-
Notifications
You must be signed in to change notification settings - Fork 94
/
CL_DM.py
162 lines (125 loc) · 7.49 KB
/
CL_DM.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
import os
import numpy as np
import torch
import argparse
from utils import get_dataset, get_network, get_eval_pool, evaluate_synset, ParamDiffAug, TensorDataset
import copy
import gc
def main():
parser = argparse.ArgumentParser(description='Parameter Processing')
parser.add_argument('--method', type=str, default='random', help='random/herding/DSA/DM')
parser.add_argument('--dataset', type=str, default='CIFAR100', help='dataset')
parser.add_argument('--model', type=str, default='ConvNet', help='model')
parser.add_argument('--ipc', type=int, default=20, help='image(s) per class')
parser.add_argument('--steps', type=int, default=5, help='5/10-step learning')
parser.add_argument('--num_eval', type=int, default=3, help='evaluation number')
parser.add_argument('--epoch_eval_train', type=int, default=1000, help='epochs to train a model with synthetic data')
parser.add_argument('--lr_net', 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('--data_path', type=str, default='./../data', help='dataset path')
args = parser.parse_args()
args.device = 'cuda' if torch.cuda.is_available() else 'cpu'
args.dsa_param = ParamDiffAug()
args.dsa = True # augment images for all methods
args.dsa_strategy = 'color_crop_cutout_flip_scale_rotate' # for CIFAR10/100
if not os.path.exists(args.data_path):
os.mkdir(args.data_path)
channel, im_size, num_classes, class_names, mean, std, dst_train, dst_test, testloader = get_dataset(args.dataset, args.data_path)
''' all training data '''
images_all = []
labels_all = []
indices_class = [[] for c in range(num_classes)]
images_all = [torch.unsqueeze(dst_train[i][0], dim=0) for i in range(len(dst_train))]
labels_all = [dst_train[i][1] for i in range(len(dst_train))]
for i, lab in enumerate(labels_all):
indices_class[lab].append(i)
images_all = torch.cat(images_all, dim=0).to(args.device)
labels_all = torch.tensor(labels_all, dtype=torch.long, device=args.device)
# for c in range(num_classes):
# print('class c = %d: %d real images' % (c, len(indices_class[c])))
def get_images(c, n): # get random n images from class c
idx_shuffle = np.random.permutation(indices_class[c])[:n]
return images_all[idx_shuffle]
print()
print('==================================================================================')
print('method: ', args.method)
results = np.zeros((args.steps, 5*args.num_eval))
for seed_cl in range(5):
num_classes_step = num_classes // args.steps
np.random.seed(seed_cl)
class_order = np.random.permutation(num_classes).tolist()
print('=========================================')
print('seed: ', seed_cl)
print('class_order: ', class_order)
print('augmentation strategy: \n', args.dsa_strategy)
print('augmentation parameters: \n', args.dsa_param.__dict__)
if args.method == 'random':
images_train_all = []
labels_train_all = []
for step in range(args.steps):
classes_current = class_order[step * num_classes_step: (step + 1) * num_classes_step]
images_train_all += [torch.cat([get_images(c, args.ipc) for c in classes_current], dim=0)]
labels_train_all += [torch.tensor([c for c in classes_current for i in range(args.ipc)], dtype=torch.long, device=args.device)]
elif args.method == 'herding':
fname = os.path.join(args.data_path, 'metasets', 'cl_data', 'cl_herding_CIFAR100_ConvNet_20ipc_%dsteps_seed%d.pt'%(args.steps, seed_cl))
data = torch.load(fname, map_location='cpu')['data']
images_train_all = [data[step][0] for step in range(args.steps)]
labels_train_all = [data[step][1] for step in range(args.steps)]
print('use data: ', fname)
elif args.method == 'DSA':
fname = os.path.join(args.data_path, 'metasets', 'cl_data', 'cl_res_DSA_CIFAR100_ConvNet_20ipc_%dsteps_seed%d.pt'%(args.steps, seed_cl))
data = torch.load(fname, map_location='cpu')['data']
images_train_all = [data[step][0] for step in range(args.steps)]
labels_train_all = [data[step][1] for step in range(args.steps)]
print('use data: ', fname)
elif args.method == 'DM':
fname = os.path.join(args.data_path, 'metasets', 'cl_data', 'cl_DM_CIFAR100_ConvNet_20ipc_%dsteps_seed%d.pt'%(args.steps, seed_cl))
data = torch.load(fname, map_location='cpu')['data']
images_train_all = [data[step][0] for step in range(args.steps)]
labels_train_all = [data[step][1] for step in range(args.steps)]
print('use data: ', fname)
else:
exit('unknown method: %s'%args.method)
for step in range(args.steps):
print('\n-----------------------------\nmethod %s seed %d step %d ' % (args.method, seed_cl, step))
classes_seen = class_order[: (step+1)*num_classes_step]
print('classes_seen: ', classes_seen)
''' train data '''
images_train = torch.cat(images_train_all[:step+1], dim=0).to(args.device)
labels_train = torch.cat(labels_train_all[:step+1], dim=0).to(args.device)
print('train data size: ', images_train.shape)
''' test data '''
images_test = []
labels_test = []
for i in range(len(dst_test)):
lab = int(dst_test[i][1])
if lab in classes_seen:
images_test.append(torch.unsqueeze(dst_test[i][0], dim=0))
labels_test.append(dst_test[i][1])
images_test = torch.cat(images_test, dim=0).to(args.device)
labels_test = torch.tensor(labels_test, dtype=torch.long, device=args.device)
dst_test_current = TensorDataset(images_test, labels_test)
testloader = torch.utils.data.DataLoader(dst_test_current, batch_size=256, shuffle=False, num_workers=0)
print('test set size: ', images_test.shape)
''' train model on the newest memory '''
accs = []
for ep_eval in range(args.num_eval):
net_eval = get_network(args.model, channel, num_classes, im_size)
net_eval = net_eval.to(args.device)
img_syn_eval = copy.deepcopy(images_train.detach())
lab_syn_eval = copy.deepcopy(labels_train.detach())
_, acc_train, acc_test = evaluate_synset(ep_eval, net_eval, img_syn_eval, lab_syn_eval, testloader, args)
del net_eval, img_syn_eval, lab_syn_eval
gc.collect() # to reduce memory cost
accs.append(acc_test)
results[step, seed_cl*args.num_eval + ep_eval] = acc_test
print('Evaluate %d random %s, mean = %.4f std = %.4f' % (len(accs), args.model, np.mean(accs), np.std(accs)))
results_str = ''
for step in range(args.steps):
results_str += '& %.1f$\pm$%.1f ' % (np.mean(results[step]) * 100, np.std(results[step]) * 100)
print('\n\n')
print('%d step learning %s perforamnce:'%(args.steps, args.method))
print(results_str)
print('Done')
if __name__ == '__main__':
main()