深度学习框架 源码阅读 调试 分析 c++ tiny-dnn (1)

layer类层次

前向传播调用栈举例

后向传播调用栈举例

前向传播

后向传播

参与后向传播的有:

1. 网络network(图graph或序列seq),
2. 层layer（激活函数层, cnn层, rnn层，全连接层, … )

参与后前传播的有：

1. 网络network(图graph或序列seq),
2. 层layer（激活函数层, cnn层, rnn层，全连接层, … )

去掉并行选项, 看到完整调用栈

//config.h

//去掉以下这行的注释,  即 去掉了并行选项,  变成单线程直接调用. 
 #define CNN_SINGLE_THREAD

去掉并行选项后, 再编译, 调式, 看到以下完整调用栈

#完整调用栈文本

tiny_dnn::kernels::avx_conv2d_5x5_kernel<tiny_dnn::aligned_allocator<float,64> >(const tiny_dnn::core::conv_params &,const std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,const std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,const std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,const bool) conv2d_op_avx.h:178
<lambda_93cd752631eaddbf3fd5164277579d9c>::operator()(unsigned long long) conv2d_op_avx.h:461
tiny_dnn::for_i<unsigned __int64,<lambda_93cd752631eaddbf3fd5164277579d9c> >(bool,unsigned long long,<lambda_93cd752631eaddbf3fd5164277579d9c>,unsigned long long) parallel_for.h:187
tiny_dnn::kernels::conv2d_op_avx(const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > &,const std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,const std::vector<float,tiny_dnn::aligned_allocator<float,64> > &,std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > &,const tiny_dnn::core::conv_params &,const bool) conv2d_op_avx.h:460
tiny_dnn::Conv2dOp::compute(tiny_dnn::core::OpKernelContext &) conv2d_op.h:46
tiny_dnn::convolutional_layer::forward_propagation(const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *> > &,std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *> > &) convolutional_layer.h:298
tiny_dnn::layer::forward() layer.h:554
tiny_dnn::sequential::forward(const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > >,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > > > &) nodes.h:296
tiny_dnn::network<tiny_dnn::sequential>::fprop(const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > >,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > > > &) network.h:203
tiny_dnn::network<tiny_dnn::sequential>::train_onebatch<tiny_dnn::mse,tiny_dnn::adagrad>(tiny_dnn::adagrad &,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,int,const int,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *) network.h:924
tiny_dnn::network<tiny_dnn::sequential>::train_once<tiny_dnn::mse,tiny_dnn::adagrad>(tiny_dnn::adagrad &,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *,int,const int,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > *) network.h:897
tiny_dnn::network<tiny_dnn::sequential>::fit<tiny_dnn::mse,tiny_dnn::adagrad,<lambda_2167b27d0bf24d6ba6fbafdc977e3835>,<lambda_dfd0aadcd2585f275bac64861e8acb9b> >(tiny_dnn::adagrad &,const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > >,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > > > &,const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > >,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > > > &,unsigned long long,int,<lambda_2167b27d0bf24d6ba6fbafdc977e3835>,<lambda_dfd0aadcd2585f275bac64861e8acb9b>,const bool,const int,const std::vector<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > >,std::allocator<std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > > > &) network.h:863
tiny_dnn::network<tiny_dnn::sequential>::train<tiny_dnn::mse,tiny_dnn::adagrad,<lambda_2167b27d0bf24d6ba6fbafdc977e3835>,<lambda_dfd0aadcd2585f275bac64861e8acb9b> >(tiny_dnn::adagrad &,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > &,const std::vector<unsigned __int64,std::allocator<unsigned __int64> > &,unsigned long long,int,<lambda_2167b27d0bf24d6ba6fbafdc977e3835>,<lambda_dfd0aadcd2585f275bac64861e8acb9b>,const bool,const int,const std::vector<std::vector<float,tiny_dnn::aligned_allocator<float,64> >,std::allocator<std::vector<float,tiny_dnn::aligned_allocator<float,64> > > > &) network.h:305
train_lenet(const std::basic_string<char,std::char_traits<char>,std::allocator<char> > &,double,const int,const int,backend_t) train.cpp:115
main(int,char **) train.cpp:212
invoke_main() 0x00007ff6939e9629
__scrt_common_main_seh() 0x00007ff6939e950e
__scrt_common_main() 0x00007ff6939e93ce
mainCRTStartup(void *) 0x00007ff6939e96be
BaseThreadInitThunk 0x00007ff903867034
RtlUserThreadStart 0x00007ff9043e2651

很明显的层次:

1. train
   1. train_lenet
   2. sequential.train
   3. sequential.fit
   4. sequential.train_once
   5. sequential.train_onebatch
2. forward
   1. sequential.fprop : 这是缩写 forward_propagation
   2. sequential.forward
   3. layer.forward
   4. convolutional_layer.forward_propagation
   5. Conv2dOp.compute
3. kernel
   1. kernels.conv2d_op_avx
   2. for_i
   3. operator()
   4. kernels.avx_conv2d_5x5_kernel

由此可见, 所谓kernel就是实际干活的, 实际执行具体计算过程的. 所以应该有各种各样的kernel

完成各种各样具体计算任务的部件 叫 kernel

tiny-dnn中有以下这些kernel:

注意 pytorch 中的 kernel 也是这个意思

nodes: 图或序列

nodes : 多个节点, 即 表示 网络结构 的顶层标记类. 网络结构 有 : 图结构、序列结构

compute: 前向计算入口统称, 且 即将转向kernel