-
Notifications
You must be signed in to change notification settings - Fork 1
/
mnist_training.py
44 lines (33 loc) · 1.28 KB
/
mnist_training.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import logging, os
logging.basicConfig(level=logging.INFO)
from deepy.trainers import AdamTrainer, LearningRateAnnealer, FineTuningAdaGradTrainer
from deepy.dataset import BinarizedMnistDataset, MiniBatches
from core import DrawModel
model_path = os.path.join(os.path.dirname(__file__), "models", "mnist1.gz")
if __name__ == '__main__':
from argparse import ArgumentParser
ap = ArgumentParser()
ap.add_argument("--load", default="", help="pre-trained model path")
ap.add_argument("--finetune", action="store_true")
args = ap.parse_args()
model = DrawModel(image_width=28, image_height=28, attention_times=64)
if args.load:
model.load_params(args.load)
conf = {
"gradient_clipping": 10,
"learning_rate": LearningRateAnnealer.learning_rate(0.004),
"weight_l2": 0
}
# conf.avoid_nan = True
# from deepy import DETECT_NAN_MODE
# conf.theano_mode = DETECT_NAN_MODE
# TODO: Find out the problem causing NaN
if args.finetune:
trainer = FineTuningAdaGradTrainer(model, conf)
else:
trainer = AdamTrainer(model, conf)
mnist = MiniBatches(BinarizedMnistDataset(), batch_size=100)
trainer.run(mnist, controllers=[])
model.save_params(model_path)