-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_script.py
197 lines (162 loc) · 7.89 KB
/
train_script.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
from __future__ import print_function
from __future__ import absolute_import
import os
import numpy as np
import argparse
import importlib
import time
import logging
import json
from collections import OrderedDict
import importlib
import torch
import torch.nn as nn
from torch.utils.data.dataset import Dataset
from torchvision import datasets, transforms
import models
import utils.robust_train_data_utils as data
import utils.robust_train_utils as robust_utils
import model_trainers
from utils.robust_train_utils import *
# import utils
# from utils import *
class create_epsilon_scheduler:
def __init__(self, max_epsilon, epochs, args):
self.max_epsilon = max_epsilon
self.args = args
self.epoch = 0
def step(self):
if self.epoch < 5:
self.args.epsilon = self.max_epsilon // 4
elif self.epoch < 10:
self.args.epsilon = self.max_epsilon // 2
else:
self.args.epsilon = self.max_epsilon
self.epoch += 1
def main():
parser = argparse.ArgumentParser(description="Robust learning")
parser.add_argument("--configs", type=str, default="./configs/configs_cifar.yml")
parser.add_argument(
"--results-dir", type=str, default="./results/",
)
parser.add_argument("--exp-name", type=str, default="temp")
parser.add_argument("--arch", type=str)
parser.add_argument("--lr", type=float)
parser.add_argument("--epochs", type=int)
parser.add_argument("--batch-size", type=int)
# training
parser.add_argument("--trainer", type=str, default="baseline", choices=("baseline", "madry", "adv"))
parser.add_argument("--val-method", type=str, default="baseline", choices=("baseline", "adv", "auto"))
parser.add_argument("--local_rank", default=0, type=int)
parser.add_argument("--accelerator", type=str, default="dp", choices=("dp"))
# use soft-labels
parser.add_argument("--opt-probs", action="store_true", default=False, help="use optimal probs as soft labels")
parser.add_argument("--clip-soft-labels", action="store_true", default=False, help="Clip soft-labels")
parser.add_argument("--drop-soft-labels", action="store_true", default=False, help="Drop soft-labels")
# adv
parser.add_argument("--epsilon", type=float)
parser.add_argument("--num-steps", type=int)
parser.add_argument("--step-size", type=float)
parser.add_argument("--beta", type=float)
# misc
parser.add_argument("--trial", type=int, default=0)
parser.add_argument("--ckpt", type=str, help="checkpoint path for pretrained classifier")
parser.add_argument("--print-freq", type=int, default=50)
parser.add_argument("--seed", type=int, default=12345)
args = update_args(parser.parse_args())
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
ngpus = torch.cuda.device_count() # Control available gpus by CUDA_VISIBLE_DEVICES only
print(f"Using {ngpus} gpus")
assert args.normalize == False, "Presumption for most code is that the pixel range is [0,1]"
assert torch.backends.cudnn.enabled, "Amp requires cudnn backend to be enabled."
assert args.mode == "base", "Even small amount of data-augmentation hurts generalization badly at higher epsilons"
# seed cuda
torch.manual_seed((args.local_rank+1)*args.seed)
torch.cuda.manual_seed((args.local_rank+1)*args.seed)
torch.cuda.manual_seed_all((args.local_rank+1)*args.seed)
np.random.seed((args.local_rank+1)*args.seed)
# create resutls dir (for logs, checkpoints, etc.)
if not os.path.isdir(args.results_dir):
os.mkdir(args.results_dir)
result_main_dir = os.path.join(args.results_dir, args.exp_name)
result_sub_dir = os.path.join(result_main_dir, f"trial_{args.trial}")
create_subdirs(result_sub_dir)
# add logger
logging.basicConfig(level=logging.INFO, format="%(message)s")
logger = logging.getLogger()
logger.addHandler(
logging.FileHandler(os.path.join(result_sub_dir, "setup.log"), "a")
)
logger.info(args)
# # multi-gpu DDP
# if args.accelerator == "ddp":
# torch.distributed.init_process_group(backend='nccl',
# init_method='env://')
# world_size = torch.distributed.get_world_size()
# print("world_size =", world_size)
# # Scale learning rate based on global batch size
# args.batch_size = args.batch_size // world_size
# args.workers = args.workers // world_size
# print(f"New per-gpu batch-size = {args.batch_size}, workers = {args.batch_size}")
# create model + optimizer
model = models.__dict__[args.arch](num_classes=args.num_classes, channels=args.channels).to(device).train()
if args.ckpt is not None:
d = fix_legacy_dict(torch.load(args.ckpt, map_location="cpu"))
model.load_state_dict(d, strict=True)
print(f"Mismatched keys {set(d.keys()) ^ set(model.state_dict().keys())}")
print(f"model loaded from {args.ckpt}")
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
# half-precision support (Actually O2 in amp is mixed-precision)
# if args.fp16:
# #print("using apex synced BN")
# #model = apex.parallel.convert_syncbn_model(model)
# # O2 opt-level by default (O2 keeps batch_norm in float32)
# model, optimizer = amp.initialize(model, optimizer, opt_level='O2')
# parallelization
if ngpus > 1:
print(f"Using multiple gpus")
if args.accelerator == "dp":
model = nn.DataParallel(model).to(device)
# elif args.accelerator == "ddp":
# model = DDP(model, delay_allreduce=True)
else:
raise ValueError("accelerator not supported")
# dataloaders
train_loader, train_sampler, val_loader, val_sampler, _, _, train_transform = data.__dict__[args.dataset](args.data_dir, batch_size=args.batch_size, mode=args.mode, normalize=args.normalize, size=args.size, workers=args.workers, args=args)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs * len(train_loader), 1e-4)
criterion = nn.CrossEntropyLoss()
if args.trainer in ["madry", "adv"]:
eps_scheduler = create_epsilon_scheduler(np.copy(args.epsilon), args.epochs, args)
best_prec = 0
# Let's roll
for epoch in range(0, args.epochs):
# if args.distributed:
# train_sampler.set_epoch(epoch)
# if args.trainer in ["madry", "adv"]:
# eps_scheduler.step()
results_train = getattr(model_trainers, args.trainer)(model, device, train_loader, criterion, optimizer, lr_scheduler, epoch, args)
results_val = getattr(robust_utils, args.val_method)(model, device, val_loader, criterion, args, epoch)
if args.local_rank == 0:
# remember best prec@1 (only based on clean accuracy) and save checkpoint
if args.trainer == "baseline":
prec = results_val["top1"]
elif args.trainer in ["adv", "madry"]:
prec = results_val["top1_adv"]
else:
raise ValueError()
is_best = prec > best_prec
best_prec = max(prec, best_prec)
d = {
"epoch": epoch + 1,
"arch": args.arch,
"state_dict": model.state_dict(),
"best_prec1": best_prec,
"optimizer": optimizer.state_dict(),
}
save_checkpoint(
d, is_best, result_dir=os.path.join(result_sub_dir, "checkpoint"),
)
logger.info(f"Epoch {epoch}, " + ", ".join(["{}: {:.3f}".format(k+"_train", v) for (k,v) in results_train.items()]+["{}: {:.3f}".format(k+"_val", v) for (k,v) in results_val.items()]))
if __name__ == "__main__":
main()