mxnet 实现deepid, 参考的是mxnet 自己提供的symbol_xyz.py, xyz 代表具体的网络名字, 比如说alexnet, googlenet...
实现的deepid 的代码如下:
import find_mxnet
import mxnet as mx
def get_symbol(num_classes = 530):
input_data = mx.symbol.Variable(name="data")
# stage 1
conv1 = mx.symbol.Convolution(
data=input_data, kernel=(4, 4), stride=(1, 1), num_filter=20)
relu1 = mx.symbol.Activation(data=conv1, act_type="relu")
pool1 = mx.symbol.Pooling(
data=relu1, pool_type="max", kernel=(2, 2), stride=(2,2))
# stage 2
conv2 = mx.symbol.Convolution(
data=pool1, kernel=(3, 3), stride=(1, 1), num_filter=40)
relu2 = mx.symbol.Activation(data=conv2, act_type="relu")
pool2 = mx.symbol.Pooling(data=relu2, kernel=(2, 2), stride=(1, 1), pool_type="max")
# stage 3
conv3 = mx.symbol.Convolution(
data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=60)
relu3 = mx.symbol.Activation(data=conv3, act_type="relu")
pool3 = mx.symbol.Pooling(data=relu3, kernel=(2, 2), stride=(2, 2), pool_type="max")
conv4 = mx.symbol.Convolution(
data=pool3, kernel=(2, 2), pad=(1, 1), num_filter=80)
relu4 = mx.symbol.Activation(data=conv4, act_type="relu")
# stage 4
fc160_1 = mx.symbol.FullyConnected(data=relu4, num_hidden=160)
fc160_2 = mx.symbol.FullyConnected(data=conv3, num_hidden=160)
dropout = mx.symbol.Dropout(data=fc160_2, p=0.5)
fc_class = mx.symbol.FullyConnected(data=dropout, num_hidden=530)
# stage 6
softmax = mx.symbol.SoftmaxOutput(data=fc_class, name='softmax')
return softmax
我觉得虽然有了代码, 但是不够直观, 一图胜千言啊, 所以可视化还是很需要的.
http://josephpcohen.com/w/visualizing-cnn-architectures-side-by-side-with-mxnet/
import find_mxnet
import mxnet as mx
import importlib
name = "inception-v3"
net = importlib.import_module("symbol_" + name).get_symbol(2)
a = mx.viz.plot_network(net, shape={"data":(1, 1, 299, 299)}, node_attrs={"shape":'rect',"fixedsize":'false'})
a.render(name)
刚才链接的博主, 自己还化了其他的很多图示. 大家可以看看, 因为图示太大了, 所以我就没有粘贴过来, 想要看的, 大家去他的网站吧.
digraph plot {
convolution0 [label="Convolution\n4x4/1, 20" fillcolor="#fb8072" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
activation0 [label="Activation\nrelu" fillcolor="#ffffb3" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
pooling0 [label="Pooling\nmax, 2x2/2" fillcolor="#80b1d3" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
convolution1 [label="Convolution\n3x3/1, 40" fillcolor="#fb8072" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
activation1 [label="Activation\nrelu" fillcolor="#ffffb3" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
pooling1 [label="Pooling\nmax, 2x2/1" fillcolor="#80b1d3" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
convolution2 [label="Convolution\n3x3/1, 60" fillcolor="#fb8072" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
fullyconnected1 [label="FullyConnected\n160" fillcolor="#fb8072" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
dropout0 [label=Dropout fillcolor="#fccde5" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
fullyconnected2 [label="FullyConnected\n530" fillcolor="#fb8072" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
softmax [label=SoftmaxOutput fillcolor="#fccde5" fixedsize=false height=0.8034 shape=rect style=filled width=1.3]
activation0 -> convolution0 [label="20x61x61" arrowtail=open dir=back]
pooling0 -> activation0 [label="20x61x61" arrowtail=open dir=back]
convolution1 -> pooling0 [label="20x31x31" arrowtail=open dir=back]
activation1 -> convolution1 [label="40x29x29" arrowtail=open dir=back]
pooling1 -> activation1 [label="40x29x29" arrowtail=open dir=back]
convolution2 -> pooling1 [label="40x28x28" arrowtail=open dir=back]
fullyconnected1 -> convolution2 [label="60x28x28" arrowtail=open dir=back]
dropout0 -> fullyconnected1 [label=160 arrowtail=open dir=back]
fullyconnected2 -> dropout0 [label=160 arrowtail=open dir=back]
softmax -> fullyconnected2 [label=530 arrowtail=open dir=back]
}
我将上面的dot 文件保存为 deepid.dot, 然后调用dot来画图
dot deepid.dot -Tpng -o deepid.png
deepid (2).png
这个网络在对LFW进行对齐之后, 训练精度为94%多一点.
补充
因为 deepid 算是一个比较中规中矩的 CNN 模型, 而且训练的方式也跟 CNN 如出一辙, 所以我这里用的还是 image-classification 的框架, 但是 deepid2 还有其他的网络, 有些变形, 可能不能直接套用这个框架.
网友评论