GitHub 代码地址： https://github.com/HansRen1024/Use-Python-to-call-Caffe-module
前言#
写这一系列博文前真的想了好久，有种无从下手的感觉。还是功力太浅，越是这样越要硬着头皮写。加油！
我先将将各个函数单独放出来，最后在放出完整代码。各个函数中单独用到的模块，我在函数中单独加载，这样方便将代码独立出来单独运行。
import caffe 是全局通用模块，我就不在每个函数中单独加载了。
一、加载 caffe 模块#
为了以后方便使用，把 caffe 模块放到 python 默认路径下，这样在任意目录下就都能加载 caffe 模块了。
caffe 编译通过后运行：
make pycaffe
sudo cp -r python/caffe/ /usr/local/lib/python/dist-packages #有些朋友路径是site-packages,这个因人而异。
这个时候运行 python，import caffe，会提示找不到 caffe 的动态库。
可以将 $CAFFE_ROOT/.build_release/lib/ 加到环境变量中去，
也可以将该动态库复制到 /usr/lib/ 或者 /usr/local/lib 目录下。
此时，应该就可以在任意目录下运行 python，import caffe 了。
二、生成训练和测试 prototxt 文件#
生成网络结构文件我找到了两种方法，下面这种通过定义一个 NetSpec () 实例 n 的方法是比较好的，所以我先写的这个方法。这种方法生成的文件内，各个层的名字和输出 blob 的名字就是等号前面定义的 n 的方法。
def lenet(lmdb, batch_size, include_acc=False):
from caffe import layers as L
from caffe import params as P
n = caffe.NetSpec()
# 具体每个层内的参数，可以对照现有的prototxt文件。我这里写的并不全。
n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
transform_param=dict(scale=1./255), ntop=2) # ntop表示两个输出
n.conv1 = L.Convolution(n.data, kernel_size=5, num_output=20, weight_filler=dict(type='xavier'))
n.pool1 = L.Pooling(n.conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)
n.conv2 = L.Convolution(n.pool1, kernel_size=5, num_output=50, weight_filler=dict(type='xavier'))
n.pool2 = L.Pooling(n.conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)
n.ip1 = L.InnerProduct(n.pool2, num_output=500, weight_filler=dict(type='xavier'))
n.relu1 = L.ReLU(n.ip1, in_place=True)
n.ip2 = L.InnerProduct(n.relu1, num_output=10, weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
if include_acc: # 生成测试文件时，需要有计算准确率的层。
n.acc = L.Accuracy(n.ip2, n.label)
return n.to_proto() # 注意这里的to_proto()不用带参数。
def write_lenet():
with open('./doc/train_lenet.prototxt','w') as f:
f.write(str(lenet('./doc/mnist_train_lmdb', 64)))
with open('./doc/test_lenet.prototxt', 'w') as f:
f.write(str(lenet('./doc/mnist_test_lmdb', 100, True)))
三、生成 deploy 文件#
这里我放出生成网络结构文件的第二种方法，虽然我并不推荐使用这个方法。其实 deploy 直接拿上面生成好的文件改就行，很简单。
def deploy():
from caffe import layers as L
from caffe import params as P
from caffe import to_proto
# deploy文件没有数据层
conv1 = L.Convolution(bottom='data', kernel_size=5, num_output=20, weight_filler=dict(type='xavier'))
pool1 = L.Pooling(conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)
conv2 = L.Convolution(pool1, kernel_size=5, num_output=50, weight_filler=dict(type='xavier'))
pool2 = L.Pooling(conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)
ip1 = L.InnerProduct(pool2, num_output=500, weight_filler=dict(type='xavier'))
relu1 = L.ReLU(ip1, in_place=True)
ip2 = L.InnerProduct(relu1, num_output=10, weight_filler=dict(type='xavier'))
prob = L.Softmax(ip2) # 最后一层不用计算loss，输出概率。
return to_proto(prob) # 这里需要带参数
def write_deploy():
with open('doc/deploy_lenet.prototxt', 'w') as f:
f.write('name: "Lenet"\n')
f.write('input: "data"\n')
f.write('input_dim: 1\n')
f.write('input_dim: 3\n')
f.write('input_dim: 28\n')
f.write('input_dim: 28\n')
f.write(str(deploy()))
四、生成 solver 文件#
方法一是用字典生成：
def solver_dict():
solver_file='doc/solver_lenet.prototxt'
sp={}
sp['train_net']='"doc/train_lenet.prototxt"'
sp['test_net']='"doc/test_lenet.prototxt"'
sp['test_iter']='100'
sp['test_interval']='500'
sp['display']='100'
sp['max_iter']='10000'
sp['base_lr']='0.01'
sp['lr_policy']='"inv"'
sp['gamma']='0.0001'
sp['power']='0.75'
sp['momentum']='0.9'
sp['weight_decay']='0.0005'
sp['snapshot']='5000'
sp['snapshot_prefix']='"models/lenet"'
sp['solver_mode']='GPU'
sp['solver_type']='SGD'
sp['device_id']='0'
with open(solver_file, 'w') as f:
for key, value in sp.items():
if not(type(value) is str):
raise TypeError('All solver parameters must be string')
f.write('%s: %s\n' %(key, value))
方法二是调用 caffe 模块生成，这种方法生成的文件内，小数部分会有很小的损失。处女座强迫症犯了！
def solver_caffe():
from caffe.proto import caffe_pb2
s = caffe_pb2.SolverParameter()
solver_file='doc/solver_lenet.prototxt'
s.train_net = 'doc/train_lenet.prototxt'
s.test_net.append('doc/test_lenet.prototxt')
s.test_interval = 500
s.test_iter.append(100)
s.display = 100
s.max_iter = 10000
s.base_lr = 0.01
s.lr_policy = "inv"
s.gamma = 0.0001
s.power = 0.75
s.momentum = 0.9
s.weight_decay = 0.0005
s.snapshot = 5000
s.snapshot_prefix = "models/lenet"
s.type = "SGD"
s.solver_mode = caffe_pb2.SolverParameter.GPU
with open(solver_file, 'w') as f:
f.write(str(s))
五、完整代码：#
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Sat Jul 29 11:16:29 2017
@author: hans
"""
import caffe
def lenet(lmdb, batch_size, include_acc=False):
from caffe import layers as L
from caffe import params as P
n = caffe.NetSpec()
n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
transform_param=dict(scale=1./255), ntop=2)
n.conv1 = L.Convolution(n.data, kernel_size=5, num_output=20, weight_filler=dict(type='xavier'))
n.pool1 = L.Pooling(n.conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)
n.conv2 = L.Convolution(n.pool1, kernel_size=5, num_output=50, weight_filler=dict(type='xavier'))
n.pool2 = L.Pooling(n.conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)
n.ip1 = L.InnerProduct(n.pool2, num_output=500, weight_filler=dict(type='xavier'))
n.relu1 = L.ReLU(n.ip1, in_place=True)
n.ip2 = L.InnerProduct(n.relu1, num_output=10, weight_filler=dict(type='xavier'))
n.loss = L.SoftmaxWithLoss(n.ip2, n.label)
if include_acc:
n.acc = L.Accuracy(n.ip2, n.label)
return n.to_proto()
def write_lenet():
with open('./doc/train_lenet.prototxt','w') as f:
f.write(str(lenet('./doc/mnist_train_lmdb', 64)))
with open('./doc/test_lenet.prototxt', 'w') as f:
f.write(str(lenet('./doc/mnist_test_lmdb', 100, True)))
def deploy():
from caffe import layers as L
from caffe import params as P
from caffe import to_proto
conv1 = L.Convolution(bottom='data', kernel_size=5, num_output=20, weight_filler=dict(type='xavier'))
pool1 = L.Pooling(conv1, kernel_size=2, stride=2, pool=P.Pooling.MAX)
conv2 = L.Convolution(pool1, kernel_size=5, num_output=50, weight_filler=dict(type='xavier'))
pool2 = L.Pooling(conv2, kernel_size=2, stride=2, pool=P.Pooling.MAX)
ip1 = L.InnerProduct(pool2, num_output=500, weight_filler=dict(type='xavier'))
relu1 = L.ReLU(ip1, in_place=True)
ip2 = L.InnerProduct(relu1, num_output=10, weight_filler=dict(type='xavier'))
prob = L.Softmax(ip2)
return to_proto(prob)
def write_deploy():
with open('doc/deploy_lenet.prototxt', 'w') as f:
f.write('name: "Lenet"\n')
f.write('input: "data"\n')
f.write('input_dim: 1\n')
f.write('input_dim: 3\n')
f.write('input_dim: 28\n')
f.write('input_dim: 28\n')
f.write(str(deploy()))
def solver_dict():
solver_file='doc/solver_lenet.prototxt'
sp={}
sp['train_net']='"doc/train_lenet.prototxt"'
sp['test_net']='"doc/test_lenet.prototxt"'
sp['test_iter']='100'
sp['test_interval']='500'
sp['display']='100'
sp['max_iter']='10000'
sp['base_lr']='0.01'
sp['lr_policy']='"inv"'
sp['gamma']='0.0001'
sp['power']='0.75'
sp['momentum']='0.9'
sp['weight_decay']='0.0005'
sp['snapshot']='5000'
sp['snapshot_prefix']='"models/lenet"'
sp['solver_mode']='GPU'
sp['solver_type']='SGD'
sp['device_id']='0'
with open(solver_file, 'w') as f:
for key, value in sp.items():
if not(type(value) is str):
raise TypeError('All solver parameters must be string')
f.write('%s: %s\n' %(key, value))
def solver_caffe():
from caffe.proto import caffe_pb2
s = caffe_pb2.SolverParameter()
solver_file='doc/solver_lenet.prototxt'
s.train_net = 'doc/train_lenet.prototxt'
s.test_net.append('doc/test_lenet.prototxt')
s.test_interval = 500
s.test_iter.append(100)
s.display = 100
s.max_iter = 10000
s.base_lr = 0.01
s.lr_policy = "inv"
s.gamma = 0.0001
s.power = 0.75
s.momentum = 0.9
s.weight_decay = 0.0005
s.snapshot = 5000
s.snapshot_prefix = "models/lenet"
s.type = "SGD"
s.solver_mode = caffe_pb2.SolverParameter.GPU
with open(solver_file, 'w') as f:
f.write(str(s))
def train():
caffe.set_device(0)
caffe.set_mode_gpu()
solver = caffe.SGDSolver('doc/solver_lenet.prototxt')
solver.solve()
if __name__ == '__main__':
write_lenet()
# write_deploy()
# solver_dict()
# solver_caffe()
# train()