-
Notifications
You must be signed in to change notification settings - Fork 1
/
run.py
203 lines (164 loc) · 7.81 KB
/
run.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
import os
import logging
from tqdm import tqdm
from munch import Munch, munchify
import torch
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from torch_geometric.loader import DataLoader
import numpy as np
from GOOD import register
from GOOD.utils.config_reader import load_config
from GOOD.utils.metric import Metric
from GOOD.data.dataset_manager import read_meta_info
from GOOD.utils.evaluation import eval_data_preprocess, eval_score
from GOOD.utils.train import nan2zero_get_mask
from args_parse import args_parser
from exputils import initialize_exp, set_seed, get_dump_path, describe_model, save_model, load_model
from models import MyModel
from dataset import DrugOODDataset
logger = logging.getLogger()
def set_requires_grad(nets, requires_grad=False):
"""Set requies_grad=Fasle for all the networks to avoid unnecessary computations
Parameters:
nets (network list) -- a list of networks
requires_grad (bool) -- whether the networks require gradients or not
"""
if not isinstance(nets, list):
nets = [nets]
for net in nets:
if net is not None:
for param in net.parameters():
param.requires_grad = requires_grad
class Runner:
def __init__(self, args, writer, logger_path):
self.args = args
self.device = torch.device(f'cuda')
if args.dataset.startswith('GOOD'):
# for GOOD, load Config
cfg_path = os.path.join(args.config_path, args.dataset, args.domain, args.shift, 'base.yaml')
cfg, _, _ = load_config(path=cfg_path)
cfg = munchify(cfg)
cfg.device = self.device
dataset, meta_info = register.datasets[cfg.dataset.dataset_name].load(dataset_root=args.data_root,
domain=cfg.dataset.domain,
shift=cfg.dataset.shift_type,
generate=cfg.dataset.generate)
read_meta_info(meta_info, cfg)
# cfg.dropout
# cfg.bs
# update dropout & bs
cfg.model.dropout_rate = args.dropout
cfg.train.train_bs = args.bs
cfg.random_seed = args.random_seed
loader = register.dataloader[cfg.dataset.dataloader_name].setup(dataset, cfg)
self.train_loader = loader['train']
self.valid_loader = loader['val']
self.test_loader = loader['test']
self.metric = Metric()
self.metric.set_score_func(dataset['metric'] if type(dataset) is dict else getattr(dataset, 'metric'))
self.metric.set_loss_func(dataset['task'] if type(dataset) is dict else getattr(dataset, 'task'))
cfg.metric = self.metric
else:
# DrugOOD
dataset = DrugOODDataset(name=args.dataset, root=args.data_root)
self.train_set = dataset[dataset.train_index]
self.valid_set = dataset[dataset.valid_index]
self.test_set = dataset[dataset.test_index]
self.train_loader = DataLoader(self.train_set, batch_size=args.bs, shuffle=True, drop_last=True)
self.valid_loader = DataLoader(self.valid_set, batch_size=args.bs, shuffle=False)
self.test_loader = DataLoader(self.test_set, batch_size=args.bs, shuffle=False)
self.metric = Metric()
self.metric.set_loss_func(task_name='Binary classification')
self.metric.set_score_func(metric_name='ROC-AUC')
cfg = Munch()
cfg.metric = self.metric
cfg.model = Munch()
cfg.model.model_level = 'graph'
self.model = MyModel(args=args, config=cfg).to(self.device)
self.opt = torch.optim.Adam(self.model.parameters(), lr=args.lr)
self.total_step = 0
self.writer = writer
describe_model(self.model, path=logger_path)
self.logger_path = logger_path
self.cfg = cfg
def run(self):
if self.metric.lower_better == 1:
best_valid_score, best_test_score = float('inf'), float('inf')
else:
best_valid_score, best_test_score = -1, -1
for e in range(self.args.epoch):
self.train_step(e)
valid_score = self.test_step(self.valid_loader)
logger.info(f"E={e}, valid={valid_score:.5f}, test-score={best_test_score:.5f}")
# if valid_score > best_valid_score:
if (valid_score > best_valid_score and self.metric.lower_better == -1) or \
(valid_score < best_valid_score and self.metric.lower_better == 1):
test_score = self.test_step(self.test_loader)
best_valid_score = valid_score
best_test_score = test_score
logger.info(f"UPDATE test-score={best_test_score:.5f}")
logger.info(f"test-score={best_test_score:.5f}")
def train_step(self, epoch):
self.model.train()
if epoch % 4 in range(1):
# train separator
set_requires_grad([self.model.separator], requires_grad=True)
set_requires_grad([self.model.encoder], requires_grad=False)
else:
# train classifier
set_requires_grad([self.model.separator], requires_grad=False)
set_requires_grad([self.model.encoder], requires_grad=True)
pbar = tqdm(self.train_loader, desc=f"E [{epoch}]")
for data in pbar:
data = data.to(self.device)
c_logit, c_f, s_f, cmt_loss, reg_loss = self.model(data)
# classification loss on c
mask, target = nan2zero_get_mask(data, 'None', self.cfg)
cls_loss = self.metric.loss_func(c_logit, target.float(), reduction='none') * mask
cls_loss = cls_loss.sum() / mask.sum()
mix_f = self.model.mix_cs_proj(c_f, s_f)
inv_loss = self.simsiam_loss(c_f, mix_f)
# inv_w: lambda_1
# reg_w: lambda_2
loss = cls_loss + cmt_loss + self.args.inv_w * inv_loss + self.args.reg_w * reg_loss
self.opt.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5)
self.opt.step()
pbar.set_postfix_str(f"loss={loss.item():.4f}")
self.writer.add_scalar('loss', loss.item(), self.total_step)
self.writer.add_scalar('cls-loss', cls_loss.item(), self.total_step)
self.writer.add_scalar('cmt-loss', cmt_loss.item(), self.total_step)
self.writer.add_scalar('reg-loss', reg_loss.item(), self.total_step)
self.total_step += 1
@torch.no_grad()
def test_step(self, loader):
self.model.eval()
y_pred, y_gt = [], []
for data in loader:
data = data.to(self.device)
logit, _, _, _, _ = self.model(data)
mask, _ = nan2zero_get_mask(data, 'None', self.cfg)
pred, target = eval_data_preprocess(data.y, logit, mask, self.cfg)
y_pred.append(pred)
y_gt.append(target)
score = eval_score(y_pred, y_gt, self.cfg)
return score
def simsiam_loss(self, causal_rep, mix_rep):
causal_rep = causal_rep.detach()
causal_rep = F.normalize(causal_rep, dim=1)
mix_rep = F.normalize(mix_rep, dim=1)
return -(causal_rep * mix_rep).sum(dim=1).mean()
def main():
args = args_parser()
torch.cuda.set_device(int(args.gpu))
logger = initialize_exp(args)
set_seed(args.random_seed)
logger_path = get_dump_path(args)
writer = SummaryWriter(log_dir=os.path.join(logger_path, 'tensorboard'))
runner = Runner(args, writer, logger_path)
runner.run()
writer.close()
if __name__ == '__main__':
main()