-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_aviator_benchmark.py
359 lines (314 loc) · 15.5 KB
/
train_aviator_benchmark.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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
import argparse
import os.path as osp
import numpy as np
from copy import deepcopy
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
from model.dataloader.samplers import CategoriesSampler
from model.models.mamlp_fc import MAMLP_FC
from model.utils import pprint, set_gpu, ensure_path, Averager, Timer, count_acc, euclidean_metric, compute_confidence_interval
from tensorboardX import SummaryWriter
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--max_epoch', type=int, default=500)
parser.add_argument('--way', type=int, default=5)
parser.add_argument('--shot', type=int, default=1)
parser.add_argument('--query', type=int, default=15)
parser.add_argument('--gd_lr', type=float, default=0.01) # lr for gd updates
parser.add_argument('--lr', type=float, default=0.001) # lr for meta updates
parser.add_argument('--lr_mul', type=float, default=10) # lr is the basic learning rate, while lr * lr_mul is the lr for other parts
parser.add_argument('--inner_iters', type=int, default=1)
parser.add_argument('--step_size', type=int, default=20)
parser.add_argument('--gamma', type=float, default=0.5)
parser.add_argument('--temperature', type=float, default=1)
parser.add_argument('--model_type', type=str, default='ConvNet', choices=['ConvNet'])
parser.add_argument('--dataset', type=str, default='MiniImageNet', choices=['MiniImageNet', 'CUB'])
# MiniImageNet, ConvNet, './saves/initialization/miniimagenet/con-pre-noaug.pth'
# MiniImageNet, ResNet, './saves/initialization/miniimagenet/res-pre.pth'
# CUB, ConvNet, './saves/initialization/cub/con-pre.pth'
parser.add_argument('--init_weights', type=str, default=None)
# parser.add_argument('--init_weights', type=str, default='saves/initialization/miniimagenet/con-pre.pth')
# parser.add_argument('--init_weights', type=str, default='saves/initialization/cub/con-pre.pth')
parser.add_argument('--comment', type=str, default='temp') # The temp name to save the reulst file
parser.add_argument('--gpu', default='0')
args = parser.parse_args()
pprint(vars(args))
set_gpu(args.gpu)
if args.init_weights is not None:
is_pretrain = 'T'
else:
is_pretrain = 'F'
save_path1 = '-'.join([args.dataset, args.model_type, 'AVIATOR_benchmark', str(args.shot), str(args.way)])
save_path2 = '_'.join([str(args.gd_lr), str(args.lr), str(args.lr_mul), str(args.inner_iters),
str(args.temperature), str(args.step_size), str(args.gamma), is_pretrain])
args.save_path = osp.join(save_path1, save_path2)
ensure_path(save_path1, remove=False)
ensure_path(args.save_path)
if args.dataset == 'MiniImageNet':
# Handle MiniImageNet
from model.dataloader.mini_imagenet import MiniImageNet as Dataset
elif args.dataset == 'CUB':
from model.dataloader.cub import CUB as Dataset
else:
raise ValueError('Non-supported Dataset.')
trainset = Dataset('train', args)
train_sampler = CategoriesSampler(trainset.label, 100, args.way, args.shot + args.query)
train_loader = DataLoader(dataset=trainset, batch_sampler=train_sampler, num_workers=8, pin_memory=True)
valset = Dataset('val', args)
val_sampler = CategoriesSampler(valset.label, 500, args.way, args.shot + args.query)
val_loader = DataLoader(dataset=valset, batch_sampler=val_sampler, num_workers=8, pin_memory=True)
model = MAMLP_FC(args)
if args.init_weights is not None:
# get previous layers
basic_para = list()
top_para = list()
for e in model.named_parameters():
if 'encoder' not in e[0] or 'FC' in e[0]:
top_para.append(e[1])
else:
basic_para.append(e[1])
if args.model_type == 'ConvNet':
optimizer = torch.optim.Adam([{'params': basic_para},
{'params': top_para, 'lr': args.lr * args.lr_mul}], lr=args.lr)
elif args.model_type == 'ResNet':
optimizer = torch.optim.SGD([{'params': basic_para},
{'params': top_para, 'lr': args.lr * args.lr_mul}], lr=args.lr, momentum=0.9, nesterov=True, weight_decay=0.0005)
else:
raise ValueError('No Such Encoder')
else:
if args.model_type == 'ConvNet':
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
elif args.model_type == 'ResNet':
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.9, nesterov=True, weight_decay=0.0005)
else:
raise ValueError('No Such Encoder')
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.step_size, gamma=args.gamma)
# load pre-trained model (no FC weights)
model_dict = model.state_dict()
if args.init_weights is not None:
if args.model_type == 'ConvNet':
refined_dict = {}
pretrained_dict = torch.load(args.init_weights)['params']
for e in pretrained_dict:
if 'encoder' in e:
new_name = e.split('.')
new_name = 'encoder.' + '_'.join(new_name[1:3]) + '.' + new_name[-1]
refined_dict[new_name] = pretrained_dict[e]
pretrained_dict = {k: v for k, v in refined_dict.items() if k in model_dict}
print(pretrained_dict.keys())
model_dict.update(pretrained_dict)
elif args.model_type == 'ResNet':
pretrained_dict = torch.load(args.init_weights)['params']
# remove the FC layer
pretrained_dict = {k: v for k, v in pretrained_dict.items() if 'FC' not in k}
pretrained_dict = {'encoder.' + k: v for k, v in pretrained_dict.items() if 'encoder.' + k in model_dict}
print(pretrained_dict.keys())
model_dict.update(pretrained_dict)
else:
raise ValueError('No Such Encoder')
model.load_state_dict(model_dict)
# record the index of running mean and variance
running_dict = {}
for e in model_dict:
if 'running' in e:
key_name = '.'.join(e.split('.')[1:-1])
if key_name in running_dict:
continue
else:
running_dict[key_name] = {}
# find the position of BN modules
component = model.encoder
for att in key_name.split('.'):
if att.isdigit():
component = component[int(att)]
else:
component = getattr(component, att)
running_dict[key_name]['mean'] = component.running_mean
running_dict[key_name]['var'] = component.running_var
if torch.cuda.is_available():
torch.backends.cudnn.benchmark = True
model = model.cuda()
def save_model(name):
torch.save(dict(params=model.state_dict()), osp.join(args.save_path, name + '.pth'))
trlog = {}
trlog['args'] = vars(args)
trlog['train_loss'] = []
trlog['val_loss'] = []
trlog['train_acc'] = []
trlog['val_acc'] = []
trlog['max_acc'] = 0.0
trlog['val_train_acc'] = 0.0
trlog['min_loss'] = 10000
trlog['val_train_loss'] = 10000
trlog['max_acc_epoch'] = 0
timer = Timer()
global_count = 0
writer = SummaryWriter(log_dir=args.save_path)
label = torch.arange(args.way).repeat(args.query)
if torch.cuda.is_available():
label = label.type(torch.cuda.LongTensor)
else:
label = label.type(torch.LongTensor)
for epoch in range(1, args.max_epoch + 1):
lr_scheduler.step()
model.train()
tl = Averager()
ta = Averager()
for i, batch in enumerate(train_loader, 1):
global_count = global_count + 1
if torch.cuda.is_available():
data, _ = [_.cuda() for _ in batch]
else:
data = batch[0]
p = args.shot * args.way
data_shot, data_query = data[:p], data[p:]
logits = model(data_shot, data_query) # KqN x KN x 1
# compute loss
loss = F.cross_entropy(logits, label)
acc = count_acc(logits, label)
writer.add_scalar('data/loss', float(loss), global_count)
writer.add_scalar('data/acc', float(acc), global_count)
if (i-1) % 50 == 0:
print('epoch {}, train {}/{}, loss={:.4f} acc={:.4f}'
.format(epoch, i, len(train_loader), loss.item(), acc))
tl.add(loss.item())
ta.add(acc)
optimizer.zero_grad()
loss.backward()
optimizer.step()
tl = tl.item()
ta = ta.item()
vl = Averager()
va = Averager()
print('best epoch {}, best val acc={:.4f}'.format(trlog['max_acc_epoch'], trlog['max_acc']))
model.eval()
model.encoder.is_training = True
# record the runing mean and variance before validation
for e in running_dict:
running_dict[e]['mean_copy'] = deepcopy(running_dict[e]['mean'])
running_dict[e]['var_copy'] = deepcopy(running_dict[e]['var'])
for i, batch in enumerate(val_loader, 1):
if torch.cuda.is_available():
data, _ = [_.cuda() for _ in batch]
else:
data = batch[0]
p = args.shot * args.way
data_shot, data_query = data[:p], data[p:]
logits = model(data_shot, data_query) # KqN x KN x 1
loss = F.cross_entropy(logits, label)
acc = count_acc(logits, label)
vl.add(loss.item())
va.add(acc)
# reset the running mean and variance
for e in running_dict:
running_dict[e]['mean'] = deepcopy(running_dict[e]['mean_copy'])
running_dict[e]['var'] = deepcopy(running_dict[e]['var_copy'])
vl = vl.item()
va = va.item()
writer.add_scalar('data/val_loss', float(vl), epoch)
writer.add_scalar('data/val_acc', float(va), epoch)
print('epoch {}, val, loss={:.4f} acc={:.4f}'.format(epoch, vl, va))
if va > trlog['max_acc']:
trlog['max_acc'] = va
trlog['min_loss'] = vl
trlog['max_acc_epoch'] = epoch
save_model('max_acc')
trlog['train_loss'].append(tl)
trlog['train_acc'].append(ta)
trlog['val_loss'].append(vl)
trlog['val_acc'].append(va)
torch.save(trlog, osp.join(args.save_path, 'trlog'))
save_model('epoch-last')
print('ETA:{}/{}'.format(timer.measure(), timer.measure(epoch / args.max_epoch)))
writer.close()
# Test Phase
# compute the training loss
trlog = torch.load(osp.join(args.save_path, 'trlog'))
val_train_sampler = CategoriesSampler(trainset.label, 500, args.way, args.shot + args.query)
val_train_loader = DataLoader(dataset=trainset, batch_sampler=val_train_sampler, num_workers=8, pin_memory=True)
model.load_state_dict(torch.load(osp.join(args.save_path, 'max_acc' + '.pth'))['params'])
model.eval()
model.encoder.is_training = True
vtl = Averager()
vta = Averager()
# record the runing mean and variance before validation
for e in running_dict:
running_dict[e]['mean_copy'] = deepcopy(running_dict[e]['mean'])
running_dict[e]['var_copy'] = deepcopy(running_dict[e]['var'])
for i, batch in enumerate(val_train_loader, 1):
if torch.cuda.is_available():
data, _ = [_.cuda() for _ in batch]
else:
data = batch[0]
p = args.shot * args.way
data_shot, data_query = data[:p], data[p:]
logits = model(data_shot, data_query) # KqN x KN x 1
loss = F.cross_entropy(logits, label)
acc = count_acc(logits, label)
vtl.add(loss.item())
vta.add(acc)
# reset the running mean and variance
for e in running_dict:
running_dict[e]['mean'] = deepcopy(running_dict[e]['mean_copy'])
running_dict[e]['var'] = deepcopy(running_dict[e]['var_copy'])
vtl = vtl.item()
vta = vta.item()
print('val_train, loss={:.4f} acc={:.4f}'.format(vtl, vta))
trlog['val_train_acc'] = vta
trlog['val_train_loss'] = vtl
# test phase
test_set = Dataset('test', args)
sampler = CategoriesSampler(test_set.label, 10000, args.way, args.shot + args.query)
loader = DataLoader(test_set, batch_sampler=sampler, num_workers=8, pin_memory=True)
basic_inner_step = args.inner_iters
test_acc = []
for iter_mul in [1,2,3]:
# try finetune with different step sizes
test_acc_record = np.zeros((10000,))
args.inner_iters = basic_inner_step * iter_mul
model.load_state_dict(torch.load(osp.join(args.save_path, 'max_acc' + '.pth'))['params'])
model.eval()
model.encoder.is_training = True
# record the runing mean and variance before validation
for e in running_dict:
running_dict[e]['mean_copy'] = deepcopy(running_dict[e]['mean'])
running_dict[e]['var_copy'] = deepcopy(running_dict[e]['var'])
ave_acc = Averager()
label = torch.arange(args.way).repeat(args.query)
if torch.cuda.is_available():
label = label.type(torch.cuda.LongTensor)
else:
label = label.type(torch.LongTensor)
for i, batch in enumerate(loader, 1):
if torch.cuda.is_available():
data, _ = [_.cuda() for _ in batch]
else:
data = batch[0]
p = args.shot * args.way
data_shot, data_query = data[:p], data[p:]
logits = model(data_shot, data_query) # KqN x KN x 1
acc = count_acc(logits, label)
ave_acc.add(acc)
test_acc_record[i-1] = acc
print('batch {}: {:.2f}({:.2f})'.format(i, ave_acc.item() * 100, acc * 100))
# reset the running mean and variance
for e in running_dict:
running_dict[e]['mean'] = deepcopy(running_dict[e]['mean_copy'])
running_dict[e]['var'] = deepcopy(running_dict[e]['var_copy'])
m, pm = compute_confidence_interval(test_acc_record)
test_acc.append((m,pm))
# show results for different iterations
print('Val Best Epoch {}, Best Val Acc {:.5f}, Val Loss {:.5f}'.format(trlog['max_acc_epoch'], trlog['max_acc'], trlog['min_loss']))
print('Val Best Epoch {}, Best Train Acc {:.5f}, Train Loss {:.5f}'.format(trlog['max_acc_epoch'], trlog['val_train_acc'], trlog['val_train_loss']))
for i, iter_mul in enumerate([1,2,3]):
print('Inner Iter {}, Test Acc {:.5f} + {:.5f}'.format(basic_inner_step * iter_mul, test_acc[i][0], test_acc[i][1]))
import pdb
pdb.set_trace()
with open(args.comment+'.txt', 'w') as f:
# save temp results
f.write(','.join([str(trlog['max_acc_epoch']), str(trlog['max_acc']), str(trlog['min_loss']),
str(trlog['val_train_acc']), str(trlog['val_train_loss']),
'{:.5f} + {:.5f}'.format(test_acc[0][0], test_acc[0][1]),
'{:.5f} + {:.5f}'.format(test_acc[1][0], test_acc[1][1]),
'{:.5f} + {:.5f}'.format(test_acc[2][0], test_acc[2][1])]))