-
Notifications
You must be signed in to change notification settings - Fork 53
/
buffer.py
136 lines (103 loc) · 6.16 KB
/
buffer.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
import os
import argparse
import torch
import torch.nn as nn
from tqdm import tqdm
from utils import get_dataset, get_network, get_daparam,\
TensorDataset, epoch, ParamDiffAug
import copy
import warnings
warnings.filterwarnings("ignore", category=DeprecationWarning)
def main(args):
args.dsa = True if args.dsa == 'True' else False
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
args.dsa_param = ParamDiffAug()
channel, im_size, num_classes, class_names, mean, std, dst_train, dst_test, testloader, loader_train_dict, class_map, class_map_inv = get_dataset(args.dataset, args.data_path, args.batch_real, args.subset, args=args)
# print('\n================== Exp %d ==================\n '%exp)
print('Hyper-parameters: \n', args.__dict__)
save_dir = os.path.join(args.buffer_path, args.dataset)
if args.dataset == "ImageNet":
save_dir = os.path.join(save_dir, args.subset, str(args.res))
if args.dataset in ["CIFAR10", "CIFAR100"] and not args.zca:
save_dir += "_NO_ZCA"
save_dir = os.path.join(save_dir, args.model)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
''' organize the real dataset '''
images_all = []
labels_all = []
indices_class = [[] for c in range(num_classes)]
print("BUILDING DATASET")
for i in tqdm(range(len(dst_train))):
sample = dst_train[i]
images_all.append(torch.unsqueeze(sample[0], dim=0))
labels_all.append(class_map[torch.tensor(sample[1]).item()])
for i, lab in tqdm(enumerate(labels_all)):
indices_class[lab].append(i)
images_all = torch.cat(images_all, dim=0).to("cpu")
labels_all = torch.tensor(labels_all, dtype=torch.long, device="cpu")
for c in range(num_classes):
print('class c = %d: %d real images'%(c, len(indices_class[c])))
for ch in range(channel):
print('real images channel %d, mean = %.4f, std = %.4f'%(ch, torch.mean(images_all[:, ch]), torch.std(images_all[:, ch])))
criterion = nn.CrossEntropyLoss().to(args.device)
trajectories = []
dst_train = TensorDataset(copy.deepcopy(images_all.detach()), copy.deepcopy(labels_all.detach()))
trainloader = torch.utils.data.DataLoader(dst_train, batch_size=args.batch_train, shuffle=True, num_workers=0)
''' set augmentation for whole-dataset training '''
args.dc_aug_param = get_daparam(args.dataset, args.model, args.model, None)
args.dc_aug_param['strategy'] = 'crop_scale_rotate' # for whole-dataset training
print('DC augmentation parameters: \n', args.dc_aug_param)
for it in range(0, args.num_experts):
''' Train synthetic data '''
teacher_net = get_network(args.model, channel, num_classes, im_size).to(args.device) # get a random model
teacher_net.train()
lr = args.lr_teacher
teacher_optim = torch.optim.SGD(teacher_net.parameters(), lr=lr, momentum=args.mom, weight_decay=args.l2) # optimizer_img for synthetic data
teacher_optim.zero_grad()
timestamps = []
timestamps.append([p.detach().cpu() for p in teacher_net.parameters()])
lr_schedule = [args.train_epochs // 2 + 1]
for e in range(args.train_epochs):
train_loss, train_acc = epoch("train", dataloader=trainloader, net=teacher_net, optimizer=teacher_optim,
criterion=criterion, args=args, aug=True)
test_loss, test_acc = epoch("test", dataloader=testloader, net=teacher_net, optimizer=None,
criterion=criterion, args=args, aug=False)
print("Itr: {}\tEpoch: {}\tTrain Acc: {}\tTest Acc: {}".format(it, e, train_acc, test_acc))
timestamps.append([p.detach().cpu() for p in teacher_net.parameters()])
if e in lr_schedule and args.decay:
lr *= 0.1
teacher_optim = torch.optim.SGD(teacher_net.parameters(), lr=lr, momentum=args.mom, weight_decay=args.l2)
teacher_optim.zero_grad()
trajectories.append(timestamps)
if len(trajectories) == args.save_interval:
n = 0
while os.path.exists(os.path.join(save_dir, "replay_buffer_{}.pt".format(n))):
n += 1
print("Saving {}".format(os.path.join(save_dir, "replay_buffer_{}.pt".format(n))))
torch.save(trajectories, os.path.join(save_dir, "replay_buffer_{}.pt".format(n)))
trajectories = []
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Parameter Processing')
parser.add_argument('--dataset', type=str, default='CIFAR10', help='dataset')
parser.add_argument('--subset', type=str, default='imagenette', help='subset')
parser.add_argument('--model', type=str, default='ConvNet', help='model')
parser.add_argument('--res', type=int, default=128, help='resolution for imagenet')
parser.add_argument('--num_experts', type=int, default=100, help='training iterations')
parser.add_argument('--lr_teacher', type=float, default=0.01, help='learning rate for updating network parameters')
parser.add_argument('--batch_train', type=int, default=256, help='batch size for training networks')
parser.add_argument('--batch_real', type=int, default=256, help='batch size for real loader')
parser.add_argument('--dsa', type=str, default='True', choices=['True', 'False'],
help='whether to use differentiable Siamese augmentation.')
parser.add_argument('--dsa_strategy', type=str, default='color_crop_cutout_flip_scale_rotate',
help='differentiable Siamese augmentation strategy')
parser.add_argument('--data_path', type=str, default='data', help='dataset path')
parser.add_argument('--buffer_path', type=str, default='./buffers', help='buffer path')
parser.add_argument('--train_epochs', type=int, default=50)
parser.add_argument('--zca', action='store_true')
parser.add_argument('--decay', action='store_true')
parser.add_argument('--mom', type=float, default=0, help='momentum')
parser.add_argument('--l2', type=float, default=0, help='l2 regularization')
parser.add_argument('--save_interval', type=int, default=10)
args = parser.parse_args()
main(args)