论文名称:Densely Connected Convolutional Networks(CVPR 2017, Best Paper Award)
论文链接:https://arxiv.org/pdf/1608.06993.pdf
源码链接:https://github.com/liuzhuang13/DenseNet
caffe版源码: https://github.com/liuzhuang13/DenseNetCaffe
本篇博文实现了在Windows环境下DenseNet网络的复现(caffe版)(应用到自己的数据集)
一、下载源码
源码中共有六个文件,分别进行解释。
1.make_densenet.py
make_densenet.py主要用来生成train_densenet.prototxt文件、test_densenet.prototxt文件以及solver.prototxt文件。打开该文件然后run就可以生成以上三个文件。同时可以修改make_densenet.py其中的参数从而修改这三个文件内的参数,比如:路径等。
2.train_densenet.prototxt文件
train_densenet.prototxt文件是训练时用到的网络文件。
3.test_densenet.prototxt文件
test_densenet.prototxt文件是在训练过程中的验证部分用到的网络文件。
4.solver.prototxt文件
solver.prototxt文件是caffe常用的文件。
5.train. sh
train.sh是ubuntu系统下训练文件,在windows系统下是train.bat文件。
注意:
需要注意的是solver.prototxt文件中有两个重要的参数train_net: "train_densenet.prototxt"
和test_net: "test_densenet.prototxt"
,这里个参数分别是训练和验证时用到的网络。原来的网络这部分的参数是net: "train_val.prototxt"
。这是与原来的网络的很大的区别。
(不过train_densenet.prototxt文件、test_densenet.prototxt文件可以修改成一个文件,也就是原来我们常用的train_val.prototxt文件)
train_net: "train_densenet.prototxt"
test_net: "test_densenet.prototxt"
test_iter: 200
test_interval: 800
base_lr: 0.1
display: 1
max_iter: 230000
lr_policy: "multistep"
gamma: 0.1
momentum: 0.9
weight_decay: 0.0001
solver_mode: GPU
random_seed: 831486
stepvalue: 115000
stepvalue: 172500
type: "Nesterov"
二、修改网络训练文件
主要是修改train_densenet.prototxt文件、test_densenet.prototxt文件以及solver.prototxt文件
1.solver.prototxt文件
根据自己的数据集以及经验对改文件进行修改。
2.train_densenet.prototxt文件
主要修改mean_file路径、source路径和最后全连接层的输出参数num_output(根据自己的任务,如二分类,该值应该设为2)。
3.test_densenet.prototxt文件
同修改train_densenet.prototxt文件类似
主要修改mean_file路径、source路径和最后全连接层的输出参数num_output(根据自己的任务,如二分类,该值应该设为2)。
注意:
1.num_output参数一定要根据自己的任务进行修改。
2.根据自己电脑配置修改batch_size大小。
3.其实train_densenet.prototxt文件、test_densenet.prototxt文件可以修改成一个文件,也就是原来我们常用的train_val.prototxt文件(这里就不改了)。
三、开始训练
编辑train.bat,然后双击训练。
SET GLOG_logtostderr=1
caffe.exe路径\caffe.exe train --solver solver.prototxt路径\solver.prototxt
pause
训练时可能会出现内存不足的情况,这时候需要调整batch_size大小再进行训练。
四、测试
1.生成deploy.prototxt 文件。
我是根据test_densenet.prototxt文件自己生成了对应densenet的deploy.prototxt 文件,具体方法请查看我的另一篇博客:https://blog.csdn.net/GL3_24/article/details/90109237 。此处只贴出来对应的deploy.prototxt 文件,很长。仅供参考。
name: "DENSENET_121"
input: "Data1"
input_dim: 1
input_dim: 3
input_dim: 224
input_dim: 224
layer {
name: "Convolution1"
type: "Convolution"
bottom: "Data1"
top: "Convolution1"
convolution_param {
num_output: 16
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "BatchNorm1"
type: "BatchNorm"
bottom: "Convolution1"
top: "BatchNorm1"
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
}
layer {
name: "Scale1"
type: "Scale"
bottom: "BatchNorm1"
top: "BatchNorm1"
scale_param {
filler {
value: 1
}
bias_term: true
bias_filler {
value: 0
}
}
}
layer {
name: "ReLU1"
type: "ReLU"
bottom: "BatchNorm1"
top: "BatchNorm1"
}
layer {
name: "Convolution2"
type: "Convolution"
bottom: "BatchNorm1"
top: "Convolution2"
convolution_param {
num_output: 12
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "Dropout1"
type: "Dropout"
bottom: "Convolution2"
top: "Dropout1"
dropout_param {
dropout_ratio: 0.2
}
}
layer {
name: "Concat1"
type: "Concat"
bottom: "Convolution1"
bottom: "Dropout1"
top: "Concat1"
concat_param {
axis: 1
}
}
layer {
name: "BatchNorm2"
type: "BatchNorm"
bottom: "Concat1"
top: "BatchNorm2"
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
}
layer {
name: "Scale2"
type: "Scale"
bottom: "BatchNorm2"
top: "BatchNorm2"
scale_param {
filler {
value: 1
}
bias_term: true
bias_filler {
value: 0
}
}
}
layer {
name: "ReLU2"
type: "ReLU"
bottom: "BatchNorm2"
top: "BatchNorm2"
}
layer {
name: "Convolution3"
type: "Convolution"
bottom: "BatchNorm2"
top: "Convolution3"
convolution_param {
num_output: 12
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "Dropout2"
type: "Dropout"
bottom: "Convolution3"
top: "Dropout2"
dropout_param {
dropout_ratio: 0.2
}
}
layer {
name: "Concat2"
type: "Concat"
bottom: "Concat1"
bottom: "Dropout2"
top: "Concat2"
concat_param {
axis: 1
}
}
layer {
name: "BatchNorm3"
type: "BatchNorm"
bottom: "Concat2"
top: "BatchNorm3"
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
}
layer {
name: "Scale3"
type: "Scale"
bottom: "BatchNorm3"
top: "BatchNorm3"
scale_param {
filler {
value: 1
}
bias_term: true
bias_filler {
value: 0
}
}
}
layer {
name: "ReLU3"
type: "ReLU"
bottom: "BatchNorm3"
top: "BatchNorm3"
}
layer {
name: "Convolution4"
type: "Convolution"
bottom: "BatchNorm3"
top: "Convolution4"
convolution_param {
num_output: 12
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "Dropout3"
type: "Dropout"
bottom: "Convolution4"
top: "Dropout3"
dropout_param {
dropout_ratio: 0.2
}
}
layer {
name: "Concat3"
type: "Concat"
bottom: "Concat2"
bottom: "Dropout3"
top: "Concat3"
concat_param {
axis: 1
}
}
layer {
name: "BatchNorm4"
type: "BatchNorm"
bottom: "Concat3"
top: "BatchNorm4"
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
}
layer {
name: "Scale4"
type: "Scale"
bottom: "BatchNorm4"
top: "BatchNorm4"
scale_param {
filler {
value: 1
}
bias_term: true
bias_filler {
value: 0
}
}
}
layer {
name: "ReLU4"
type: "ReLU"
bottom: "BatchNorm4"
top: "BatchNorm4"
}
layer {
name: "Convolution5"
type: "Convolution"
bottom: "BatchNorm4"
top: "Convolution5"
convolution_param {
num_output: 12
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "Dropout4"
type: "Dropout"
bottom: "Convolution5"
top: "Dropout4"
dropout_param {
dropout_ratio: 0.2
}
}
layer {
name: "Concat4"
type: "Concat"
bottom: "Concat3"
bottom: "Dropout4"
top: "Concat4"
concat_param {
axis: 1
}
}
layer {
name: "BatchNorm5"
type: "BatchNorm"
bottom: "Concat4"
top: "BatchNorm5"
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
param {
lr_mult: 0
decay_mult: 0
}
}
layer {
name: "Scale5"
type: "Scale"
bottom: "BatchNorm5"
top: "BatchNorm5"
scale_param {
filler {
value: 1
}
bias_term: true
bias_filler {
value: 0
}
}
}
layer {
name: "ReLU5"
type: "ReLU"
bottom: "BatchNorm5"
top: "BatchNorm5"
}
layer {
name: "Convolution6"
type: "Convolution"
bottom: "BatchNorm5"
top: "Convolution6"
convolution_param {
num_output: 12
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "msra"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "Dropout5"
type: "Dropout"
bottom: "Convolution6"
top: "Dropout5"
dropout_param {
dropout_ratio: 0.2
}
}
layer {
name: "Concat5"
type: "Concat"
bottom: "Concat4"
bottom: "Dropout5"
top: "Concat5"
concat_param {
axis: 1
}
}
layer {
name: "BatchNorm6"
type: "BatchNorm"
bottom: "Concat5"
top: "BatchNorm6"
param {