/
mnist.py
205 lines (182 loc) · 7.73 KB
/
mnist.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
#!/usr/bin/env python2.7
import logging
import argparse
import mxnet as mx
import numpy as np
import matplotlib
matplotlib.use('Agg')
import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib.patheffects as PathEffects
from lsoftmax import LSoftmaxOp
logging.basicConfig(format="[%(asctime)s][%(levelname)s] %(message)s")
logger = logging.getLogger()
logger.setLevel(logging.INFO)
def get_symbol():
data = mx.sym.Variable('data')
label = mx.sym.Variable('softmax_label')
conv1 = mx.sym.Convolution(data=data, kernel=(5, 5), num_filter=32)
relu1 = mx.sym.Activation(data=conv1, act_type='relu')
pool1 = mx.sym.Pooling(data=relu1, kernel=(2, 2), stride=(2, 2), pool_type='max')
conv2 = mx.sym.Convolution(data=pool1, kernel=(5, 5), num_filter=64)
relu2 = mx.sym.Activation(data=conv2, act_type='relu')
pool2 = mx.sym.Pooling(data=relu2, kernel=(2, 2), stride=(2, 2), pool_type='max')
fc3 = mx.sym.FullyConnected(data=pool2, num_hidden=256)
relu3 = mx.sym.Activation(data=fc3, act_type='relu')
embedding = mx.sym.FullyConnected(data=relu3, num_hidden=2, name='embedding')
if not args.no_lsoftmax:
if args.op_impl == 'cpp':
fc4 = mx.sym.LSoftmax(data=embedding, label=label, num_hidden=10,
beta=args.beta, margin=args.margin, scale=args.scale,
beta_min=args.beta_min, verbose=True)
else:
fc4 = mx.sym.Custom(data=embedding, label=label, num_hidden=10,
beta=args.beta, margin=args.margin, scale=args.scale,
beta_min=args.beta_min, op_type='LSoftmax')
else:
fc4 = mx.sym.FullyConnected(data=embedding, num_hidden=10, no_bias=True)
softmax_loss = mx.sym.SoftmaxOutput(data=fc4, label=label)
return softmax_loss
def train():
ctx = mx.gpu(args.gpu) if args.gpu >=0 else mx.cpu()
train = mx.io.MNISTIter(
image='data/train-images-idx3-ubyte',
label='data/train-labels-idx1-ubyte',
input_shape=(1, 28, 28),
mean_r=128,
scale=1./128,
batch_size=args.batch_size,
shuffle=True)
val = mx.io.MNISTIter(
image='data/t10k-images-idx3-ubyte',
label='data/t10k-labels-idx1-ubyte',
input_shape=(1, 28, 28),
mean_r=128,
scale=1./128,
batch_size=args.batch_size)
symbol = get_symbol()
mod = mx.mod.Module(
symbol=symbol,
context=ctx,
data_names=('data',),
label_names=('softmax_label',))
num_examples = 60000
epoch_size = int(num_examples / args.batch_size)
optim_params = {
'learning_rate': args.lr,
'momentum': 0.9,
'wd': 0.0005,
'lr_scheduler': mx.lr_scheduler.FactorScheduler(step=10*epoch_size, factor=0.1),
}
mod.fit(train_data=train,
eval_data=val,
eval_metric=mx.metric.Accuracy(),
initializer=mx.init.Xavier(),
optimizer='sgd',
optimizer_params=optim_params,
num_epoch=args.num_epoch,
batch_end_callback=mx.callback.Speedometer(args.batch_size, 50),
epoch_end_callback=mx.callback.do_checkpoint(args.model_prefix))
def test():
ctx = mx.gpu(args.gpu) if args.gpu >=0 else mx.cpu()
val = mx.io.MNISTIter(
image='data/t10k-images-idx3-ubyte',
label='data/t10k-labels-idx1-ubyte',
input_shape=(1, 28, 28),
mean_r=128,
scale=1./128,
batch_size=1)
symbol, arg_params, aux_params = mx.model.load_checkpoint(args.model_prefix, args.num_epoch)
embedding = symbol.get_internals()['embedding_output']
mod = mx.mod.Module(
symbol=embedding,
context=ctx,
data_names=('data',))
mod.bind(data_shapes=[('data', (1, 1, 28, 28))], for_training=False)
mod.init_params(arg_params=arg_params, aux_params=aux_params)
embeds = []
labels = []
for preds, i_batch, batch in mod.iter_predict(val):
embeds.append(preds[0].asnumpy())
labels.append(batch.label[0].asnumpy())
embeds = np.vstack(embeds)
labels = np.hstack(labels)
# vis, plot code from https://github.com/pangyupo/mxnet_center_loss
num = len(labels)
names = dict()
for i in range(10):
names[i]=str(i)
palette = np.array(sns.color_palette("hls", 10))
f = plt.figure(figsize=(8, 8))
ax = plt.subplot(aspect='equal')
sc = ax.scatter(embeds[:,0], embeds[:,1], lw=0, s=40,
c=palette[labels.astype(np.int)])
ax.axis('off')
ax.axis('tight')
# We add the labels for each digit.
txts = []
for i in range(10):
# Position of each label.
xtext, ytext = np.median(embeds[labels == i, :], axis=0)
txt = ax.text(xtext, ytext, names[i])
txt.set_path_effects([
PathEffects.Stroke(linewidth=5, foreground="w"),
PathEffects.Normal()])
txts.append(txt)
margin = args.margin if not args.no_lsoftmax else 1
fname = 'mnist-lsoftmax-margin-%d.png'%margin
plt.savefig(fname)
def profile():
ctx = mx.gpu(args.gpu) if args.gpu >=0 else mx.cpu()
val = mx.io.MNISTIter(
image='data/t10k-images-idx3-ubyte',
label='data/t10k-labels-idx1-ubyte',
input_shape=(1, 28, 28),
mean_r=128,
scale=1./128,
batch_size=args.batch_size)
symbol = get_symbol()
mod = mx.mod.Module(
symbol=symbol,
context=ctx,
data_names=('data',),
label_names=('softmax_label',))
mod.bind(data_shapes=val.provide_data, label_shapes=val.provide_label, for_training=True)
mod.init_params(initializer=mx.init.Xavier())
# run a while
for nbatch, data_batch in enumerate(val):
mod.forward_backward(data_batch)
# profile
mx.profiler.profiler_set_config(mode='symbolic', filename='profile.json')
mx.profiler.profiler_set_state('run')
val.reset()
for nbatch, data_batch in enumerate(val):
mod.forward_backward(data_batch)
mx.profiler.profiler_set_state('stop')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1, help="gpu device")
parser.add_argument('--lr', type=float, default=0.01, help="learning rate")
parser.add_argument('--beta', type=float, default=1, help="beta in lsoftmax, same as lambda")
parser.add_argument('--beta-min', type=float, default=0, help="minimun beta")
parser.add_argument('--scale', type=float, default=1, help="beta scale for every mini-batch")
parser.add_argument('--batch-size', type=int, default=128, help="batch size")
parser.add_argument('--train', action='store_true', help="train mnist")
parser.add_argument('--test', action='store_true', help="test mnist and plot")
parser.add_argument('--no-lsoftmax', action='store_true', help="don't use lsoftmax layer")
parser.add_argument('--margin', type=int, default=2, help="lsoftmax margin")
parser.add_argument('--model-prefix', type=str, default='model/mnist', help="model predix")
parser.add_argument('--num-epoch', type=int, default=20, help="number of epoches to train")
parser.add_argument('--op-impl', type=str, choices=['py', 'cpp'], default='py', help="operator implementation")
parser.add_argument('--profile', action='store_true', help="do profile")
args = parser.parse_args()
print args
# check
if args.op_impl == 'cpp' and args.gpu < 0:
raise ValueError("LSoftmax in C++ currently only supports GPU")
if args.train:
train()
if args.test:
test()
if args.profile:
profile()