-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
executable file
·192 lines (152 loc) · 10.1 KB
/
main.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
import os
import copy
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm
from torchmeta.utils.data import BatchMetaDataLoader
from maml.utils import load_dataset, load_model, update_parameters, get_accuracy
def project_vec(u, v):
utu = sum(u*u)
utv = sum(u*v)
return (utv/utu)*u
def gs(X):
# Gram-Schmidt function
Q = np.zeros(X.shape, dtype=X.dtype)
Q[0] = X[0]
Q[1] = X[1] - project_vec(Q[0], X[1])
Q[2] = X[2] - project_vec(Q[0], X[2]) - project_vec(Q[1], X[2])
Q[3] = X[3] - project_vec(Q[0], X[3]) - project_vec(Q[1], X[3]) - project_vec(Q[2], X[3])
Q[4] = X[4] - project_vec(Q[0], X[4]) - project_vec(Q[1], X[4]) - project_vec(Q[2], X[4]) - project_vec(Q[3], X[4])
Q = torch.tensor(Q).type(torch.FloatTensor)
Q = Q / torch.norm(Q, dim=1, keepdim=True)
return Q
def main(args, mode, iteration=None):
dataset = load_dataset(args, mode)
dataloader = BatchMetaDataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
model.to(device=args.device)
model.train()
# To control outer update parameter
# If you want to control inner update parameter, please see update_parameters function in ./maml/utils.py
freeze_params = [p for name, p in model.named_parameters() if 'classifier' in name]
learnable_params = [p for name, p in model.named_parameters() if 'classifier' not in name]
if args.outer_fix:
meta_optimizer = torch.optim.Adam([{'params': freeze_params, 'lr': 0},
{'params': learnable_params, 'lr': args.meta_lr}])
else:
meta_optimizer = torch.optim.Adam([{'params': freeze_params, 'lr': args.meta_lr},
{'params': learnable_params, 'lr': args.meta_lr}])
if args.meta_train:
total = args.train_batches
elif args.meta_val:
total = args.valid_batches
elif args.meta_test:
total = args.test_batches
loss_logs, accuracy_logs = [], []
# Training loop
with tqdm(dataloader, total=total, leave=False) as pbar:
for batch_idx, batch in enumerate(pbar):
if args.centering:
fc_weight_mean = torch.mean(model.classifier.weight.data, dim=0)
model.classifier.weight.data -= fc_weight_mean
model.zero_grad()
support_inputs, support_targets = batch['train']
support_inputs = support_inputs.to(device=args.device)
support_targets = support_targets.to(device=args.device)
query_inputs, query_targets = batch['test']
query_inputs = query_inputs.to(device=args.device)
query_targets = query_targets.to(device=args.device)
outer_loss = torch.tensor(0., device=args.device)
accuracy = torch.tensor(0., device=args.device)
for task_idx, (support_input, support_target, query_input, query_target) in enumerate(zip(support_inputs, support_targets, query_inputs, query_targets)):
support_features, support_logit = model(support_input)
inner_loss = F.cross_entropy(support_logit, support_target)
model.zero_grad()
params = update_parameters(model,
inner_loss,
extractor_step_size=args.extractor_step_size,
classifier_step_size=args.classifier_step_size,
first_order=args.first_order)
query_features, query_logit = model(query_input, params=params)
outer_loss += F.cross_entropy(query_logit, query_target)
with torch.no_grad():
accuracy += get_accuracy(query_logit, query_target)
outer_loss.div_(args.batch_size)
accuracy.div_(args.batch_size)
loss_logs.append(outer_loss.item())
accuracy_logs.append(accuracy.item())
if args.meta_train:
outer_loss.backward()
meta_optimizer.step()
postfix = {'mode': mode, 'iter': iteration, 'acc': round(accuracy.item(), 5)}
pbar.set_postfix(postfix)
if batch_idx+1 == total:
break
# Save model
if args.meta_train:
filename = os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'models', 'epochs_{}.pt'.format((iteration+1)*total))
if (iteration+1)*total % 5000 == 0:
with open(filename, 'wb') as f:
state_dict = model.state_dict()
torch.save(state_dict, f)
return loss_logs, accuracy_logs
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser('Model-Agnostic Meta-Learning (MAML)')
parser.add_argument('--folder', type=str, help='Path to the folder the data is downloaded to.')
parser.add_argument('--dataset', type=str, help='Dataset: miniimagenet, tieredimagenet, cub, cars, cifar_fs, fc100, aircraft, vgg_flower')
parser.add_argument('--model', type=str, help='Model: 4conv, resnet')
parser.add_argument('--device', type=str, default='cuda:0', help='gpu device')
parser.add_argument('--download', action='store_true', help='Download the dataset in the data folder.')
parser.add_argument('--num-shots', type=int, default=5, help='Number of examples per class (k in "k-shot", default: 5).')
parser.add_argument('--num-ways', type=int, default=5, help='Number of classes per task (N in "N-way", default: 5).')
parser.add_argument('--meta-lr', type=float, default=1e-3, help='Learning rate of meta optimizer.')
parser.add_argument('--first-order', action='store_true', help='Use the first-order approximation of MAML.')
parser.add_argument('--extractor-step-size', type=float, default=0.5, help='Extractor step-size for the gradient step for adaptation (default: 0.5).')
parser.add_argument('--classifier-step-size', type=float, default=0.5, help='Classifier step-size for the gradient step for adaptation (default: 0.5).')
parser.add_argument('--hidden-size', type=int, default=64, help='Number of channels for each convolutional layer (default: 64).')
parser.add_argument('--blocks-type', type=str, default=None, help='Resnet block type (optional).')
parser.add_argument('--output-folder', type=str, default='./output/', help='Path to the output folder for saving the model (optional).')
parser.add_argument('--save-name', type=str, default=None, help='Name of model (optional).')
parser.add_argument('--batch-size', type=int, default=4, help='Number of tasks in a mini-batch of tasks (default: 4).')
parser.add_argument('--batch-iter', type=int, default=300, help='Number of times to repeat train batches (i.e., total epochs = batch_iter * train_batches) (default: 300).')
parser.add_argument('--train-batches', type=int, default=100, help='Number of batches the model is trained over (i.e., validation save steps) (default: 100).')
parser.add_argument('--valid-batches', type=int, default=25, help='Number of batches the model is validated over (default: 25).')
parser.add_argument('--test-batches', type=int, default=2500, help='Number of batches the model is tested over (default: 2500).')
parser.add_argument('--num-workers', type=int, default=1, help='Number of workers for data loading (default: 1).')
parser.add_argument('--centering', action='store_true', help='Parallel shift operation in the head.')
parser.add_argument('--ortho-init', action='store_true', help='Use the head from the orthononal model.')
parser.add_argument('--outer-fix', action='store_true', help='Fix the head during outer updates.')
args = parser.parse_args()
os.makedirs(os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'logs'), exist_ok=True)
os.makedirs(os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'models'), exist_ok=True)
arguments_txt = ""
for k, v in args.__dict__.items():
arguments_txt += "{}: {}\n".format(str(k), str(v))
filename = os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'logs', 'arguments.txt')
with open(filename, 'w') as f:
f.write(arguments_txt[:-1])
args.device = torch.device(args.device)
model = load_model(args)
if args.ortho_init:
X = np.random.randn(5, 1600)
Q = gs(X)
model.classifier.weight.data = nn.Parameter(Q)
log_pd = pd.DataFrame(np.zeros([args.batch_iter*args.train_batches, 6]),
columns=['train_error', 'train_accuracy', 'valid_error', 'valid_accuracy', 'test_error', 'test_accuracy'])
for iteration in tqdm(range(args.batch_iter)):
meta_train_loss_logs, meta_train_accuracy_logs = main(args=args, mode='meta_train', iteration=iteration)
meta_valid_loss_logs, meta_valid_accuracy_logs = main(args=args, mode='meta_valid', iteration=iteration)
log_pd['train_error'][iteration*args.train_batches:(iteration+1)*args.train_batches] = meta_train_loss_logs
log_pd['train_accuracy'][iteration*args.train_batches:(iteration+1)*args.train_batches] = meta_train_accuracy_logs
log_pd['valid_error'][(iteration+1)*args.train_batches-1] = np.mean(meta_valid_loss_logs)
log_pd['valid_accuracy'][(iteration+1)*args.train_batches-1] = np.mean(meta_valid_accuracy_logs)
filename = os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'logs', 'logs.csv')
log_pd.to_csv(filename, index=False)
meta_test_loss_logs, meta_test_accuracy_logs = main(args=args, mode='meta_test')
log_pd['test_error'][args.batch_iter*args.train_batches-1] = np.mean(meta_test_loss_logs)
log_pd['test_accuracy'][args.batch_iter*args.train_batches-1] = np.mean(meta_test_accuracy_logs)
filename = os.path.join(args.output_folder, args.dataset+'_'+args.save_name, 'logs', 'logs.csv')
log_pd.to_csv(filename, index=False)