-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain.py
153 lines (132 loc) · 6.33 KB
/
train.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
import mxnet as mx
from mxnet import gluon
from mxnet import autograd
from caps_net import CapsNet, margin_loss, mask_mse_loss
import time
from metric import LossMetric
from visdom import Visdom
import numpy as np
viz = Visdom()
class SimpleLRScheduler(mx.lr_scheduler.LRScheduler):
"""A simple lr schedule that simply return `dynamic_lr`. We will set `dynamic_lr`
dynamically based on performance on the validation set.
"""
def __init__(self, learning_rate=0.001):
super(SimpleLRScheduler, self).__init__()
self.learning_rate = learning_rate
def __call__(self, num_update):
return self.learning_rate
def train_mnist(epochs, input_shape, n_class, num_routing, recon_loss_weight, ctx = mx.gpu(0), log_interval=20, **kwargs):
batch_size, C, H, W = input_shape
capsnet = CapsNet(n_class, num_routing, input_shape)
capsnet.initialize(init=mx.init.Xavier(), ctx=ctx)
capsnet.hybridize()
#mnist = mx.test_utils.get_mnist()
#train_iter = mx.io.NDArrayIter(mnist['train_data'], mnist['train_label'], batch_size, shuffle=True)
#val_iter = mx.io.NDArrayIter(mnist['test_data'], mnist['test_label'], batch_size)
train_iter = mx.io.MNISTIter(image="data/train-images.idx3-ubyte",
label="data/train-labels.idx1-ubyte",
batch_size=batch_size, shuffle=True)
val_iter = mx.io.MNISTIter(image="data/t10k-images.idx3-ubyte",
label="data/t10k-labels.idx1-ubyte",
batch_size=batch_size, shuffle=False)
draw_num = 32
draw_batch = val_iter.next()
draw_data = draw_batch.data[0].as_in_context(ctx)
draw_label = draw_batch.label[0].as_in_context(ctx)
draw_label = mx.nd.one_hot(draw_label, n_class)
learning_rate = 0.001
lr_scheduler = SimpleLRScheduler(learning_rate)
decay = 0.9
trainer = gluon.Trainer(capsnet.collect_params(),
optimizer='adam',
optimizer_params = {'lr_scheduler': lr_scheduler})
train_plt = viz.line(Y=np.zeros((1,3)),
X=np.zeros((1,3)),
opts=dict(
xlabel='Batch',
ylabel='Loss and Acc',
title='CapsNet traning plot',
legend=['Accuracy', 'Digit Loss', 'Mask Loss']
))
val_plt = viz.line(Y=np.zeros((1,3)),
X=np.zeros((1,3)),
opts=dict(
xlabel='Epoch',
ylabel='Loss and Acc',
title='CapsNet validation plot',
legend=['Accuracy', 'Digit Loss', 'Mask Loss']
))
mask_plt = viz.images(
np.random.randn(draw_num*2, 1, 28, 28),
opts=dict(title='Mask images', caption='Mask'))
hist_acc = 0
loss_metric = LossMetric(batch_size, 1)
val_metric = LossMetric(batch_size, 1)
batches_one_epoch = 60000 / batch_size
for epoch in range(epochs):
train_iter.reset()
val_iter.reset()
loss_metric.reset()
for i, batch in enumerate(train_iter):
tic = time.time()
x = batch.data[0].as_in_context(ctx)
y = batch.label[0].as_in_context(ctx)
y_ori = y
y = mx.nd.one_hot(y, n_class)
with autograd.record():
out_caps, out_mask = capsnet(x, y)
margin_loss_ = margin_loss(mx.nd, y, out_caps)
mask_loss_ = mask_mse_loss(mx.nd, x, out_mask)
loss = (1-recon_loss_weight)*margin_loss_ + recon_loss_weight*mask_loss_
loss.backward()
trainer.step(batch_size)
loss_metric.update([y_ori], [out_caps, loss, mask_loss_])
if i % log_interval == 0:
acc, digit_loss, mask_loss = loss_metric.get_name_value()
viz.line(Y=np.array([acc, digit_loss, mask_loss]).reshape((1,3)),
X=np.ones((1,3))*batches_one_epoch*epoch+i,
win=train_plt,
update='append')
take_num = min(draw_num, batch_size)
pred_label, pred_mask = capsnet(draw_data, draw_label)
draw = np.concatenate([draw_data[:take_num].asnumpy(), pred_mask[:take_num].asnumpy()])
viz.images(
draw,
win=mask_plt
)
elasp = time.time()-tic
print 'Epoch %2d, train %s %.5f, time %.1f sec, %d samples/s' % (epoch, "acc", acc, elasp, int(batch_size/elasp))
lr_scheduler.learning_rate = learning_rate * (decay ** (epoch+1))
val_metric.reset()
for i, batch in enumerate(val_iter):
x = batch.data[0].as_in_context(ctx)
y = batch.label[0].as_in_context(ctx)
y_ori = y
y = mx.nd.one_hot(y, n_class)
out_caps, out_mask = capsnet(x, y)
margin_loss_ = margin_loss(mx.nd, y, out_caps)
mask_loss_ = mask_mse_loss(mx.nd, x, out_mask)
loss = (1-recon_loss_weight)*margin_loss_ + recon_loss_weight*mask_loss_
val_metric.update([y_ori], [out_caps, loss, mask_loss_])
acc, digit_loss, mask_loss = val_metric.get_name_value()
viz.line(Y=np.array([acc, digit_loss, mask_loss]).reshape((1,3)),
X=np.ones((1,3))*epoch,
win=val_plt,
update='append')
if acc > hist_acc:
hist_acc = acc
capsnet.save_params("model/capsnet_%f.params"%acc)
print 'Epoch %2d, validation %s %.5f' % (epoch, "acc", acc)
if __name__ == "__main__":
from easydict import EasyDict as edict
params = edict()
params.epochs = 100
params.batch_size = 80
params.input_shape = (params.batch_size, 1, 28, 28)
params.n_class = 10
params.num_routing = 3
params.recon_loss_weight = 0.392
params.log_interval = 1
params.ctx = mx.gpu(0)
train_mnist(**params)