dummy
from utils import get_inf_iterator, save_model, get_inf_iterator, print_and_log
from evaluate import evaluate
from torch.utils.tensorboard import SummaryWriter
from datetime import datetime
import numpy as np
def pseudo_label(net, device, target_dataset, batch_size, threshold=0.9):
assert net.num_heads>1, 'Number of pseudo-heads must be 2 or more'
net.eval()
device_loc = device#"cpu"
target_dataloader = DataLoader(target_dataset,batch_size=batch_size,
shuffle=False)
excerpt_final = torch.IntTensor().to(device_loc)
pseudo_labels_final = torch.IntTensor().to(device_loc)
for batch_counter, (img,_,_) in enumerate(target_dataloader):
_, p_outs = net(img.to(device)) # num_heads x batch_size x num_classes
p_outs = [p_out.data.to(device_loc) for p_out in p_outs]
_, pred_0 = torch.max(p_outs[0], 1)
mask_common = torch.ones(p_outs[0].shape[0]).to(device_loc) # batch_size,
for i in range(1,net.num_heads):
_, pred = torch.max(p_outs[i],1)
mask_common = mask_common * torch.eq(pred_0,pred)
equal_idx = torch.nonzero(mask_common)
max_conf, _ = torch.max(p_outs[0],1)
for i in range(1,net.num_heads):
conf,_ = torch.max(p_outs[i],1)
max_conf,_ = torch.max(torch.stack([max_conf, conf], 1), 1)
mask_conf = max_conf > threshold
mask_final = max_conf*mask_common
excerpt = torch.nonzero(mask_final).squeeze(-1)
_, pseudo_labels = torch.max(p_outs[0][excerpt, :], 1)
excerpt_final = torch.cat([excerpt_final, excerpt+batch_counter*batch_size])
pseudo_labels_final = torch.cat([pseudo_labels_final, pseudo_labels])
if excerpt_final.shape[0] == 0:
is_empty = True
else:
is_empty = False
excerpt_final = excerpt_final.cpu()
pseudo_labels_final = pseudo_labels_final.cpu()
target_dataset_labelled = get_dummy(target_dataset, excerpt_final,
pseudo_labels_final, need_dataset=True)
return target_dataset_labelled, is_empty
def source_train_bootstrap(net,device, train_dataset,val_dataset,batch_size,num_epochs,
model_dir,log_file,epoch_offset=0,threshold=0.9):
criterion = nn.CrossEntropyLoss()
optimizer_enc = torch.optim.Adam(net.enc.parameters(), lr=1e-4)
scheduler_enc = torch.optim.lr_scheduler.ExponentialLR(optimizer_enc, gamma=0.96)
optimizer_tHead = torch.optim.Adam(net.tHead.parameters(), lr=1e-3)
scheduler_tHead = torch.optim.lr_scheduler.ExponentialLR(optimizer_tHead, gamma=0.96)
optimizer_pHeads_arr = []
scheduler_pHeads_arr = []
if net.num_heads > 0:
optimizer_pHeads_arr = [torch.optim.Adam(pHead.parameters(), lr=net.num_heads*1e-3)
for pHead in net.pHeads.heads]
scheduler_pHeads_arr = [torch.optim.lr_scheduler.ExponentialLR(optimizer_pHead, gamma=0.96)
for optimizer_pHead in optimizer_pHeads_arr]
writer = SummaryWriter()
for epoch in range(epoch_offset,num_epochs):
sampler = RandomSampler(train_dataset, replacement=True, num_samples=len(train_dataset))
train_loader = DataLoader(train_dataset,batch_size=batch_size,
sampler=sampler, drop_last=True)
net.train()
Loss_T = 0.0
Loss_P = 0.0
for img, label, _ in train_loader:
img = Variable(img.to(device))
label = Variable(label.to(device))
optimizer_enc.zero_grad()
optimizer_tHead.zero_grad()
t_out, p_outs = net(img)
loss_t = criterion(t_out, label)
loss = loss_t
Loss_T +=loss_t
rand_phead = int(np.random.randint(net.num_heads))
if net.num_heads>0:
optimizer_pHeads_arr[rand_phead].zero_grad()
loss_p = criterion(p_outs[rand_phead], label)
loss += loss_p
Loss_P += loss_p
loss.backward()
optimizer_enc.step()
optimizer_tHead.step()
if net.num_heads>0:
optimizer_pHeads_arr[rand_phead].step()
# else:
# print("train_T_Loss={:.7f}".format(loss_t))
# print(type(Loss_T))
scheduler_enc.step()
scheduler_tHead.step()
for scheduler_pHead in scheduler_pHeads_arr:
scheduler_pHead.step()
if net.num_heads>0:
Loss_P = Loss_P/len(train_loader)
Loss_T = Loss_T/len(train_loader)
save_model(net,os.path.join(model_dir,f"source_trained_epoch_{epoch+1}.pt"))
print_and_log(message="Epoch {}/{}: train_T_Loss={:.7f}, train_P_Loss={:.7f}".format(
epoch+1,num_epochs,Loss_T,Loss_P),
log_file=log_file)
writer.add_scalar("Loss/train", Loss_T.item(), epoch)
if (epoch+1)%5 ==0:
target_loss, target_accs, target_cerr, target_pHead_stats = evaluate(net,device,val_dataset,batch_size,threshold)
com_corr_high, com_corr, com_inc, com_inc_high, disag, p_cerr = target_pHead_stats
print_and_log(message="Tgt_Loss={:.7f}, Tgt_Acc={:.7f}, Tgt_Cal Error={:.7f}".format(
target_loss, target_accs[0], target_cerr),log_file=log_file)
print_and_log(message=f"Accuracies of heads = {target_accs[1]}",log_file=log_file)
print_and_log(message="com_corr_high={:.7f}, com_corr={:.7f}, com_inc={:.7f}, com_inc_high={:.7f}, disag={:.7f}, P_Cal Error={:.7f}".format(
com_corr_high, com_corr,com_inc,com_inc_high,disag,p_cerr),log_file=log_file)
def source_train(net,device, train_dataset,val_dataset,batch_size,num_epochs,
model_dir,log_file,epoch_offset=0,threshold=0.9):
criterion = nn.CrossEntropyLoss()
optimizer_enc = torch.optim.Adam(net.enc.parameters(), lr=1e-4)
scheduler_enc = torch.optim.lr_scheduler.ExponentialLR(optimizer_enc, gamma=0.96)
optimizer_tHead = torch.optim.Adam(net.tHead.parameters(), lr=1e-3)
scheduler_tHead = torch.optim.lr_scheduler.ExponentialLR(optimizer_tHead, gamma=0.96)
if net.num_heads > 0:
optimizer_pHeads = torch.optim.Adam(net.pHeads.parameters(), lr=1e-3)
scheduler_pHeads = torch.optim.lr_scheduler.ExponentialLR(optimizer_pHeads, gamma=0.96)
train_loader = DataLoader(train_dataset,batch_size=batch_size,
shuffle=True, drop_last=True)
writer = SummaryWriter()
# train with source samples
for epoch in range(epoch_offset,num_epochs):
net.train()
Loss_T = 0.0
Loss_P = 0.0
for img, label, _ in train_loader:
img = Variable(img.to(device))
label = Variable(label.to(device))
optimizer_enc.zero_grad()
optimizer_tHead.zero_grad()
t_out, p_outs = net(img)
loss_t = criterion(t_out, label)
loss = loss_t
Loss_T +=loss_t
if net.num_heads>0:
optimizer_pHeads.zero_grad()
loss_p = [criterion(p_out, label) for p_out in p_outs]
avg_loss_p = sum(loss_p)/len(loss_p)
loss += avg_loss_p
Loss_P += avg_loss_p
loss.backward()
optimizer_enc.step()
optimizer_tHead.step()
if net.num_heads>0:
optimizer_pHeads.step()
scheduler_enc.step()
scheduler_tHead.step()
if net.num_heads>0:
scheduler_pHeads.step()
Loss_P = Loss_P/len(train_loader)
Loss_T = Loss_T/len(train_loader)
save_model(net,os.path.join(model_dir,f"source_trained_epoch_{epoch+1}.pt"))
print_and_log(message="Epoch {}/{}: train_T_Loss={:.7f}, train_P_Loss={:.7f}".format(
epoch+1,num_epochs,Loss_T,Loss_P),
log_file=log_file)
writer.add_scalar("Loss/train", Loss_T.item(), epoch)
if(epoch+1)%5 ==0:
target_loss, target_accs, target_cerr, target_pHead_stats = evaluate(net,device,val_dataset,batch_size,threshold)
com_corr_high, com_corr, com_inc, com_inc_high, disag, p_cerr = target_pHead_stats
print_and_log(message="Tgt_Loss={:.7f}, Tgt_Acc={:.7f}, Tgt_Cal Error={:.7f}".format(
target_loss, target_accs[0], target_cerr),log_file=log_file)
print_and_log(message=f"Accuracies of heads = {target_accs[1]}",log_file=log_file)
print_and_log(message="com_corr_high={:.7f}, com_corr={:.7f}, com_inc={:.7f}, com_inc_high={:.7f}, disag={:.7f}, P_Cal Error={:.7f}".format(
com_corr_high, com_corr,com_inc,com_inc_high,disag,p_cerr),log_file=log_file)
def domain_adapt(net, device, source_dataset, target_dataset,
batch_size, k_step, num_adapt_epochs, threshold, model_dir,log_file):
if net.num_heads == 0:
return
# get pseudolabels
target_dataset_labelled, is_empty = pseudo_label(net, device, target_dataset, batch_size, threshold)
if is_empty:
print("No pseudo labelled points")
return
merged_dataset = ConcatDataset([source_dataset, target_dataset_labelled])
criterion = nn.CrossEntropyLoss()
optimizer_enc = torch.optim.Adam(net.enc.parameters(), lr=1e-4*(0.96)**30)
scheduler_enc = torch.optim.lr_scheduler.ExponentialLR(optimizer_enc, gamma=0.96)
optimizer_tHead = torch.optim.Adam(net.tHead.parameters(), lr=1e-3*(0.96)**30)
scheduler_tHead = torch.optim.lr_scheduler.ExponentialLR(optimizer_tHead, gamma=0.96)
optimizer_pHeads = torch.optim.Adam(net.pHeads.parameters(), lr=1e-3*(0.96)**30)
scheduler_pHeads = torch.optim.lr_scheduler.ExponentialLR(optimizer_pHeads, gamma=0.96)
net.train()
merged_dataloader = DataLoader(merged_dataset,batch_size=batch_size,
shuffle=True, drop_last=True)
target_dataloader_labelled = get_inf_iterator(DataLoader(target_dataset_labelled,
batch_size=batch_size, shuffle=True, drop_last=True))
for epoch in range(num_adapt_epochs):
Loss_P = 0.0
Loss_T = 0.0
for img, label, _ in merged_dataloader:
pseudo_img, pseudo_target_labels, _ = next(target_dataloader_labelled)
img = Variable(img.to(device))
label = Variable(label.to(device))
pseudo_img = Variable(pseudo_img.to(device))
pseudo_target_labels = Variable(pseudo_target_labels.to(device))
optimizer_enc.zero_grad()
optimizer_pHeads.zero_grad()
# Train F, F_heads with merged dataset
t_out, p_outs = net(img)
loss_p = [criterion(p_out, label) for p_out in p_outs]
loss = sum(loss_p)/len(loss_p)
loss.backward()
Loss_P +=loss
optimizer_enc.step()
optimizer_pHeads.step()
# Train F, F_t with pseudo-labelled target data
optimizer_enc.zero_grad()
optimizer_tHead.zero_grad()
t_out, _ = net(pseudo_img)
loss_t = criterion(t_out, pseudo_target_labels)
loss_t.backward()
optimizer_enc.step()
optimizer_tHead.step()
Loss_T+=loss_t
scheduler_enc.step()
scheduler_tHead.step()
scheduler_pHeads.step()
Loss_P = Loss_P/len(merged_dataloader)
Loss_T = Loss_T/len(merged_dataloader)
save_model(net,os.path.join(model_dir,f"domain_adapt_step_{k_step}_epoch_{epoch+1}.pt"))
print_and_log(message="Domain Adapt Step {} Epoch {}/{}: Target Loss={:.7f}, Pseudo Loss={:.7f}".format(
k_step, epoch+1,num_adapt_epochs,Loss_T, Loss_P),
log_file=log_file)