caffe搭建squeezenet网络的整套工程

之前用pytorch构建了squeezenet，个人觉得pytorch是最好用的，但是有的工程就是需要caffe结构的，所以本篇也用caffe构建一个squeezenet网络。

数据处理

首先要对数据进行处理，跟pytorch不同，pytorch读取数据只需要给数据集所在目录即可直接从中读取数据，而caffe需要一个包含每张图片的绝对路径以及所在类别的txt文件，从中读取数据。写一个生成次txt文件的脚本：

import os
import randomfolder = 'cotta'  # 数据集目录相对路径
names = os.listdir(folder)f1 = open('/train_txt/train_cotta.txt', 'a')  # 生成的txt地址
f2 = open('/train_txt/test_water_workcloth.txt', 'a')for name in names:imgnames = os.listdir(folder + '/' + name)random.shuffle(imgnames)numimg = len(imgnames)for i in range(numimg):f1.write('%s %s\n' % (folder + '/' + name + '/' + imgnames[i], name[0]))# if i < int(0.9*numimg):#     f1.write('%s %s\n'%(folder + '/' + name + '/' + imgnames[i], name[0]))# else:#     f2.write('%s %s\n'%(folder + '/' + name + '/' + imgnames[i], name[0]))
# f2.close()
f1.close()

数据集的目录也要跟pytorch的一致，一个类的数据放在一个目录中，目录名为类名。且脚本与该目录同级。
运行脚本后生成的txt内容如下：

/cotta/0_other/0_1_391_572_68_68.jpg 0
/cotta/1_longSleeves/9605_1_5_565_357_82_70.jpg 1
/cotta/2_cotta/713_0.99796_1_316_162_96_87.jpg 2
......
图片相对路径 图片所属类别

网络结构配置文件

trainval.prototxt

layer {name: "data"type: "ImageData"top: "data"top: "label"transform_param {mirror: truecrop_size: 96}image_data_param {source: "/train_txt/train_cotta.txt"   # 生成的txt的相对路径root_folder: "/data/"   # 存放数据集目录的路径batch_size: 64shuffle: truenew_height: 96new_width: 96}}
layer {name: "conv1"type: "Convolution"bottom: "data"top: "conv1"convolution_param {num_output: 96kernel_size: 3stride: 1pad: 1weight_filler {type: "xavier"}}
}layer {  name: "BatchNorm1"  type: "BatchNorm" bottom: "conv1"  top: "BatchNorm1"   
}layer {name: "relu_conv1"type: "ReLU"bottom: "BatchNorm1"top: "BatchNorm1"
}
layer {name: "pool1"type: "Pooling"bottom: "BatchNorm1"top: "pool1"pooling_param {pool: MAXkernel_size: 2stride: 2}
}
layer {name: "fire2/squeeze1x1"type: "Convolution"bottom: "pool1"top: "fire2/squeeze1x1"convolution_param {num_output: 16kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire2/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire2/squeeze1x1"  top: "fire2/bn_squeeze1x1"   
}layer {name: "fire2/relu_squeeze1x1"type: "ReLU"bottom: "fire2/bn_squeeze1x1"top: "fire2/bn_squeeze1x1"
}
layer {name: "fire2/expand1x1"type: "Convolution"bottom: "fire2/bn_squeeze1x1"top: "fire2/expand1x1"convolution_param {num_output: 64kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire2/bn_expand1x1"  type: "BatchNorm" bottom: "fire2/expand1x1"  top: "fire2/bn_expand1x1"   
}layer {name: "fire2/relu_expand1x1"type: "ReLU"bottom: "fire2/bn_expand1x1"top: "fire2/bn_expand1x1"
}
layer {name: "fire2/expand3x3"type: "Convolution"bottom: "fire2/bn_expand1x1"top: "fire2/expand3x3"convolution_param {num_output: 64pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire2/bn_expand3x3"  type: "BatchNorm" bottom: "fire2/expand3x3"  top: "fire2/bn_expand3x3"   
}layer {name: "fire2/relu_expand3x3"type: "ReLU"bottom: "fire2/bn_expand3x3"top: "fire2/bn_expand3x3"
}
layer {name: "fire2/concat"type: "Concat"bottom: "fire2/bn_expand1x1"bottom: "fire2/bn_expand3x3"top: "fire2/concat"
}#fire2 ends: 128 channels
layer {name: "fire3/squeeze1x1"type: "Convolution"bottom: "fire2/concat"top: "fire3/squeeze1x1"convolution_param {num_output: 16kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire3/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire3/squeeze1x1"  top: "fire3/bn_squeeze1x1"   
}layer {name: "fire3/relu_squeeze1x1"type: "ReLU"bottom: "fire3/bn_squeeze1x1"top: "fire3/bn_squeeze1x1"
}
layer {name: "fire3/expand1x1"type: "Convolution"bottom: "fire3/bn_squeeze1x1"top: "fire3/expand1x1"convolution_param {num_output: 64kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire3/bn_expand1x1"  type: "BatchNorm" bottom: "fire3/expand1x1"  top: "fire3/bn_expand1x1"   
}layer {name: "fire3/relu_expand1x1"type: "ReLU"bottom: "fire3/bn_expand1x1"top: "fire3/bn_expand1x1"
}
layer {name: "fire3/expand3x3"type: "Convolution"bottom: "fire3/bn_expand1x1"top: "fire3/expand3x3"convolution_param {num_output: 64pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire3/bn_expand3x3"  type: "BatchNorm" bottom: "fire3/expand3x3"  top: "fire3/bn_expand3x3"   
}layer {name: "fire3/relu_expand3x3"type: "ReLU"bottom: "fire3/bn_expand3x3"top: "fire3/bn_expand3x3"
}
layer {name: "fire3/concat"type: "Concat"bottom: "fire3/bn_expand1x1"bottom: "fire3/bn_expand3x3"top: "fire3/concat"
}#fire3 ends: 128 channelslayer {name: "bypass_23"type: "Eltwise"bottom: "fire2/concat"bottom: "fire3/concat"top: "fire3_EltAdd"
}layer {name: "fire4/squeeze1x1"type: "Convolution"bottom: "fire3_EltAdd"top: "fire4/squeeze1x1"convolution_param {num_output: 32kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire4/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire4/squeeze1x1"  top: "fire4/bn_squeeze1x1"   
}layer {name: "fire4/relu_squeeze1x1"type: "ReLU"bottom: "fire4/bn_squeeze1x1"top: "fire4/bn_squeeze1x1"
}
layer {name: "fire4/expand1x1"type: "Convolution"bottom: "fire4/bn_squeeze1x1"top: "fire4/expand1x1"convolution_param {num_output: 128kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire4/bn_expand1x1"  type: "BatchNorm" bottom: "fire4/expand1x1"  top: "fire4/bn_expand1x1"   
}layer {name: "fire4/relu_expand1x1"type: "ReLU"bottom: "fire4/bn_expand1x1"top: "fire4/bn_expand1x1"
}
layer {name: "fire4/expand3x3"type: "Convolution"bottom: "fire4/bn_expand1x1"top: "fire4/expand3x3"convolution_param {num_output: 128pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire4/bn_expand3x3"  type: "BatchNorm" bottom: "fire4/expand3x3"  top: "fire4/bn_expand3x3"   
}layer {name: "fire4/relu_expand3x3"type: "ReLU"bottom: "fire4/bn_expand3x3"top: "fire4/bn_expand3x3"
}
layer {name: "fire4/concat"type: "Concat"bottom: "fire4/bn_expand1x1"bottom: "fire4/bn_expand3x3"top: "fire4/concat"
}
#fire4 ends: 256 channelslayer {name: "pool4"type: "Pooling"bottom: "fire4/concat"top: "pool4"pooling_param {pool: MAXkernel_size: 2stride: 2}
}
#fire4 ends: 256 channels / pooled
layer {name: "fire5/squeeze1x1"type: "Convolution"bottom: "pool4"top: "fire5/squeeze1x1"convolution_param {num_output: 32kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire5/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire5/squeeze1x1"  top: "fire5/bn_squeeze1x1"   
}layer {name: "fire5/relu_squeeze1x1"type: "ReLU"bottom: "fire5/bn_squeeze1x1"top: "fire5/bn_squeeze1x1"
}
layer {name: "fire5/expand1x1"type: "Convolution"bottom: "fire5/bn_squeeze1x1"top: "fire5/expand1x1"convolution_param {num_output: 128kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire5/bn_expand1x1"  type: "BatchNorm" bottom: "fire5/expand1x1"  top: "fire5/bn_expand1x1"   
}layer {name: "fire5/relu_expand1x1"type: "ReLU"bottom: "fire5/bn_expand1x1"top: "fire5/bn_expand1x1"
}
layer {name: "fire5/expand3x3"type: "Convolution"bottom: "fire5/bn_expand1x1"top: "fire5/expand3x3"convolution_param {num_output: 128pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire5/bn_expand3x3"  type: "BatchNorm" bottom: "fire5/expand3x3"  top: "fire5/bn_expand3x3"   
}layer {name: "fire5/relu_expand3x3"type: "ReLU"bottom: "fire5/bn_expand3x3"top: "fire5/bn_expand3x3"
}
layer {name: "fire5/concat"type: "Concat"bottom: "fire5/bn_expand1x1"bottom: "fire5/bn_expand3x3"top: "fire5/concat"
}#fire5 ends: 256 channels
layer {name: "bypass_45"type: "Eltwise"bottom: "pool4"bottom: "fire5/concat"top: "fire5_EltAdd"
}layer {name: "fire6/squeeze1x1"type: "Convolution"bottom: "fire5_EltAdd"top: "fire6/squeeze1x1"convolution_param {num_output: 48kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire6/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire6/squeeze1x1"  top: "fire6/bn_squeeze1x1"   
}layer {name: "fire6/relu_squeeze1x1"type: "ReLU"bottom: "fire6/bn_squeeze1x1"top: "fire6/bn_squeeze1x1"
}
layer {name: "fire6/expand1x1"type: "Convolution"bottom: "fire6/bn_squeeze1x1"top: "fire6/expand1x1"convolution_param {num_output: 192kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire6/bn_expand1x1"  type: "BatchNorm" bottom: "fire6/expand1x1"  top: "fire6/bn_expand1x1"   
}layer {name: "fire6/relu_expand1x1"type: "ReLU"bottom: "fire6/bn_expand1x1"top: "fire6/bn_expand1x1"
}
layer {name: "fire6/expand3x3"type: "Convolution"bottom: "fire6/bn_expand1x1"top: "fire6/expand3x3"convolution_param {num_output: 192pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire6/bn_expand3x3"  type: "BatchNorm" bottom: "fire6/expand3x3"  top: "fire6/bn_expand3x3"   
}layer {name: "fire6/relu_expand3x3"type: "ReLU"bottom: "fire6/bn_expand3x3"top: "fire6/bn_expand3x3"
}
layer {name: "fire6/concat"type: "Concat"bottom: "fire6/bn_expand1x1"bottom: "fire6/bn_expand3x3"top: "fire6/concat"
}
#fire6 ends: 384 channelslayer {name: "fire7/squeeze1x1"type: "Convolution"bottom: "fire6/concat"top: "fire7/squeeze1x1"convolution_param {num_output: 48kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire7/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire7/squeeze1x1"  top: "fire7/bn_squeeze1x1"   
}layer {name: "fire7/relu_squeeze1x1"type: "ReLU"bottom: "fire7/bn_squeeze1x1"top: "fire7/bn_squeeze1x1"
}
layer {name: "fire7/expand1x1"type: "Convolution"bottom: "fire7/bn_squeeze1x1"top: "fire7/expand1x1"convolution_param {num_output: 192kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire7/bn_expand1x1"  type: "BatchNorm" bottom: "fire7/expand1x1"  top: "fire7/bn_expand1x1"   
}layer {name: "fire7/relu_expand1x1"type: "ReLU"bottom: "fire7/bn_expand1x1"top: "fire7/bn_expand1x1"
}
layer {name: "fire7/expand3x3"type: "Convolution"bottom: "fire7/bn_expand1x1"top: "fire7/expand3x3"convolution_param {num_output: 192pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire7/bn_expand3x3"  type: "BatchNorm" bottom: "fire7/expand3x3"  top: "fire7/bn_expand3x3"   
}layer {name: "fire7/relu_expand3x3"type: "ReLU"bottom: "fire7/bn_expand3x3"top: "fire7/bn_expand3x3"
}
layer {name: "fire7/concat"type: "Concat"bottom: "fire7/bn_expand1x1"bottom: "fire7/bn_expand3x3"top: "fire7/concat"
}
#fire7 ends: 384 channels
layer {name: "bypass_67"type: "Eltwise"bottom: "fire6/concat"bottom: "fire7/concat"top: "fire7_EltAdd"
}layer {name: "fire8/squeeze1x1"type: "Convolution"bottom: "fire7_EltAdd"top: "fire8/squeeze1x1"convolution_param {num_output: 64kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire8/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire8/squeeze1x1"  top: "fire8/bn_squeeze1x1"   
}layer {name: "fire8/relu_squeeze1x1"type: "ReLU"bottom: "fire8/bn_squeeze1x1"top: "fire8/bn_squeeze1x1"
}
layer {name: "fire8/expand1x1"type: "Convolution"bottom: "fire8/bn_squeeze1x1"top: "fire8/expand1x1"convolution_param {num_output: 256kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire8/bn_expand1x1"  type: "BatchNorm" bottom: "fire8/expand1x1"  top: "fire8/bn_expand1x1"   
}layer {name: "fire8/relu_expand1x1"type: "ReLU"bottom: "fire8/bn_expand1x1"top: "fire8/bn_expand1x1"
}
layer {name: "fire8/expand3x3"type: "Convolution"bottom: "fire8/bn_expand1x1"top: "fire8/expand3x3"convolution_param {num_output: 256pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire8/bn_expand3x3"  type: "BatchNorm" bottom: "fire8/expand3x3"  top: "fire8/bn_expand3x3"   
}layer {name: "fire8/relu_expand3x3"type: "ReLU"bottom: "fire8/bn_expand3x3"top: "fire8/bn_expand3x3"
}
layer {name: "fire8/concat"type: "Concat"bottom: "fire8/bn_expand1x1"bottom: "fire8/bn_expand3x3"top: "fire8/concat"
}
#fire8 ends: 512 channelslayer {name: "pool8"type: "Pooling"bottom: "fire8/concat"top: "pool8"pooling_param {pool: MAXkernel_size: 2stride: 2}
}
#fire8 ends: 512 channels
layer {name: "fire9/squeeze1x1"type: "Convolution"bottom: "pool8"top: "fire9/squeeze1x1"convolution_param {num_output: 64kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire9/bn_squeeze1x1"  type: "BatchNorm" bottom: "fire9/squeeze1x1"  top: "fire9/bn_squeeze1x1"   
}layer {name: "fire9/relu_squeeze1x1"type: "ReLU"bottom: "fire9/bn_squeeze1x1"top: "fire9/bn_squeeze1x1"
}
layer {name: "fire9/expand1x1"type: "Convolution"bottom: "fire9/bn_squeeze1x1"top: "fire9/expand1x1"convolution_param {num_output: 256kernel_size: 1weight_filler {type: "xavier"}}
}layer {  name: "fire9/bn_expand1x1"  type: "BatchNorm" bottom: "fire9/expand1x1"  top: "fire9/bn_expand1x1"   
}layer {name: "fire9/relu_expand1x1"type: "ReLU"bottom: "fire9/bn_expand1x1"top: "fire9/bn_expand1x1"
}
layer {name: "fire9/expand3x3"type: "Convolution"bottom: "fire9/bn_expand1x1"top: "fire9/expand3x3"convolution_param {num_output: 256pad: 1kernel_size: 3weight_filler {type: "xavier"}}
}layer {  name: "fire9/bn_expand3x3"  type: "BatchNorm" bottom: "fire9/expand3x3"  top: "fire9/bn_expand3x3"   
}layer {name: "fire9/relu_expand3x3"type: "ReLU"bottom: "fire9/bn_expand3x3"top: "fire9/bn_expand3x3"
}
layer {name: "fire9/concat"type: "Concat"bottom: "fire9/bn_expand1x1"bottom: "fire9/bn_expand3x3"top: "fire9/concat"
}
#fire9 ends: 512 channelslayer {name: "conv10_new"type: "Convolution"bottom: "fire9/concat"top: "conv10"convolution_param {num_output: 3kernel_size: 1weight_filler {type: "gaussian"mean: 0.0std: 0.01}}
}layer {name: "pool10"type: "Pooling"bottom: "conv10"top: "pool10"pooling_param {pool: AVEglobal_pooling: true}
}# loss, top1, top5
layer {name: "loss"type: "SoftmaxWithLoss"bottom: "pool10"bottom: "label"top: "loss"include {
#    phase: TRAIN}
}
layer {name: "accuracy"type: "Accuracy"bottom: "pool10"bottom: "label"top: "accuracy"#include {#  phase: TEST#}
}

在最后一层卷积层conv10中的num_output修改类别数量。

模型超参配置文件

solver.prototxt

test_iter: 2000 #not subject to iter_size
test_interval: 1000000
# base_lr: 0.0001
base_lr: 0.005       # 学习率
display: 40
# max_iter: 600000
max_iter: 200000    # 迭代数
iter_size: 2 #global batch size = batch_size * iter_size
lr_policy: "poly"
power: 1.0 #linearly decrease LR
momentum: 0.9
weight_decay: 0.0002
snapshot: 10000     # 每多少次迭代保存一个模型
snapshot_prefix: "/data/zxc/classfication/model/model_cotta/cotta_"   # 模型保存路径
solver_mode: GPU
random_seed: 42
net: "./trainNets_drive/trainval.prototxt"   # 网络结构配置文件的路径 
test_initialization: false
average_loss: 40

• max_iter：caffe用的是迭代数而不是pytorch的轮数。pytorch中训练完全部的训练集为一轮，而caffe中训练完一个batch_size的数据为一个迭代。如果想要等价与轮数的话，一轮就等于：len(train_data) / batch_size。如果有余数就要看pytorch里的dataloader里面设置舍去还是为一个batch，如果舍去就是向下取整，如果不舍去就是向上取整；
• snapshot_prefix：最后一部分为每个保存模型的前缀，如图：
  

运行命令

将运行命令写入bash文件中：
train.sh

/home/seg/anaconda3/envs/zxc/bin/caffe train -gpu 1 -solver ./solvers/solver_3.prototxt -weights=/data/classfication/model/model_cotta/cotta__iter_200000.caffemodel 2>&1 | tee log_3_4_class.txt 

• -gpu：选择哪块卡，如果就一块就是0；
• -solver：后面跟网络超参配置文件路径；
• -weights：后面跟预训练模型，可以用官方给的squeezenet的caffe版本的预训练模型，我这里是训练中断从断点继续训练

编写完成后source activate 环境名称进入source环境，然后source train.sh运行bash文件就能开始训练。