forked from markdtw/meta-learning-lstm-pytorch
/
main.py
284 lines (241 loc) · 11 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
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
#!/home/miranda9/.conda/envs/automl-meta-learning_wmlce-v1.7.0-py3.7/bin/python3.7
#SBATCH --job-name="miranda9job"
#SBATCH --output="demo.%j.%N.out"
#SBATCH --error="demo.%j.%N.err"
#SBATCH --partition=gpu
#SBATCH --time=24:00:00
#SBATCH --nodes=1
#SBATCH --sockets-per-node=1
#SBATCH --cores-per-socket=8
#SBATCH --threads-per-core=4
#SBATCH --mem-per-cpu=1200
#SBATCH --export=ALL
#SBATCH --gres=gpu:1
#SBATCH --mail-user=brando.science@gmail.com
#SBATCH --mail-type=ALL
from __future__ import division, print_function, absolute_import
import os
import pdb
import copy
import random
import argparse
import torch
import torch.nn as nn
import numpy as np
from tqdm import tqdm
from learner import Learner
from metalearner import MetaLearner
from dataloader import prepare_data
from utils import *
from pathlib import Path
from pdb import set_trace as st
FLAGS = argparse.ArgumentParser()
FLAGS.add_argument('--mode', choices=['train', 'test'])
# Hyper-parameters
FLAGS.add_argument('--n-shot', type=int,
help="How many examples per class for training (k, n_support)")
FLAGS.add_argument('--n-eval', type=int,
help="How many examples per class for evaluation (n_query)")
FLAGS.add_argument('--n-class', type=int,
help="How many classes (N, n_way)")
FLAGS.add_argument('--input-size', type=int,
help="Input size for the first LSTM")
FLAGS.add_argument('--hidden-size', type=int,
help="Hidden size for the first LSTM")
FLAGS.add_argument('--lr', type=float,
help="Learning rate")
FLAGS.add_argument('--episode', type=int,
help="Episodes to train")
FLAGS.add_argument('--episode-val', type=int,
help="Episodes to eval")
FLAGS.add_argument('--epoch', type=int,
help="Epoch to train for an episode")
FLAGS.add_argument('--batch-size', type=int,
help="Batch size when training an episode")
FLAGS.add_argument('--image-size', type=int,
help="Resize image to this size")
FLAGS.add_argument('--grad-clip', type=float,
help="Clip gradients larger than this number")
FLAGS.add_argument('--bn-momentum', type=float,
help="Momentum parameter in BatchNorm2d")
FLAGS.add_argument('--bn-eps', type=float,
help="Eps parameter in BatchNorm2d")
# Paths
FLAGS.add_argument('--data', choices=['miniimagenet'],
help="Name of dataset")
FLAGS.add_argument('--data-root', type=str,
help="Location of data")
FLAGS.add_argument('--resume', type=str,
help="Location to pth.tar")
FLAGS.add_argument('--save', type=str, default='logs',
help="Location to logs and ckpts")
# Others
FLAGS.add_argument('--cpu', action='store_true',
help="Set this to use CPU, default use CUDA")
FLAGS.add_argument('--n-workers', type=int, default=4,
help="How many processes for preprocessing")
FLAGS.add_argument('--pin-mem', type=bool, default=False,
help="DataLoader pin_memory")
FLAGS.add_argument('--log-freq', type=int, default=100,
help="Logging frequency")
FLAGS.add_argument('--val-freq', type=int, default=1000,
help="Validation frequency")
FLAGS.add_argument('--seed', type=int,
help="Random seed")
def meta_test(eps, eval_loader, learner_w_grad, learner_wo_grad, metalearner, args, logger):
for subeps, (episode_x, episode_y) in enumerate(tqdm(eval_loader, ascii=True)):
train_input = episode_x[:, :args.n_shot].reshape(-1, *episode_x.shape[-3:]).to(args.dev) # [n_class * n_shot, :]
train_target = torch.LongTensor(np.repeat(range(args.n_class), args.n_shot)).to(args.dev) # [n_class * n_shot]
test_input = episode_x[:, args.n_shot:].reshape(-1, *episode_x.shape[-3:]).to(args.dev) # [n_class * n_eval, :]
test_target = torch.LongTensor(np.repeat(range(args.n_class), args.n_eval)).to(args.dev) # [n_class * n_eval]
# Train learner with metalearner
learner_w_grad.reset_batch_stats()
learner_wo_grad.reset_batch_stats()
learner_w_grad.train()
learner_wo_grad.eval()
cI = train_learner(learner_w_grad, metalearner, train_input, train_target, args)
learner_wo_grad.transfer_params(learner_w_grad, cI)
output = learner_wo_grad(test_input)
loss = learner_wo_grad.criterion(output, test_target)
acc = accuracy(output, test_target)
logger.log_batch_info(loss=loss.item(), acc=acc, phase='eval')
return logger.log_batch_info(eps=eps, totaleps=args.episode_val, phase='evaldone')
def train_learner(learner_w_grad, metalearner, train_input, train_target, args):
cI = metalearner.metalstm.cI.data
hs = [None]
for _ in range(args.epoch):
for i in range(0, len(train_input), args.batch_size):
x = train_input[i:i+args.batch_size]
y = train_target[i:i+args.batch_size]
# get the loss/grad
learner_w_grad.copy_flat_params(cI)
output = learner_w_grad(x)
loss = learner_w_grad.criterion(output, y)
acc = accuracy(output, y)
learner_w_grad.zero_grad()
loss.backward()
# grad = torch.cat([w.grad.data.view(-1) / args.batch_size for w in learner_w_grad.parameters()], 0)
grad = []
for w in learner_w_grad.parameters():
g = w.grad.data.view(-1) / args.batch_size
grad.append(g)
grad = torch.cat(grad,0)
# preprocess grad & loss and metalearner forward
grad_prep = preprocess_grad_loss(grad) # [n_learner_params, 2]
loss_prep = preprocess_grad_loss(loss.data.unsqueeze(0)) # [1, 2]
metalearner_input = [loss_prep, grad_prep, grad.unsqueeze(1)]
cI, h = metalearner(metalearner_input, hs[-1])
hs.append(h)
#print("training loss: {:8.6f} acc: {:6.3f}, mean grad: {:8.6f}".format(loss, acc, torch.mean(grad)))
return cI
def brandos_load(args):
args.mode = "train"
args.n_shot = 5
args.n_eval = 15
args.n_class = 5
args.input_size = 4
args.hidden_size = 20
args.lr = 1e-3
args.episode = 50000
args.episode_val = 100
args.epoch = 8
args.batch_size = 25 # N*K = 5*5
args.image_size = 84
args.grad_clip = 0.25
args.bn_momentum = 0.95
args.bn_eps = 1e-3
args.data = "miniimagenet"
# args.data_root = Path('~/data/miniimagenet_meta_lstm/miniImagenet/').expanduser()
args.data_root = Path('~/data/miniimagenet_meta_lstm/miniImagenet/').expanduser()
args.pin_mem = True
args.log_freq = 50
args.val_freq = 10
args.cpu = True
args.seed = None
args.save = Path('~/data/meta_lstm_logs/').expanduser()
args.save.mkdir(parents=True, exist_ok=True)
args.save = str(args.save)
return args
def main():
args, unparsed = FLAGS.parse_known_args()
args = brandos_load(args)
if len(unparsed) != 0:
raise NameError("Argument {} not recognized".format(unparsed))
if args.seed is None:
args.seed = random.randint(0, 1e3)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
#args.dev = torch.device('cpu')
if args.cpu:
args.dev = torch.device('cpu')
args.gpu_name = args.dev
else:
if not torch.cuda.is_available():
raise RuntimeError("GPU unavailable.")
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
args.dev = torch.device('cuda')
try:
args.gpu_name = torch.cuda.get_device_name(0)
except:
args.gpu_name = args.dev
print(f'device {args.dev}')
logger = GOATLogger(args)
# Get data
train_loader, val_loader, test_loader = prepare_data(args)
# Set up learner, meta-learner
learner_w_grad = Learner(args.image_size, args.bn_eps, args.bn_momentum, args.n_class).to(args.dev)
learner_wo_grad = copy.deepcopy(learner_w_grad)
metalearner = MetaLearner(args.input_size, args.hidden_size, learner_w_grad.get_flat_params().size(0)).to(args.dev)
metalearner.metalstm.init_cI(learner_w_grad.get_flat_params())
# Set up loss, optimizer, learning rate scheduler
optim = torch.optim.Adam(metalearner.parameters(), args.lr)
if args.resume:
logger.loginfo("Initialized from: {}".format(args.resume))
last_eps, metalearner, optim = resume_ckpt(metalearner, optim, args.resume, args.dev)
if args.mode == 'test':
_ = meta_test(last_eps, test_loader, learner_w_grad, learner_wo_grad, metalearner, args, logger)
return
best_acc = 0.0
logger.loginfo("---> Start training")
# Meta-training
for eps, (episode_x, episode_y) in enumerate(train_loader): # sample data set split episode_x = D = (D^{train},D^{test})
print(f'episode = {eps}')
#print(f'episode_y = {episode_y}')
# print(f'episide_x.size() = {episode_x.size()}') # episide_x.size() = torch.Size([5, 20, 3, 84, 84]) i.e. N classes for K shot task with K_eval query examples
# print(f'episode_x.mean() = {episode_x.mean()}')
# episode_x.shape = [n_class, n_shot + n_eval, c, h, w]
# episode_y.shape = [n_class, n_shot + n_eval] --> NEVER USED
train_input = episode_x[:, :args.n_shot].reshape(-1, *episode_x.shape[-3:]).to(args.dev) # [n_class * n_shot, :]
train_target = torch.LongTensor(np.repeat(range(args.n_class), args.n_shot)).to(args.dev) # [n_class * n_shot]
test_input = episode_x[:, args.n_shot:].reshape(-1, *episode_x.shape[-3:]).to(args.dev) # [n_class * n_eval, :]
test_target = torch.LongTensor(np.repeat(range(args.n_class), args.n_eval)).to(args.dev) # [n_class * n_eval]
# Train learner with metalearner
learner_w_grad.reset_batch_stats()
learner_wo_grad.reset_batch_stats()
learner_w_grad.train()
learner_wo_grad.train()
cI = train_learner(learner_w_grad, metalearner, train_input, train_target, args)
# Train meta-learner with validation loss
learner_wo_grad.transfer_params(learner_w_grad, cI)
output = learner_wo_grad(test_input)
loss = learner_wo_grad.criterion(output, test_target)
acc = accuracy(output, test_target)
optim.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(metalearner.parameters(), args.grad_clip)
optim.step()
logger.batch_info(eps=eps, totaleps=args.episode, loss=loss.item(), acc=acc, phase='train')
# Meta-validation
if eps % args.val_freq == 0 and eps != 0:
save_ckpt(eps, metalearner, optim, args.save)
acc = meta_test(eps, val_loader, learner_w_grad, learner_wo_grad, metalearner, args, logger)
if acc > best_acc:
best_acc = acc
logger.loginfo(f"* Best accuracy so far {acc}*\n")
logger.loginfo(f'acc: {acc}')
logger.loginfo(f"* Best accuracy so far {acc}*\n")
logger.loginfo("Done")
if __name__ == '__main__':
main()