-
Notifications
You must be signed in to change notification settings - Fork 0
/
test.py
224 lines (200 loc) · 10.8 KB
/
test.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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
import numpy as np
import torch
from torchvision import transforms
from tqdm import tqdm
import os
# from torchmeta.datasets import CUB
from cub import CUB
from torchmeta.utils.data import BatchMetaDataLoader
from torchmeta.transforms import Categorical, ClassSplitter
from gbml.maml_higher import MAML
from gbml.fomaml_higher import FOMAML
from mbml.protonet import ProtoNet
from utils import set_seed, set_gpu, check_dir, dict2json, ImageJitter
def train(args, model, dataloader):
# model.network.train()
loss_list = []
acc_list = []
with tqdm(dataloader, total=args.num_train_batches) as pbar:
for batch_idx, batch in enumerate(pbar):
loss_log, acc_log, _ = model.outer_loop(batch, is_train=True)
loss_list.append(loss_log)
acc_list.append(acc_log)
pbar.set_description('loss = {:.4f} || acc={:.4f}'.format(np.mean(loss_list), np.mean(acc_list)))
if batch_idx >= args.num_train_batches:
break
loss_mean = np.round(np.mean(loss_list), 4)
loss_err = np.round((1.96 * np.std(loss_list) / np.sqrt(args.num_train_batches)), 2)
acc_mean = np.round(np.mean(acc_list) * 100, 2)
acc_err = np.round((1.96 * np.std(acc_list) / np.sqrt(args.num_train_batches)) * 100, 2)
return loss_mean, loss_err, acc_mean, acc_err
# @torch.no_grad()
def valid(args, model, dataloader):
# model.network.eval()
loss_list = []
acc_list = []
with tqdm(dataloader, total=args.num_valid_batches) as pbar:
for batch_idx, batch in enumerate(pbar):
loss_log, acc_log = model.outer_loop(batch, is_train=False)
loss_list.append(loss_log)
acc_list.append(acc_log)
pbar.set_description('loss = {:.4f} || acc={:.4f}'.format(np.mean(loss_list), np.mean(acc_list)))
if batch_idx >= args.num_valid_batches:
break
loss_mean = np.round(np.mean(loss_list), 4)
loss_err = np.round((1.96 * np.std(loss_list) / np.sqrt(args.num_valid_batches)), 2)
acc_mean = np.round(np.mean(acc_list) * 100, 2)
acc_err = np.round((1.96 * np.std(acc_list) / np.sqrt(args.num_valid_batches)) * 100, 2)
return loss_mean, loss_err, acc_mean, acc_err
def main(args):
if args.alg=='MAML':
model = MAML(args)
elif args.alg=='FOMAML':
model = FOMAML(args)
elif args.alg=='ProtoNet':
model = ProtoNet(args)
else:
raise ValueError('Not implemented Meta-Learning Algorithm')
if args.dataset=='CUB':
dataclass = CUB
print("loaded cub")
else:
raise ValueError('Not used Data-set ')
if args.load:
model.load()
elif args.load_encoder:
model.load_encoder()
train_dataset = dataclass(args.data_path, num_classes_per_task=args.num_way,
meta_split='train',
transform=transforms.Compose([
transforms.RandomResizedCrop(args.image_size),
ImageJitter(dict(Brightness=0.4, Contrast=0.4, Color=0.4)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(
np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225])),
]),
target_transform=Categorical(num_classes=args.num_way)
)
train_dataset = ClassSplitter(train_dataset, shuffle=True, num_train_per_class=args.num_shot, num_test_per_class=args.num_query)
train_loader = BatchMetaDataLoader(train_dataset, batch_size=args.batch_size,
shuffle=True, pin_memory=args.pin_memory, num_workers=args.num_workers)
valid_dataset = dataclass(args.data_path, num_classes_per_task=args.num_way,
meta_split='val',
transform=transforms.Compose([
transforms.Resize([int(args.image_size * 1.15), int(args.image_size * 1.15)]),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Normalize(
np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225]))
]),
target_transform=Categorical(num_classes=args.num_way)
)
valid_dataset = ClassSplitter(valid_dataset, shuffle=True, num_train_per_class=args.num_shot, num_test_per_class=args.num_query)
valid_loader = BatchMetaDataLoader(valid_dataset, batch_size=args.batch_size,
shuffle=True, pin_memory=args.pin_memory, num_workers=args.num_workers)
test_dataset = dataclass(args.data_path, num_classes_per_task=args.num_way,
meta_split='test',
transform=transforms.Compose([
transforms.Resize([int(args.image_size * 1.15), int(args.image_size * 1.15)]),
transforms.CenterCrop(args.image_size),
transforms.ToTensor(),
transforms.Normalize(
np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225]))
]),
target_transform=Categorical(num_classes=args.num_way)
)
test_dataset = ClassSplitter(test_dataset, shuffle=True, num_train_per_class=args.num_shot, num_test_per_class=args.num_query)
test_loader = BatchMetaDataLoader(test_dataset, batch_size=args.batch_size,
shuffle=True, pin_memory=args.pin_memory, num_workers=args.num_workers)
print(len(train_loader), len(valid_loader), len(test_loader))
start_epoch = 0
resume_file = os.path.join(args.save_path, 'best_model.pth')
if resume_file is not None:
tmp = torch.load(resume_file)
model.network.load_state_dict(tmp['state'])
for epoch in range(start_epoch, args.num_epoch):
log_dict = {'epoch': epoch}
test_loss, test_loss_err, test_acc, test_acc_err = valid(args, model, test_loader)
log_dict['te_loss'] = test_loss
log_dict['te_loss_err'] = test_loss_err
log_dict['te_acc'] = test_acc
log_dict['te_acc_err'] = test_acc_err
dict2json(os.path.join(args.save_path, 'test_log.json'), log_dict)
return None
def parse_args():
import argparse
parser = argparse.ArgumentParser('Meta-Learning Algorithms')
# experimental settings
parser.add_argument('--seed', type=int, default=2022, help='Random seed.')
parser.add_argument('--dataset', type=str, default='CUB', help='dataset')
parser.add_argument('--data_path', type=str, default='./data', help='Path of dataset.')
parser.add_argument('--result_path', type=str, default='./result')
parser.add_argument('--log_path', type=str, default='result.tsv')
parser.add_argument('--save_path', type=str, default='')
parser.add_argument('--load', type=lambda x: (str(x).lower() == 'true'), default=False)
parser.add_argument('--load_encoder', type=lambda x: (str(x).lower() == 'true'), default=False)
parser.add_argument('--load_path', type=str, default='best_model.pth')
parser.add_argument('--resume', action='store_true', help='continue from previous trained model with largest epoch')
parser.add_argument('--device', type=int, nargs='+', default=[0], help='0 = CPU.')
parser.add_argument('--num_workers', type=int, default=4,
help='Number of workers for data loading (default: 4).')
parser.add_argument('--pin_memory', action='store_false', default=True, help='if or not use pin_memory')
# training settings
parser.add_argument('--num_epoch', type=int, default=1, help='Number of epochs for meta test.')
parser.add_argument('--batch_size', type=int, default=1, help='Number of tasks in a mini-batch of tasks (default: 1).')
parser.add_argument('--num_train_batches', type=int, default= 100, help='Number of batches the model is trained over (default: 100).')
parser.add_argument('--num_valid_batches', type=int, default= 600, help='Number of batches the model is trained over (default: 600).')
# meta-learning settings
parser.add_argument('--num_shot', type=int, default=1, help='Number of support examples per class (k in "k-shot", default: 1).')
parser.add_argument('--num_query', type=int, default=15, help='Number of query examples per class (k in "k-query", default: 15).')
parser.add_argument('--num_way', type=int, default=5, help='Number of classes per task (N in "N-way", default: 5).')
parser.add_argument('--alg', type=str, default='MAML')
# algorithm settings
parser.add_argument('--n_inner', type=int, default=5)
parser.add_argument('--inner_lr', type=float, default=1e-2)
parser.add_argument('--inner_opt', type=str, default='SGD')
parser.add_argument('--outer_lr', type=float, default=1e-3)
parser.add_argument('--outer_opt', type=str, default='Adam')
parser.add_argument('--lr_sched', type=str, default='None')
# network settings
parser.add_argument('--net', type=str, default='ConvNet')
parser.add_argument('--n_conv', type=int, default=4)
parser.add_argument('--n_dense', type=int, default=0)
parser.add_argument('--hidden_dim', type=int, default=64)
parser.add_argument('--in_channels', type=int, default=3)
parser.add_argument('--hidden_channels', type=int, default=64, help='Number of channels for each convolutional layer (default: 64).')
parser.add_argument('--image_size', type=int, default=84)
# catfish settings
parser.add_argument('--net_aug', type=str, default='',
choices=['mask-layer-snip-mbml', 'mask-layer-snip-mbml-small', 'mask-layer-snip-fomaml',
'mask-layer-snip-fomaml-small'])
parser.add_argument('--max_width', type=float, default=1.)
parser.add_argument('--min_width', type=float, default=0.8)
parser.add_argument('--num_subnet', type=int, default=3)
parser.add_argument('--shot_aug', type=str, default='', choices=['resize'])
parser.add_argument('--resos', type=list, default=[84, 64, 48])
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
set_seed(args.seed)
set_gpu(args.device)
if 'Conv' in args.net:
args.image_size = 84
args.resos = [84, 64, 48]
else:
args.image_size = 224
args.resos = [224, 192, 160, 140]
save_str = args.alg + '_' + args.dataset + '_' + args.net + '_' + str(args.num_way) + 'w' + str(
args.num_shot) + 's_' + str(args.net_aug) + '_(' + str(args.min_width) + ',' + str(args.max_width) + ')' + '_' +args.shot_aug
if args.shot_aug == 'resize':
save_str = save_str + '_' + str(args.resos)
args.save_path = os.path.join(args.result_path, save_str)
check_dir(args.save_path)
if 'maml' in str(args.alg).lower():
args.batch_size = 4
main(args)