-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
157 lines (118 loc) · 6.59 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
import argparse
import os
import random
import time
import warnings
import numpy as np
import torch
import torch.distributed as dist
import torch.nn.parallel as parallel
import torch.optim as optim
import torch.utils.data as data
import engine
import utils
import lr_lambda
def get_args_parser():
parser = argparse.ArgumentParser()
parser.add_argument('--random_seed', default=19970423, type=int)
parser.add_argument('--dataset', default='AdienceFace', help='{AdienceFace, HistoricalColor, ImageAesthetics}')
parser.add_argument('--adience_face_root', default='../../DataSet/AdienceFace', type=str)
parser.add_argument('--historical_color_root', default='../../DataSet/HistoricalColor', type=str)
parser.add_argument('--image_aesthetics_root', default='../../DataSet/ImageAesthetics', type=str)
parser.add_argument('--image_aesthetics_cat', default='nature', type=str, help='{nature, animals, urban, people}')
parser.add_argument('--constraint', default='S-P', type=str, help='{S-P, S-B, H-L, H-S}')
parser.add_argument('--feature_extractor', default='V16', type=str, help='{V16, R50, R101}')
parser.add_argument('--metric_method', default='E', type=str, help='{E, C}')
parser.add_argument('--feature_dim', default=512, type=int)
parser.add_argument('--cosine_scale', default=6., type=float)
parser.add_argument('--tau', default=0.11, type=float)
parser.add_argument('--loss_lam', default=5., type=float)
parser.add_argument('--epochs', default=48, type=int)
parser.add_argument('--batch_size', default=32, type=int)
parser.add_argument('--lr', default=0.01, type=float)
parser.add_argument('--lr_pl_mul', default=10, type=float)
parser.add_argument('--weight_decay', default=1e-4, type=float)
parser.add_argument('--warm_up_epoch', default=1, type=int)
parser.add_argument('--warm_up_ratio', default=0.333, type=float)
parser.add_argument('--milestones', default=[36, 44], type=list)
parser.add_argument('--step_gamma', default=0.1, type=float)
parser.add_argument('--local_rank', default=0, type=int)
parser.add_argument('--master_rank', default=0, type=int)
return parser.parse_args()
def set_random_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def get_model_name(args, fold_idx):
model_name = f'model_{args.dataset}_'
if args.dataset == 'IA':
model_name += f'{args.ia_cat}_'
model_name += f'{args.constraint}_fold_{fold_idx}.pkl'
return model_name
def run_fold(args, fold_idx):
train_dataset, val_dataset, test_dataset, num_ranks = utils.get_train_val_test_datasets(args, fold_idx)
train_dist_sampler = data.distributed.DistributedSampler(train_dataset)
val_dist_sampler = data.distributed.DistributedSampler(val_dataset)
train_data_loader = data.DataLoader(train_dataset, args.batch_size, num_workers=8, sampler=train_dist_sampler, pin_memory=True, drop_last=True)
val_data_loader = data.DataLoader(val_dataset, args.batch_size, num_workers=8, sampler=val_dist_sampler, pin_memory=True)
test_data_loader = data.DataLoader(test_dataset, args.batch_size, num_workers=8, pin_memory=True)
model, criterion = utils.get_model_criterion(num_ranks, args)
model = parallel.DistributedDataParallel(model.cuda(), device_ids=[dist.get_rank()], find_unused_parameters=True)
optim_parameters = [
{'params': [p for n, p in model.module.named_parameters() if n.startswith('feature_extractor') and p.requires_grad]},
{'params': [p for n, p in model.module.named_parameters() if not n.startswith('feature_extractor') and p.requires_grad], 'lr': args.lr * args.lr_pl_mul}
]
optimizer = optim.SGD(optim_parameters, lr=args.lr, weight_decay=args.weight_decay, momentum=0.9)
ll = lr_lambda.get_warm_up_multi_step_lr_lambda(len(train_dataset), args.warm_up_epoch, args.warm_up_ratio, args.milestones, args.step_gamma)
lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, ll)
ckpt_path = './ckpt'
if not os.path.exists(ckpt_path):
os.mkdir(ckpt_path)
model_path = os.path.join(ckpt_path, 'model.pkl')
best_val_acc, best_val_mae = 0, 100
for epoch_idx in range(args.epochs):
train_dist_sampler.set_epoch(epoch_idx)
epoch_loss, t_train = engine.train(model, criterion, optimizer, lr_scheduler, train_data_loader)
dist.reduce(epoch_loss, dst=args.master_rank, op=dist.ReduceOp.SUM)
val_acc, val_mae, val_t = engine.val(model, val_data_loader)
dist.reduce(val_acc, dst=args.master_rank, op=dist.ReduceOp.SUM)
dist.reduce(val_mae, dst=args.master_rank, op=dist.ReduceOp.SUM)
if dist.get_rank() == args.master_rank:
epoch_loss = epoch_loss.item() / dist.get_world_size()
val_acc = val_acc.item() / dist.get_world_size()
val_mae = val_mae.item() / dist.get_world_size()
print_str = f'F:{fold_idx};E:{epoch_idx + 1}/{args.epochs};Train:[Loss={epoch_loss:.4f};Time={t_train:.2f}s];Val:[ACC={val_acc:.3f};MAE={val_mae:.3f};Time={val_t:.2f}s].'
if val_mae < best_val_mae:
best_val_mae = val_mae
torch.save(model.module.state_dict(), model_path)
print_str += ' :-)'
print(print_str)
test_acc, test_mae = 0, 0
if dist.get_rank() == args.master_rank:
model.module.load_state_dict(torch.load(model_path))
test_acc, test_mae, test_t, test_plot = engine.val(model, test_data_loader)
test_acc, test_mae = test_acc.item(), test_mae.item()
print(f'Fold:{fold_idx};Test:[ACC={test_acc:.3f};MAE={test_mae:.3f}];Time={test_t:.1f}s.')
return test_acc, test_mae
def main():
st = time.time()
args = get_args_parser()
dist.init_process_group(backend='nccl')
set_random_seed(args.random_seed + dist.get_rank())
torch.cuda.set_device(torch.device(f'cuda:{dist.get_rank()}'))
warnings.filterwarnings('ignore')
num_folds = utils.get_num_folds(args)
test_acc_list, test_mae_list = [], []
for fold_idx in range(num_folds):
test_acc, test_mae = run_fold(args, fold_idx)
test_acc_list.append(test_acc)
test_mae_list.append(test_mae)
if dist.get_rank() == args.master_rank:
test_acc_mean = np.array(test_acc_list).mean()
test_acc_std = np.array(test_acc_list).std()
test_mae_mean = np.array(test_mae_list).mean()
test_mae_std = np.array(test_mae_list).std()
print(f'FinalTest:[ACC:{test_acc_mean:.3f}±{test_acc_std:.3f};MAE:{test_mae_mean:.3f}±{test_mae_std:.3f}].Time={time.time() - st:.2f}s.')
if __name__ == '__main__':
main()