-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
181 lines (149 loc) · 8.23 KB
/
main.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
import argparse
from datetime import datetime
import torch
import random
import wandb
import os
import json
from dataloader import get_dataloaders
from utils import get_model, save_config,set_seed,ddp_setup
from torch.distributed import destroy_process_group
import torch.multiprocessing as mp
from train import Trainer, VanillaTrainer
from fixed_main import main as f_main
from fvcore.nn import FlopCountAnalysis, parameter_count, flop_count_table
def main(args):
print(f"PID: {os.getpid()}")
if args.wandb:
config = json.load(open('config.json'))
os.environ["WANDB_API_KEY"] = config['WANDB_API_KEY']
wandb.login(key=os.environ['WANDB_API_KEY'])
wandb.init(project=args.project, config=args, name=args.experiment)
train_dl, valid_dl, test_dl = get_dataloaders(args)
model = get_model(args)
# flops
input = torch.rand((1, 3, args.size, args.size)).to(args.device)
flops = FlopCountAnalysis(model, input)
args.flops = flops.total()
args.n_params = parameter_count(model)['']
os.makedirs(os.path.join(args.output, args.experiment), exist_ok=True)
with open(os.path.join(args.output, args.experiment, "flops_table.txt"), "w") as text_file:
text_file.write(f"{flop_count_table(flops, max_depth=0)}")
save_config(args)
trainer = Trainer(model, args)
trainer.fit(train_dl, valid_dl, test_dl, args)
if args.retrain_fixed:
args.path = os.path.join(args.output, args.project, args.experiment)
f_main(args)
def distributed_main(rank, world_size, args):
print(f"PID: {os.getpid()}")
args.device = rank
args.world_size = world_size
ddp_setup(rank, world_size)
train_dl, valid_dl, test_dl = get_dataloaders(args)
model = get_model(args)
input = torch.rand((1, 3, args.size, args.size)).to(args.device)
flops = FlopCountAnalysis(model, input)
args.flops = flops.total()
args.n_params = parameter_count(model)['']
os.makedirs(os.path.join(args.output, args.experiment), exist_ok=True)
with open(os.path.join(args.output, args.experiment, "flops_table.txt"), "w") as text_file:
text_file.write(f"{flop_count_table(flops, max_depth=0)}")
save_config(args)
trainer = Trainer(model, args)
trainer.fit(train_dl, valid_dl, test_dl, args)
destroy_process_group()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# data
parser.add_argument('--project', type=str, default=f"search_mixer")
parser.add_argument('--output', type=str, default=f"./out")
parser.add_argument('--experiment', type=str, default=f"experiment_{datetime.now().strftime('%Y-%m-%d_%H:%M:%S')}")
parser.add_argument('--seed', type=int, default=random.randint(1, 1e5))
parser.add_argument('--dataset', required=True, choices=['c10', 'c100', 'svhn', 'stl10', 'imagenet','imagenet100'])
parser.add_argument('--img-size', type=int, dest='size')
parser.add_argument('--padding', type=int)
parser.add_argument('--valid-ratio', type=float, default=0.5)
parser.add_argument('--batch-size', type=int, default=128)
parser.add_argument('--num-workers', type=int, default=4)
parser.add_argument('--eval-batch-size', type=int, default=1024)
parser.add_argument('--autoaugment', action='store_true')
# model
parser.add_argument('--model', required=True, choices=['mlp_mixer', 'search_mixer'])
parser.add_argument('--hidden-size', type=int, default=128)
parser.add_argument('--patch-size', type=int, default=4)
parser.add_argument('--hidden-s-candidates', nargs='+', help='<Required> Set flag', required=True)
parser.add_argument('--hidden-c-candidates', nargs='+', help='<Required> Set flag', required=True)
parser.add_argument('--n-cells', type=int, default=8)
parser.add_argument('--drop-p', type=float, default=0.)
parser.add_argument('--off-act', action='store_true', help='Disable activation function')
parser.add_argument('--is-cls-token', action='store_true', help='Introduce a class token.')
# training
parser.add_argument('--clip-grad', type=float, default=0, help="0 means disabling clip-grad")
parser.add_argument('--cutmix-beta', type=float, default=1.0)
parser.add_argument('--cutmix-prob', type=float, default=0.)
parser.add_argument('--label-smoothing', type=float, default=0.1)
parser.add_argument('--w01', type=float, default=1.)
parser.add_argument('--mu', type=float, default=0.)
parser.add_argument('--l', type=float, default=0.)
parser.add_argument('--w-optimizer', default='Adam', choices=['Adam', 'SGD','Adamw'])
parser.add_argument('--w-lr', type=float, default=1e-3)
parser.add_argument('--w-momentum', type=float, default=0.9)
parser.add_argument('--w-weight-decay', type=float, default=5e-5)
parser.add_argument('--off-nesterov', action='store_true')
parser.add_argument('--w-beta1', type=float, default=0.9)
parser.add_argument('--w-beta2', type=float, default=0.99)
parser.add_argument('--a-optimizer', default='Adam', choices=['Adam', 'SGD'])
parser.add_argument('--a-lr', type=float, default=1e-3)
parser.add_argument('--a-beta1', type=float, default=0.9)
parser.add_argument('--a-beta2', type=float, default=0.99)
parser.add_argument('--a-weight-decay', type=float, default=5e-5)
parser.add_argument('--w-scheduler', default='cosine', choices=['step', 'cosine'])
parser.add_argument('--epochs', type=int, default=300)
parser.add_argument('--w-gamma', type=float, default=0.1)
parser.add_argument('--w-min-lr', type=float, default=1e-6)
parser.add_argument('--warmup-epochs', type=int, default=5)
parser.add_argument('--unrolled', action='store_true')
parser.add_argument('--wandb', action='store_true')
parser.add_argument('--use-amp', action='store_true')
# Retrain Fixed Arch
parser.add_argument('--use-timm-transform', action='store_true', help='Use timm.data transforms (for Imagenet only)')
parser.add_argument('--retrain-fixed', action='store_true')
parser.add_argument('--fixed-batch-size', type=int, default=128)
parser.add_argument('--fixed-eval-batch-size', type=int, default=1024)
parser.add_argument('--fixed-seed', type=int, default=random.randint(1, 1e5))
parser.add_argument('--fixed-epochs', type=int, default=300)
parser.add_argument('--fixed-th', type=float, default=0.5)
parser.add_argument('--fixed-lr', type=float, default=1e-3)
parser.add_argument('--fixed-min-lr', type=float, default=1e-6)
parser.add_argument('--fixed-momentum', type=float, default=0.9)
parser.add_argument('--fixed-optimizer', default='adam', choices=['Adam', 'SGD'])
parser.add_argument('--fixed-scheduler', default='cosine', choices=['step', 'cosine'])
parser.add_argument('--fixed-beta1', type=float, default=0.9)
parser.add_argument('--fixed-beta2', type=float, default=0.99)
parser.add_argument('--fixed-weight-decay', type=float, default=5e-5)
parser.add_argument('--fixed-off-nesterov', action='store_true')
parser.add_argument('--fixed-label-smoothing', type=float, default=0.1)
parser.add_argument('--fixed-gamma', type=float, default=0.1)
parser.add_argument('--fixed-warmup-epoch', type=int, default=5)
parser.add_argument('--fixed-autoaugment', action='store_true')
parser.add_argument('--fixed-clip-grad', type=float, default=0, help="0 means disabling clip-grad")
parser.add_argument('--fixed-cutmix-beta', type=float, default=1.0)
parser.add_argument('--fixed-cutmix-prob', type=float, default=0.)
parser.add_argument('--discrete', action='store_true')
parser.add_argument('--distributed', action='store_true')
args = parser.parse_args()
args.hidden_s_candidates = list(map(lambda x: int(x), args.hidden_s_candidates))
args.hidden_c_candidates = list(map(lambda x: int(x), args.hidden_c_candidates))
args.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
args.nesterov = not args.off_nesterov
args.pid = os.getpid()
args.subexperiment=f"experiment_{datetime.now().strftime('%Y-%m-%d_%H:%M:%S')}"
# torch.random.manual_seed(args.seed)
set_seed(args.seed)
# main(args)
if args.distributed:
world_size = torch.cuda.device_count()
mp.spawn(distributed_main, args=(world_size, args), nprocs=world_size)
else:
main(args)