参考赵永科的博客,这里我们实现一个新 Layer,名称为 AllPassLayer,顾名思义就是全通 Layer,“全通”借鉴于信号处理中的全通滤波器,将信号无失真地从输入转到输出。
虽然这个 Layer 并没有什么卵用,但是在这个基础上增加你的处理是非常简单的事情。另外也是出于实验考虑,全通层的 Forward/Backward 函数非常简单不需要读者有任何高等数学和求导的背景知识。读者使用该层时可以插入到任何已有网络中,而不会影响训练、预测的准确性。
首先,要把你的实现,要像正常的 Layer 类一样,分解为声明部分和实现部分,分别放在 .hpp 与 .cpp、.cu 中。Layer 名称要起一个能区别于原版实现的新名称。.hpp 文件置于 $CAFFE_ROOT/include/caffe/layers/,而 .cpp 和 .cu 置于 $CAFFE_ROOT/src/caffe/layers/,这样你在 $CAFFE_ROOT 下执行 make 编译时,会自动将这些文件加入构建过程,省去了手动设置编译选项的繁琐流程。
其次,在 $CAFFE_ROOT/src/caffe/proto/caffe.proto 中,增加新 LayerParameter 选项,这样你在编写 train.prototxt 或者 test.prototxt 或者 deploy.prototxt 时就能把新 Layer 的描述写进去,便于修改网络结构和替换其他相同功能的 Layer 了。
最后也是最容易忽视的一点,在 Layer 工厂注册新 Layer 加工函数,不然在你运行过程中可能会报如下错误:
首先看头文件(all_pass_layer.hpp):
#ifndef CAFFE_ALL_PASS_LAYER_HPP_
#define CAFFE_ALL_PASS_LAYER_HPP_ #include <vector> #include "caffe/blob.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h" #include "caffe/layers/neuron_layer.hpp" namespace caffe {
template <typename Dtype>
class AllPassLayer : public NeuronLayer<Dtype> {
public:
explicit AllPassLayer(const LayerParameter& param)
: NeuronLayer<Dtype>(param) {} virtual inline const char* type() const { return "AllPass"; } protected: virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
}; } // namespace caffe
再看源文件(all_pass_layer.cpp和all_pass_layer.cu,这两个文件我暂时还不知道有什么区别,我是直接复制粘贴的,两个文件内容一样):
#include <algorithm>
#include <vector> #include "caffe/layers/all_pass_layer.hpp"
#include <iostream> using namespace std;
#define DEBUG_AP(str) cout << str << endl;
namespace caffe{ template <typename Dtype>
void AllPassLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom , const vector<Blob<Dtype>*>& top)
{
const Dtype* bottom_data = bottom[] -> cpu_data(); //cpu_data()只读访问cpu data
Dtype* top_data = top[] -> mutable_cpu_data(); //mutable_cpu_data读写访问cpu data
const int count = bottom[] -> count(); //计算Blob中的元素总数
for (int i = ; i < count ; i ++)
{
top_data[i] = bottom_data[i]; //只是单纯的通过,全通
}
DEBUG_AP("Here is All Pass Layer , forwarding.");
DEBUG_AP(this -> layer_param_.all_pass_param().key()); //读取prototxt预设值
} template <typename Dtype>
void AllPassLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top , const vector<bool>& propagate_down , const vector<Blob<Dtype>*>& bottom)
{
if (propagate_down[]) //propagate_down[0] = 1则指定计算权值的剃度 , ...[1] = 1则指定计算偏置项的剃度,这里是计算权值的剃度
{
const Dtype* bottom_data = bottom[] -> cpu_data();
const Dtype* top_diff = top[] -> cpu_diff();
Dtype* bottom_diff = bottom[] -> mutable_cpu_diff();
const int count = bottom[] -> count();
for ( int i = ; i < count ; i ++ )
{
bottom_diff[i] = top_diff[i];
}
}
DEBUG_AP("Here is All Pass Layer , backwarding.");
DEBUG_AP(this -> layer_param_.all_pass_param().key());
} #ifndef CPU_ONLY
#define CPU_ONLY
#endif INSTANTIATE_CLASS(AllPassLayer);
REGISTER_LAYER_CLASS(AllPass);
} //namespace caffe
时间考虑,我没有实现 GPU 模式的 forward、backward,故本文例程仅支持 CPU_ONLY 模式。
编辑 caffe.proto,找到 LayerParameter 描述,增加一项:
message LayerParameter {
optional string name = ; // the layer name
optional string type = ; // the layer type
repeated string bottom = ; // the name of each bottom blob
repeated string top = ; // the name of each top blob // The train / test phase for computation.
optional Phase phase = ; // The amount of weight to assign each top blob in the objective.
// Each layer assigns a default value, usually of either 0 or 1,
// to each top blob.
repeated float loss_weight = ; // Specifies training parameters (multipliers on global learning constants,
// and the name and other settings used for weight sharing).
repeated ParamSpec param = ; // The blobs containing the numeric parameters of the layer.
repeated BlobProto blobs = ; // Specifies on which bottoms the backpropagation should be skipped.
// The size must be either 0 or equal to the number of bottoms.
repeated bool propagate_down = ; // Rules controlling whether and when a layer is included in the network,
// based on the current NetState. You may specify a non-zero number of rules
// to include OR exclude, but not both. If no include or exclude rules are
// specified, the layer is always included. If the current NetState meets
// ANY (i.e., one or more) of the specified rules, the layer is
// included/excluded.
repeated NetStateRule include = ;
repeated NetStateRule exclude = ; // Parameters for data pre-processing.
optional TransformationParameter transform_param = ; // Parameters shared by loss layers.
optional LossParameter loss_param = ; // Layer type-specific parameters.
//
// Note: certain layers may have more than one computational engine
// for their implementation. These layers include an Engine type and
// engine parameter for selecting the implementation.
// The default for the engine is set by the ENGINE switch at compile-time.
optional AccuracyParameter accuracy_param = ;
optional ArgMaxParameter argmax_param = ;
optional BatchNormParameter batch_norm_param = ;
optional BiasParameter bias_param = ;
optional ConcatParameter concat_param = ;
optional ContrastiveLossParameter contrastive_loss_param = ;
optional ConvolutionParameter convolution_param = ;
optional CropParameter crop_param = ;
optional DataParameter data_param = ;
optional DropoutParameter dropout_param = ;
optional DummyDataParameter dummy_data_param = ;
optional EltwiseParameter eltwise_param = ;
optional ELUParameter elu_param = ;
optional EmbedParameter embed_param = ;
optional ExpParameter exp_param = ;
optional FlattenParameter flatten_param = ;
optional HDF5DataParameter hdf5_data_param = ;
optional HDF5OutputParameter hdf5_output_param = ;
optional HingeLossParameter hinge_loss_param = ;
optional ImageDataParameter image_data_param = ;
optional InfogainLossParameter infogain_loss_param = ;
optional InnerProductParameter inner_product_param = ;
optional InputParameter input_param = ;
optional LogParameter log_param = ;
optional LRNParameter lrn_param = ;
optional MemoryDataParameter memory_data_param = ;
optional MVNParameter mvn_param = ;
optional PoolingParameter pooling_param = ;
optional PowerParameter power_param = ;
optional PReLUParameter prelu_param = ;
optional PythonParameter python_param = ;
optional ReductionParameter reduction_param = ;
optional ReLUParameter relu_param = ;
optional ReshapeParameter reshape_param = ;
optional ScaleParameter scale_param = ;
optional SigmoidParameter sigmoid_param = ;
optional SoftmaxParameter softmax_param = ;
optional SPPParameter spp_param = ;
optional SliceParameter slice_param = ;
optional TanHParameter tanh_param = ;
optional ThresholdParameter threshold_param = ;
optional TileParameter tile_param = ;
optional WindowDataParameter window_data_param = ;
optional AllPassParameter all_pass_param = ;
}
注意新增数字不要和以前的 Layer 数字重复。
仍然在 caffe.proto 中,增加 AllPassParameter 声明,位置任意。我设定了一个参数,可以用于从 prototxt 中读取预设值。
message AllPassParameter {
optional float key = [default = ];
}
这句来读取 prototxt 预设值。
在 $CAFFE_ROOT 下执行 make clean,然后重新 make all。要想一次编译成功,务必规范代码,对常见错误保持敏锐的嗅觉并加以避免。
万事具备,只欠 prototxt 了。
不难,我们写个最简单的 deploy.prototxt,不需要 data layer 和 softmax layer,just for fun。
name: "AllPassTest"
layer {
name: "data"
type: "Input"
top: "data"
input_param { shape: { dim: dim: dim: dim: } }
}
layer {
name: "ap"
type: "AllPass"
bottom: "data"
top: "conv1"
all_pass_param {
key: 12.88
}
}
注意,这里的 type :后面写的内容,应该是你在 .hpp 中声明的新类 class name 去掉 Layer 后的名称。
上面设定了 key 这个参数的预设值为 12.88,嗯,你想到了刘翔对不对。
为了检验该 Layer 是否能正常创建和执行 forward, backward,我们运行 caffe time 命令并指定刚刚实现的 prototxt :
$ ./build/tools/caffe.bin time -model deploy.prototxt
I1002 ::41.667682 caffe.cpp:] Use CPU.
I1002 ::41.671360 net.cpp:] Initializing net from parameters:
name: "AllPassTest"
state {
phase: TRAIN
}
layer {
name: "data"
type: "Input"
top: "data"
input_param {
shape {
dim:
dim:
dim:
dim:
}
}
}
layer {
name: "ap"
type: "AllPass"
bottom: "data"
top: "conv1"
all_pass_param {
key: 12.88
}
}
I1002 ::41.671463 layer_factory.hpp:] Creating layer data
I1002 ::41.671484 net.cpp:] Creating Layer data
I1002 ::41.671499 net.cpp:] data -> data
I1002 ::41.671555 net.cpp:] Setting up data
I1002 ::41.671566 net.cpp:] Top shape: ()
I1002 ::41.671592 net.cpp:] Memory required for data:
I1002 ::41.671605 layer_factory.hpp:] Creating layer ap
I1002 ::41.671620 net.cpp:] Creating Layer ap
I1002 ::41.671630 net.cpp:] ap <- data
I1002 ::41.671644 net.cpp:] ap -> conv1
I1002 ::41.671663 net.cpp:] Setting up ap
I1002 ::41.671674 net.cpp:] Top shape: ()
I1002 ::41.671685 net.cpp:] Memory required for data:
I1002 ::41.671695 net.cpp:] ap does not need backward computation.
I1002 ::41.671705 net.cpp:] data does not need backward computation.
I1002 ::41.671710 net.cpp:] This network produces output conv1
I1002 ::41.671720 net.cpp:] Network initialization done.
I1002 ::41.671746 caffe.cpp:] Performing Forward
Here is All Pass Layer, forwarding.
12.88
I1002 ::41.679689 caffe.cpp:] Initial loss:
I1002 ::41.679714 caffe.cpp:] Performing Backward
I1002 ::41.679738 caffe.cpp:] *** Benchmark begins ***
I1002 ::41.679746 caffe.cpp:] Testing for iterations.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.681139 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.682394 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.683653 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.685096 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.686326 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.687713 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.689038 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.690251 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.691548 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.692805 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.694056 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.695264 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.696761 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.698225 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.699653 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.700945 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.702761 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.704056 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.706471 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.708784 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.710043 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.711272 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.712528 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.713964 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.715248 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.716487 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.717725 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.718962 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.720289 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.721837 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.723042 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.724261 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.725587 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.726771 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.728013 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.729249 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.730716 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.732275 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.733809 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.735049 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.737144 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.739090 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.741575 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.743450 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.744732 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.745970 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.747185 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.748430 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.749826 caffe.cpp:] Iteration: forward-backward time: ms.
Here is All Pass Layer, forwarding.
12.88
Here is All Pass Layer, backwarding.
12.88
I1002 ::41.751124 caffe.cpp:] Iteration: forward-backward time: ms.
I1002 ::41.751147 caffe.cpp:] Average time per layer:
I1002 ::41.751157 caffe.cpp:] data forward: 0.00108 ms.
I1002 ::41.751183 caffe.cpp:] data backward: 0.001 ms.
I1002 ::41.751194 caffe.cpp:] ap forward: 1.37884 ms.
I1002 ::41.751205 caffe.cpp:] ap backward: 0.01156 ms.
I1002 ::41.751220 caffe.cpp:] Average Forward pass: 1.38646 ms.
I1002 ::41.751231 caffe.cpp:] Average Backward pass: 0.0144 ms.
I1002 ::41.751240 caffe.cpp:] Average Forward-Backward: 1.42 ms.
I1002 ::41.751250 caffe.cpp:] Total Time: ms.
I1002 ::41.751260 caffe.cpp:] *** Benchmark ends ***
可见该 Layer 可以正常创建、加载预设参数、执行 forward、backward 函数。
实际上对于算法 Layer,还要写 Test Case 保证功能正确。由于我们选择了极为简单的全通 Layer,故这一步可以省去。我这里偷点懒,您省点阅读时间。
补充:make all后还要重新配置python接口,make pycaffe ,然后再按之前的操作就行了。。。
参考链接:http://blog.csdn.net/kkk584520/article/details/52721838