-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_maml_cifar100.py
191 lines (154 loc) · 6.26 KB
/
train_maml_cifar100.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
#!/usr/bin/env python
# coding: utf-8
"""
Usage:
python train_maml_cifar100.py \
--gpu_id 3 \
--max_meta_iter 1000 \
--k_shot 4 --n_way 4 \
--train_bs 32 --eval_bs 32 --max_iter_for_eval 10 \
--lr_meta 1e-3 --lr_task 0.1
"""
# # Train MAML in pytorch-lightning
# - 2022-01-13 (r)
import os, sys
from pathlib import Path
import argparse
from argparse import ArgumentParser
import torch
import pytorch_lightning as pl
from pytorch_lightning import loggers as pl_loggers
# Callbacks
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint
def main(args):
from reprlearn.utils.misc import info, now2str, today2str, get_next_version_path, n_iter_per_epoch
from reprlearn.data.datamodules.kshot_datamodule import KShotDataModule
# ## Set Path
# - DATA_ROOT:
# - Use '/data/hayley-old/Tenanbaum2000/data' for MNIST, Mono-MNIST, Rotated-MNIST, Teapots
# - Use `/data/hayley-old/maptiles_v2' folder for Maptile dataset
ROOT = Path('/data/hayley-old/Tenanbaum2000')
# ### Init DataModule
dataset_name = 'cifar100'
data_root = Path(args.data_root)
in_shape = args.in_shape #(32, 32, 3)
max_meta_iter = args.max_meta_iter #10000
k_shot = args.k_shot #4
n_way = args.n_way #5
num_tasks_per_iter_for_train = args.train_bs #16
num_tasks_per_iter_for_eval = args.eval_bs #16
max_iter_for_train = max_meta_iter # must be >= max-meta-iter
max_iter_for_eval = args.max_iter_for_eval #10 # totoal number of loss_q's to be averaged over is num_tasks_per_iter_for_eval * max_iter_for_eval
dm_config = {
'dataset_name': dataset_name,
'data_root': data_root,
'k_shot': k_shot,
'n_way': n_way,
'num_tasks_per_iter_for_train': num_tasks_per_iter_for_train,
'max_iter_for_train': max_iter_for_train,
'num_tasks_per_iter_for_eval': num_tasks_per_iter_for_eval,
'max_iter_for_eval': max_iter_for_eval,
}
dm = KShotDataModule(**dm_config)
print(dm.name)
# ### Init pl.Module
from reprlearn.models.plmodules.meta_learning import MAML
model_kwargs = {
'lr_meta': args.lr_meta,
'lr_task': args.lr_task,
'use_averaged_meta_loss': args.use_averaged_meta_loss,
'num_inner_steps': args.num_inner_steps,
'log_every': args.log_every,
}
# net = get_densenet(output_size=n_way)
net = None
model = MAML(
in_shape=args.in_shape,
k_shot=args.k_shot,
n_way=args.n_way,
net=net,
**model_kwargs,
)
print('lr_meta, lr_task', model.lr_meta, model.lr_task)
print('num_inner_steps: ', model.num_inner_steps)
# ### Init pl.Trainer
# Init. callbacks
# Model Checkpoint criterion
ckpt_metric, ckpt_mode = 'val/loss_q', 'min'
# ckpt_metric, ckpt_mode = 'loss', 'min'
ckpt_callback = ModelCheckpoint(
monitor=ckpt_metric,
mode=ckpt_mode,
save_top_k=5,
)
stop_metric, stop_mode = 'val/loss_q', 'min'
# stop_metric, stop_mode = 'loss', 'min'
stop_patience = 100
early_stopping_callback = EarlyStopping(
monitor=stop_metric,
patience=stop_patience,
mode=stop_mode,
)
callbacks = [ckpt_callback]
# callbacks = [ckpt_callback, early_stopping_callback]
# Init. Tensorboard logger
exp_name = f'{model.name}_{dm.name}'
tb_logger = pl_loggers.TensorBoardLogger(
save_dir=f'{ROOT}/lightning_logs/{today2str()}',
name=exp_name,
default_hp_metric=False, # todo: what is this param's effect?
)
log_dir = Path(tb_logger.log_dir)
print(tb_logger.log_dir)
if not log_dir.exists():
log_dir.mkdir(parents=True)
print("Created: ", log_dir)
# Init. pl.Trainer
trainer_config = {
'gpus': 1,
'max_epochs': args.max_meta_iter,
'progress_bar_refresh_rate': 0,
'terminate_on_nan': True,
'check_val_every_n_epoch': 1,
'logger': tb_logger,
'callbacks': callbacks,
}
# trainer = pl.Trainer(fast_dev_run=3) # for test/debug
trainer = pl.Trainer(**trainer_config) # for real training
# Fit model
trainer.fit(model, dm)
print(f"Finished at ep {trainer.current_epoch}")
if __name__ == '__main__':
parser = argparse.ArgumentParser(add_help=False) # add_help=False is important!
parser.add_argument('--gpu_id', type=str, required=True)
parser.add_argument('--max_meta_iter', type=int, required=True,
help="Number of meta-updates (aka. max_epochs)")
# DataModule parameters ##todo: make to kshot datamoudle's add-args method
parser.add_argument('--data_root', type=str, default='/data/hayley-old/Tenanbaum2000/data')
parser.add_argument('--in_shape', nargs=3, type=int, default=[32, 32, 3])
parser.add_argument('--k_shot', type=int, required=True)
parser.add_argument('--n_way', type=int, required=True)
parser.add_argument('--train_bs', type=int, required=True)
parser.add_argument('--eval_bs',type=int, required=True)
parser.add_argument('--max_iter_for_eval', type=int, required=True)
# Model params # todo: move to module's add_args method
parser.add_argument('--lr_meta', type=float, default=1e-3)
parser.add_argument('--lr_task', type=float, default=0.1)
parser.add_argument('--num_inner_steps', type=int, default=1)
# -- whether to use meta-loss as the average of the sum of the loss_q's in the
# -- batch of tasks
group = parser.add_mutually_exclusive_group(required=False)
group.add_argument('--use_averaged_meta_loss', dest='use_averaged_meta_loss', action='store_true')
group.add_argument('--no_use_averaged_meta_loss', dest='use_averaged_meta_loss', action='store_false')
parser.set_defaults(use_averaged_meta_loss=True)
parser.add_argument('--log_every', type=int, default=10)
args = parser.parse_args()
print("Final args: ")
# ------------------------------------------------------------------------
# Initialize model, datamodule, trainer using the parsered arg dict
# ------------------------------------------------------------------------
# Select Visible GPU
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
main(args)