在飞驰的列车上，无法入眠。外面阴雨绵绵，思绪被拉扯到天边。
翻看之前聊天，想起还欠一个读者一篇博客。
于是花了点时间整理一下之前学习 Caffe 时增加自定义 Layer 及自定义 ProtoBuffer 参数的简单例程，希望对初学者有借鉴意义。
博客内容基于新书《深度学习：21 天实战 Caffe》，书中课后习题答案欢迎读者留言讨论。以下进入正文。
在使用 Caffe 过程中经常会有这样的需求：已有 Layer 不符合我的应用场景；我需要这样这样的功能，原版代码没有实现；或者已经实现但效率太低，我有更好的实现。
方案一：简单粗暴的解法——偷天换日
如果你对 ConvolutionLayer 的实现不满意，那就直接改这两个文件：$CAFFE_ROOT/include/caffe/layers/conv_layer.hpp 和 $CAFFE_ROOT/src/caffe/layers/conv_layer.cpp 或 conv_layer.cu ，将 im2col + gemm 替换为你自己的实现（比如基于 winograd 算法的实现）。
优点：快速迭代，不需要对 Caffe 框架有过多了解，糙快狠准。
缺点：代码难维护，不能 merge 到 caffe master branch，容易给使用代码的人带来困惑（效果和 #define TRUE false 差不多）。
方案二：稍微温柔的解法——千人千面
和方案一类似，只是通过预编译宏来确定使用哪种实现。例如可以保留 ConvolutionLayer 默认实现，同时在代码中增加如下段：
#ifdef SWITCH_MY_IMPLEMENTATION// 你的实现代码
#else
// 默认代码
#endif
这样可以在需要使用该 Layer 的代码中，增加宏定义：
#define SWITCH_MY_IMPLEMENTATION
就可以使用你的实现。而未定义该宏的代码，仍然使用原版实现。
优点：可以在新旧实现代码之间灵活切换；
缺点：每次切换需要重新编译；
方案三：优雅转身——山路十八弯
同一个功能的 Layer 有不同实现，希望能灵活切换又不需要重新编译代码，该如何实现？
这时不得不使用 ProtoBuffer 工具了。
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：
F1002 01:51:22.656038 1954701312 layer_factory.hpp:81] Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: AllPass (known types: AbsVal, Accuracy, ArgMax, BNLL, BatchNorm, BatchReindex, Bias, Concat, ContrastiveLoss, Convolution, Crop, Data, Deconvolution, Dropout, DummyData, ELU, Eltwise, Embed, EuclideanLoss, Exp, Filter, Flatten, HDF5Data, HDF5Output, HingeLoss, Im2col, ImageData, InfogainLoss, InnerProduct, Input, LRN, Log, MVN, MemoryData, MultinomialLogisticLoss, PReLU, Pooling, Power, ReLU, Reduction, Reshape, SPP, Scale, Sigmoid, SigmoidCrossEntropyLoss, Silence, Slice, Softmax, SoftmaxWithLoss, Split, TanH, Threshold, Tile, WindowData)*** Check failure stack trace: *** @ 0x10243154e google::LogMessage::Fail() @ 0x102430c53 google::LogMessage::SendToLog() @ 0x1024311a9 google::LogMessage::Flush() @ 0x1024344d7 google::LogMessageFatal::~LogMessageFatal() @ 0x10243183b google::LogMessageFatal::~LogMessageFatal() @ 0x102215356 caffe::LayerRegistry<>::CreateLayer() @ 0x102233ccf caffe::Net<>::Init() @ 0x102235996 caffe::Net<>::Net() @ 0x102118d8b time() @ 0x102119c9a main @ 0x7fff851285ad start @ 0x4 (unknown)Abort trap: 6
下面给出一个实际案例，走一遍方案三的流程。
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。
首先看头文件：
#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#endif // CAFFE_ALL_PASS_LAYER_HPP_
再看源文件：
#include <algorithm>#include <vector>#include "caffe/layers/all_pass_layer.hpp"#include <iostream>using namespace std;#define DEBUG_AP(str) cout<<str<<endlnamespace caffe {template <typename Dtype>void AllPassLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { const Dtype* bottom_data = bottom[0]->cpu_data(); Dtype* top_data = top[0]->mutable_cpu_data(); const int count = bottom[0]->count(); for (int i = 0; i < count; ++i) { top_data[i] = bottom_data[i]; } DEBUG_AP("Here is All Pass Layer, forwarding."); DEBUG_AP(this->layer_param_.all_pass_param().key());}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[0]) { const Dtype* bottom_data = bottom[0]->cpu_data(); const Dtype* top_diff = top[0]->cpu_diff(); Dtype* bottom_diff = bottom[0]->mutable_cpu_diff(); const int count = bottom[0]->count(); for (int i = 0; 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());}#ifdef CPU_ONLYSTUB_GPU(AllPassLayer);#endifINSTANTIATE_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 = 1; // the layer name optional string type = 2; // the layer type repeated string bottom = 3; // the name of each bottom blob repeated string top = 4; // the name of each top blob // The train / test phase for computation. optional Phase phase = 10; // 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 = 5; // Specifies training parameters (multipliers on global learning constants, // and the name and other settings used for weight sharing). repeated ParamSpec param = 6; // The blobs containing the numeric parameters of the layer. repeated BlobProto blobs = 7; // 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 = 11; // 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 = 8; repeated NetStateRule exclude = 9; // Parameters for data pre-processing. optional TransformationParameter transform_param = 100; // Parameters shared by loss layers. optional LossParameter loss_param = 101; // 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 = 102; optional ArgMaxParameter argmax_param = 103; optional BatchNormParameter batch_norm_param = 139; optional BiasParameter bias_param = 141; optional ConcatParameter concat_param = 104; optional ContrastiveLossParameter contrastive_loss_param = 105; optional ConvolutionParameter convolution_param = 106; optional CropParameter crop_param = 144; optional DataParameter data_param = 107; optional DropoutParameter dropout_param = 108; optional DummyDataParameter dummy_data_param = 109; optional EltwiseParameter eltwise_param = 110; optional ELUParameter elu_param = 140; optional EmbedParameter embed_param = 137; optional ExpParameter exp_param = 111; optional FlattenParameter flatten_param = 135; optional HDF5DataParameter hdf5_data_param = 112; optional HDF5OutputParameter hdf5_output_param = 113; optional HingeLossParameter hinge_loss_param = 114; optional ImageDataParameter image_data_param = 115; optional InfogainLossParameter infogain_loss_param = 116; optional InnerProductParameter inner_product_param = 117; optional InputParameter input_param = 143; optional LogParameter log_param = 134; optional LRNParameter lrn_param = 118; optional MemoryDataParameter memory_data_param = 119; optional MVNParameter mvn_param = 120; optional PoolingParameter pooling_param = 121; optional PowerParameter power_param = 122; optional PReLUParameter prelu_param = 131; optional PythonParameter python_param = 130; optional ReductionParameter reduction_param = 136; optional ReLUParameter relu_param = 123; optional ReshapeParameter reshape_param = 133; optional ScaleParameter scale_param = 142; optional SigmoidParameter sigmoid_param = 124; optional SoftmaxParameter softmax_param = 125; optional SPPParameter spp_param = 132; optional SliceParameter slice_param = 126; optional TanHParameter tanh_param = 127; optional ThresholdParameter threshold_param = 128; optional TileParameter tile_param = 138; optional WindowDataParameter window_data_param = 129; optional AllPassParameter all_pass_param = 155;}
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 = 1 [default = 0];}
在 cpp 代码中，通过
this->layer_param_.all_pass_param().key()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: 10 dim: 3 dim: 227 dim: 227 } }}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.prototxtI1002 02:03:41.667682 1954701312 caffe.cpp:312] Use CPU.I1002 02:03:41.671360 1954701312 net.cpp:49] Initializing net from parameters:name: "AllPassTest"state { phase: TRAIN}layer { name: "data" type: "Input" top: "data" input_param { shape { dim: 10 dim: 3 dim: 227 dim: 227 } }}layer { name: "ap" type: "AllPass" bottom: "data" top: "conv1" all_pass_param { key: 12.88 }}I1002 02:03:41.671463 1954701312 layer_factory.hpp:77] Creating layer dataI1002 02:03:41.671484 1954701312 net.cpp:91] Creating Layer dataI1002 02:03:41.671499 1954701312 net.cpp:399] data -> dataI1002 02:03:41.671555 1954701312 net.cpp:141] Setting up dataI1002 02:03:41.671566 1954701312 net.cpp:148] Top shape: 10 3 227 227 (1545870)I1002 02:03:41.671592 1954701312 net.cpp:156] Memory required for data: 6183480I1002 02:03:41.671605 1954701312 layer_factory.hpp:77] Creating layer apI1002 02:03:41.671620 1954701312 net.cpp:91] Creating Layer apI1002 02:03:41.671630 1954701312 net.cpp:425] ap <- dataI1002 02:03:41.671644 1954701312 net.cpp:399] ap -> conv1I1002 02:03:41.671663 1954701312 net.cpp:141] Setting up apI1002 02:03:41.671674 1954701312 net.cpp:148] Top shape: 10 3 227 227 (1545870)I1002 02:03:41.671685 1954701312 net.cpp:156] Memory required for data: 12366960I1002 02:03:41.671695 1954701312 net.cpp:219] ap does not need backward computation.I1002 02:03:41.671705 1954701312 net.cpp:219] data does not need backward computation.I1002 02:03:41.671710 1954701312 net.cpp:261] This network produces output conv1I1002 02:03:41.671720 1954701312 net.cpp:274] Network initialization done.I1002 02:03:41.671746 1954701312 caffe.cpp:320] Performing ForwardHere is All Pass Layer, forwarding.12.88I1002 02:03:41.679689 1954701312 caffe.cpp:325] Initial loss: 0I1002 02:03:41.679714 1954701312 caffe.cpp:326] Performing BackwardI1002 02:03:41.679738 1954701312 caffe.cpp:334] *** Benchmark begins ***I1002 02:03:41.679746 1954701312 caffe.cpp:335] Testing for 50 iterations.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.681139 1954701312 caffe.cpp:363] Iteration: 1 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.682394 1954701312 caffe.cpp:363] Iteration: 2 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.683653 1954701312 caffe.cpp:363] Iteration: 3 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.685096 1954701312 caffe.cpp:363] Iteration: 4 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.686326 1954701312 caffe.cpp:363] Iteration: 5 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.687713 1954701312 caffe.cpp:363] Iteration: 6 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.689038 1954701312 caffe.cpp:363] Iteration: 7 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.690251 1954701312 caffe.cpp:363] Iteration: 8 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.691548 1954701312 caffe.cpp:363] Iteration: 9 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.692805 1954701312 caffe.cpp:363] Iteration: 10 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.694056 1954701312 caffe.cpp:363] Iteration: 11 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.695264 1954701312 caffe.cpp:363] Iteration: 12 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.696761 1954701312 caffe.cpp:363] Iteration: 13 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.698225 1954701312 caffe.cpp:363] Iteration: 14 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.699653 1954701312 caffe.cpp:363] Iteration: 15 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.700945 1954701312 caffe.cpp:363] Iteration: 16 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.702761 1954701312 caffe.cpp:363] Iteration: 17 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.704056 1954701312 caffe.cpp:363] Iteration: 18 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.706471 1954701312 caffe.cpp:363] Iteration: 19 forward-backward time: 2 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.708784 1954701312 caffe.cpp:363] Iteration: 20 forward-backward time: 2 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.710043 1954701312 caffe.cpp:363] Iteration: 21 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.711272 1954701312 caffe.cpp:363] Iteration: 22 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.712528 1954701312 caffe.cpp:363] Iteration: 23 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.713964 1954701312 caffe.cpp:363] Iteration: 24 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.715248 1954701312 caffe.cpp:363] Iteration: 25 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.716487 1954701312 caffe.cpp:363] Iteration: 26 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.717725 1954701312 caffe.cpp:363] Iteration: 27 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.718962 1954701312 caffe.cpp:363] Iteration: 28 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.720289 1954701312 caffe.cpp:363] Iteration: 29 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.721837 1954701312 caffe.cpp:363] Iteration: 30 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.723042 1954701312 caffe.cpp:363] Iteration: 31 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.724261 1954701312 caffe.cpp:363] Iteration: 32 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.725587 1954701312 caffe.cpp:363] Iteration: 33 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.726771 1954701312 caffe.cpp:363] Iteration: 34 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.728013 1954701312 caffe.cpp:363] Iteration: 35 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.729249 1954701312 caffe.cpp:363] Iteration: 36 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.730716 1954701312 caffe.cpp:363] Iteration: 37 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.732275 1954701312 caffe.cpp:363] Iteration: 38 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.733809 1954701312 caffe.cpp:363] Iteration: 39 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.735049 1954701312 caffe.cpp:363] Iteration: 40 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.737144 1954701312 caffe.cpp:363] Iteration: 41 forward-backward time: 2 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.739090 1954701312 caffe.cpp:363] Iteration: 42 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.741575 1954701312 caffe.cpp:363] Iteration: 43 forward-backward time: 2 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.743450 1954701312 caffe.cpp:363] Iteration: 44 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.744732 1954701312 caffe.cpp:363] Iteration: 45 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.745970 1954701312 caffe.cpp:363] Iteration: 46 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.747185 1954701312 caffe.cpp:363] Iteration: 47 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.748430 1954701312 caffe.cpp:363] Iteration: 48 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.749826 1954701312 caffe.cpp:363] Iteration: 49 forward-backward time: 1 ms.Here is All Pass Layer, forwarding.12.88Here is All Pass Layer, backwarding.12.88I1002 02:03:41.751124 1954701312 caffe.cpp:363] Iteration: 50 forward-backward time: 1 ms.I1002 02:03:41.751147 1954701312 caffe.cpp:366] Average time per layer:I1002 02:03:41.751157 1954701312 caffe.cpp:369] dataforward: 0.00108 ms.I1002 02:03:41.751183 1954701312 caffe.cpp:372] databackward: 0.001 ms.I1002 02:03:41.751194 1954701312 caffe.cpp:369] apforward: 1.37884 ms.I1002 02:03:41.751205 1954701312 caffe.cpp:372] apbackward: 0.01156 ms.I1002 02:03:41.751220 1954701312 caffe.cpp:377] Average Forward pass: 1.38646 ms.I1002 02:03:41.751231 1954701312 caffe.cpp:379] Average Backward pass: 0.0144 ms.I1002 02:03:41.751240 1954701312 caffe.cpp:381] Average Forward-Backward: 1.42 ms.I1002 02:03:41.751250 1954701312 caffe.cpp:383] Total Time: 71 ms.I1002 02:03:41.751260 1954701312 caffe.cpp:384] *** 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。
感谢各位读者提出的宝贵建议和意见，这些都是无价的有监督学习数据集，是激励我不断 update 的 back prop 源动力。
祝各位同学国庆快乐！