MXNet动手学深度学习笔记：VGG神经网络实现

1#coding:utf-8
2'''
3VGG网络
4'''
5from mxnet.gluon import nn
6from mxnet import ndarray as nd
7import mxnet as mx
8from mxnet import init
9import os
10import sys
11sys.path.append(os.getcwd())
12import utils
13
14def vgg_block(num_convs,channels):
15    out = nn.Sequential()
16    for _ in range(num_convs):
17        out.add(
18            nn.Conv2D(channels=channels,kernel_size=3,padding=1,activation='relu')
19        )
20
21    out.add(nn.MaxPool2D(pool_size=2,strides=2))
22
23    return out
24
25# 将 vgg_block堆起来
26def vgg_stack(arhitechure):
27    out = nn.Sequential()
28    for(num_convs,channels) in arhitechure:
29        out.add(vgg_block(num_convs,channels))
30
31    return out
32
33blk = vgg_block(2,128)
34blk.initialize()
35
36x = nd.random.uniform(shape=(2,3,16,16))
37y = blk(x)
38print(y.shape)
39
40# 定义一个最简单的VGG结构，8个卷积层，3个全连接层，称为VGG11
41ctx = mx.cpu() #utils.try_gpu()
42num_outputs = 10
43architechure = ((1,64),(1,128),(2,256),(2,512),(2,512))
44net = nn.Sequential()
45with net.name_scope():
46    net.add(
47        vgg_stack(architechure),
48        nn.Flatten(),
49        nn.Dense(4096,activation='relu'),
50        nn.Dropout(0.5),
51        nn.Dense(4096,activation='relu'),
52        nn.Dropout(0.5),
53        nn.Dense(num_outputs)
54    )
55
56# 训练模型
57train_data,test_data = utils.load_data_fashion_mnist(batch_size=64,
58            resize=96)
59ctx = utils.try_gpu()
60net.initialize(ctx=ctx,init=init.Xavier())
61
62loss = mx.gluon.loss.SoftmaxCrossEntropyLoss()
63trainer = mx.gluon.Trainer(net.collect_params(),'sgd',{'learning_rate':0.05})
64
65utils.train(train_data,test_data,net,loss,trainer,ctx,num_epochs=5)