[caffe学习笔记][03][生成配置文件]

说明：

caffe通过配置文件prototxt来描述网路结构，通过Python接口来生成网路配置文件比较简单。这里生成train.prototxt和test.prototxt，分别用于训练阶段和验证阶段。

步骤：

1.生成配置文件

touch create_train_val_prototxt.py

spyder create_train_val_prototxt.py

# -*- coding: utf-8 -*-
"""
yuandanfei Editor

This is a temporary script file.
"""

from caffe import layers as L, params as P, to_proto
path = '/home/yuandanfei/work/caffe/mnist/' #root path
train_lmdb = path + 'train_lmdb'           #train_lmdb path
test_lmdb = path + 'test_lmdb'             #test_lmdb path
mean_file = path + 'mean.binaryproto'      #mean.binaryproto path
train_proto = path + 'train.prototxt'      #train.prototxt path
test_proto = path + 'test.prototxt'        #test.prototxt path

def create_net(lmdb, batch_size, include_acc=False):
    #input layer
    data, label = L.Data(source=lmdb, backend=P.Data.LMDB, batch_size=batch_size, ntop=2, 
                         transform_param=dict(crop_size=28, mean_file=mean_file, mirror=True))
    #conv1 layer   n*c*w*h;c1=num_output;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1;if stride=1 and pad=(kernel_size-1)/2, then w1/h1=w0/h0;
    conv1 = L.Convolution(data, kernel_size=5, stride=1, pad=2, num_output=16, weight_filler=dict(type='xavier'))
    #reul1 layer
    relu1 = L.ReLU(conv1, in_place=True)
    #pool1 layer   n*c*w*h;c1=c0;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1;
    pool1 = L.Pooling(relu1, pool=P.Pooling.MAX, kernel_size=3, stride=2)
    #conv2 layer
    conv2 = L.Convolution(pool1, kernel_size=3, stride=1, pad=1, num_output=32, weight_filler=dict(type='xavier'))
    #relu2 layer
    relu2 = L.ReLU(conv2, in_place=True)
    #pool2 layer
    pool2 = L.Pooling(relu2, pool=P.Pooling.MAX, kernel_size=3, stride=2)
    #fc3 layer
    fc3 = L.InnerProduct(pool2, num_output=1024, weight_filler=dict(type='xavier'))
    #relu3 layer
    relu3 = L.ReLU(fc3, in_place=True)
    #drop3 layer
    drop3 = L.Dropout(relu3, in_place=True)
    #fc4 layer
    fc4 = L.InnerProduct(drop3, num_output=10, weight_filler=dict(type='xavier'))
    #softmax-loss layer
    loss = L.SoftmaxWithLoss(fc4, label)
    #accuracy layer
    if include_acc: #test
        acc = L.Accuracy(fc4, label)
        return to_proto(loss, acc)
    else:           #train
        return to_proto(loss)

    
def write_net():
    #write train prototxt
    with open(train_proto, 'w') as f:
        f.write(str(create_net(train_lmdb, batch_size=64)))
    
    #write test prototxt
    with open(test_proto, 'w') as f:
        f.write(str(create_net(test_lmdb, batch_size=32, include_acc=True)))


if __name__ == '__main__':
    write_net()
    

2.绘制网络模型

touch draw_net.sh

vim draw_net.sh

#!/usr/bin/bash

DATA=train
BULID=/home/yuandanfei/caffe/python/draw_net.py

python $BULID ../out/$DATA.prototxt ../out/$DATA.png --rankdir=BT

参考资料：

https://www.cnblogs.com/denny402/p/5679037.html

https://www.cnblogs.com/denny402/p/5106764.html