-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathmain_batch_ihdp.py
267 lines (212 loc) · 8.84 KB
/
main_batch_ihdp.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
import torch
import math
import numpy as np
from models.dynamic_net import Vcnet, Drnet, TR
from data.data import get_iter
from utils.eval import curve
import os
import json
import argparse
def adjust_learning_rate(optimizer, init_lr, epoch):
if lr_type == 'cos': # cos without warm-up
lr = 0.5 * init_lr * (1 + math.cos(math.pi * epoch / num_epoch))
elif lr_type == 'exp':
step = 1
decay = 0.96
lr = init_lr * (decay ** (epoch // step))
elif lr_type == 'fixed':
lr = init_lr
else:
raise NotImplementedError
for param_group in optimizer.param_groups:
param_group['lr'] = lr
return lr
def save_checkpoint(state, model_name='', checkpoint_dir='.'):
filename = os.path.join(checkpoint_dir, model_name + '_ckpt.pth.tar')
print('=> Saving checkpoint to {}'.format(filename))
torch.save(state, filename)
# criterion
def criterion(out, y, alpha=0.5, epsilon=1e-6):
return ((out[1].squeeze() - y.squeeze()) ** 2).mean() - alpha * torch.log(out[0] + epsilon).mean()
def criterion_TR(out, trg, y, beta=1., epsilon=1e-6):
# out[1] is Q
# out[0] is g
return beta * ((y.squeeze() - trg.squeeze() / (out[0].squeeze() + epsilon) - out[1].squeeze()) ** 2).mean()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='train with ihdp data')
# i/o
parser.add_argument('--data_dir', type=str, default='dataset/ihdp', help='dir of data matrix')
parser.add_argument('--data_split_dir', type=str, default='dataset/ihdp/eval', help='dir of data split')
parser.add_argument('--save_dir', type=str, default='logs/ihdp/eval', help='dir to save result')
# common
parser.add_argument('--num_dataset', type=int, default=100, help='num of datasets to train')
# training
parser.add_argument('--n_epochs', type=int, default=800, help='num of epochs to train')
# print train info
parser.add_argument('--verbose', type=int, default=100, help='print train info freq')
args = parser.parse_args()
seed = 0
torch.manual_seed(seed)
np.random.seed(seed)
# Parameters
# optimizer
lr_type = 'fixed'
wd = 5e-3
momentum = 0.9
# targeted regularization optimizer
# original
tr_wd = 1e-3
# epoch: 800!
num_epoch = args.n_epochs
# check val loss
verbose = args.verbose
# load
load_path = args.data_split_dir
num_dataset = args.num_dataset
# save
save_path = args.save_dir
if not os.path.exists(save_path):
os.makedirs(save_path)
data_matrix = torch.load(args.data_dir + '/data_matrix.pt')
t_grid_all = torch.load(args.data_dir + '/t_grid.pt')
Result = {}
for model_name in ['Vcnet', 'Vcnet_tr', 'Tarnet', 'Tarnet_tr', 'Drnet', 'Drnet_tr']:
Result[model_name]=[]
# import model
if model_name == 'Vcnet' or model_name == 'Vcnet_tr':
cfg_density = [(25, 50, 1, 'relu'), (50, 50, 1, 'relu')]
num_grid = 10
cfg = [(50, 50, 1, 'relu'), (50, 1, 1, 'id')]
degree = 2
knots = [0.33, 0.66]
model = Vcnet(cfg_density, num_grid, cfg, degree, knots)
model._initialize_weights()
elif model_name == 'Drnet' or model_name == 'Drnet_tr':
cfg_density = [(25, 50, 1, 'relu'), (50, 50, 1, 'relu')]
num_grid = 10
cfg = [(50, 50, 1, 'relu'), (50, 1, 1, 'id')]
isenhance = 1
model = Drnet(cfg_density, num_grid, cfg, isenhance=isenhance)
model._initialize_weights()
elif model_name == 'Tarnet' or model_name == 'Tarnet_tr':
cfg_density = [(25, 50, 1, 'relu'), (50, 50, 1, 'relu')]
num_grid = 10
cfg = [(50, 50, 1, 'relu'), (50, 1, 1, 'id')]
isenhance = 0
model = Drnet(cfg_density, num_grid, cfg, isenhance=isenhance)
model._initialize_weights()
# use Target Regularization?
if model_name == 'Vcnet_tr' or model_name == 'Drnet_tr' or model_name == 'Tarnet_tr':
isTargetReg = 1
else:
isTargetReg = 0
tr_knots = list(np.arange(0.05, 1, 0.05))
tr_degree = 2
TargetReg = TR(tr_degree, tr_knots)
TargetReg._initialize_weights()
# best cfg for each model
if model_name == 'Tarnet':
init_lr = 0.02
alpha = 1.0
tr_init_lr = 0.001
beta = 1.
Result['Tarnet'] = []
elif model_name == 'Tarnet_tr':
init_lr = 0.02
alpha = 0.5
tr_init_lr = 0.001
beta = 1.
Result['Tarnet_tr'] = []
elif model_name == 'Drnet':
init_lr = 0.02
alpha = 1.
tr_init_lr = 0.001
beta = 1.
Result['Drnet'] = []
elif model_name == 'Drnet_tr':
init_lr = 0.02
alpha = 0.5
tr_init_lr = 0.001
beta = 1.
Result['Drnet_tr'] = []
elif model_name == 'Vcnet':
init_lr = 0.001
alpha = 0.5
tr_init_lr = 0.001
beta = 1.
Result['Vcnet'] = []
elif model_name == 'Vcnet_tr':
# v2
init_lr = 0.001
alpha = 0.5
tr_init_lr = 0.001
beta = 1.
Result['Vcnet_tr'] = []
for _ in range(num_dataset):
cur_save_path = save_path + '/' + str(_)
if not os.path.exists(cur_save_path):
os.makedirs(cur_save_path)
idx_train = torch.load('dataset/ihdp/eval/' + str(_) + '/idx_train.pt')
idx_test = torch.load('dataset/ihdp/eval/' + str(_) + '/idx_test.pt')
train_matrix = data_matrix[idx_train, :]
test_matrix = data_matrix[idx_test, :]
t_grid = t_grid_all[:, idx_test]
# train_matrix, test_matrix, t_grid = simu_data1(500, 200)
train_loader = get_iter(data_matrix[idx_train, :], batch_size=471, shuffle=True)
test_loader = get_iter(data_matrix[idx_test, :], batch_size=data_matrix[idx_test, :].shape[0], shuffle=False)
# reinitialize model
model._initialize_weights()
# define optimizer
optimizer = torch.optim.SGD(model.parameters(), lr=init_lr, momentum=momentum, weight_decay=wd,
nesterov=True)
if isTargetReg:
TargetReg._initialize_weights()
tr_optimizer = torch.optim.SGD(TargetReg.parameters(), lr=tr_init_lr, weight_decay=tr_wd)
print('model : ', model_name)
for epoch in range(num_epoch):
for idx, (inputs, y) in enumerate(train_loader):
t = inputs[:, 0]
x = inputs[:, 1:]
if isTargetReg:
optimizer.zero_grad()
out = model.forward(t, x)
trg = TargetReg(t)
loss = criterion(out, y, alpha=alpha) + criterion_TR(out, trg, y, beta=beta)
loss.backward()
optimizer.step()
tr_optimizer.zero_grad()
out = model.forward(t, x)
trg = TargetReg(t)
tr_loss = criterion_TR(out, trg, y, beta=beta)
tr_loss.backward()
tr_optimizer.step()
else:
optimizer.zero_grad()
out = model.forward(t, x)
loss = criterion(out, y, alpha=alpha)
loss.backward()
optimizer.step()
if epoch % verbose == 0:
print('current epoch: ', epoch)
print('loss: ', loss)
if isTargetReg:
t_grid_hat, mse = curve(model, test_matrix, t_grid, targetreg=TargetReg)
mse = float(mse)
print('current loss: ', float(loss.data))
print('current test loss: ', mse)
else:
t_grid_hat, mse = curve(model, test_matrix, t_grid)
mse = float(mse)
print('current loss: ', float(loss.data))
print('current test loss: ', mse)
print('-----------------------------------------------------------------')
save_checkpoint({
'model': model_name,
'best_test_loss': mse,
'model_state_dict': model.state_dict(),
'TR_state_dict': TargetReg.state_dict(),
}, model_name=model_name, checkpoint_dir=cur_save_path)
print('-----------------------------------------------------------------')
Result[model_name].append(mse)
with open(save_path + '/result.json', 'w') as fp:
json.dump(Result, fp)