使用netron对mnist网络结构分析「建议收藏」

全栈程序员站长

发布于 2022-11-01 16:34:11

8650

发布于 2022-11-01 16:34:11

文章被收录于专栏：全栈程序员必看

基于libonnx环境简要分析一下mnist网络算子结构，关于环境搭建可以参考前面两篇文章：

xboot大神的libonnx环境搭建

使用netron实现对onnx模型结构可视化

本文主要目的是搞清楚mnist各层之间数据shape的变化情况,关于什么是shape，引用一本书中的介绍:

“在tensorflow中，使用张量来表示计算图中的所有数据，张量在计算图的节点之间流动，张量可以看成N维数组，而数组的维数就是张量的阶数。因此，0阶张量对应标量数据，1阶张量对应一维数组，也就是向量。二阶张量对应二维数组，也就是矩阵，以此类推，N阶张量对应n维数组，例如，一张RGB图像可以表示为3阶张量，而多张RGB图构成的数据可以表示为4阶张量。shape(形状）代表的就是张量的一种属性，当然还有其他属性，比如数据类型等等”

再算子执行前面打断点，依次观察输入数据和输出数据的大小：

(gdb) b 2124
Breakpoint 2 at 0x555555560ef8: file onnx.c, line 2124.
(gdb) display n->inputs[0]->ndata
(gdb) display n->outputs[0]->ndata
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 2560
2: n->outputs[0]->ndata = 2560
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 784
2: n->outputs[0]->ndata = 6272
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 6272
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 6272
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 1568
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 1568
2: n->outputs[0]->ndata = 3136
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 3136
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 3136
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 256
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 256
2: n->outputs[0]->ndata = 256
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 256
2: n->outputs[0]->ndata = 10
(gdb) c
Continuing.

Breakpoint 2, onnx_run (ctx=0x555555a501e0) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 10
2: n->outputs[0]->ndata = 10
(gdb) c
Continuing.

可以看出一个简单的规律，就是前一级网络的输出size等于后一级网络的输入size.

对照网络，可以完全对应的上：

将shape打印出(由dims表示），可以看出和上图完全吻合。（图中一维向量表示为1*N,也看成2维的shape).

Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 2560
2: n->outputs[0]->ndata = 2560
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 2
(gdb) c
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 784
2: n->outputs[0]->ndata = 6272
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) c
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 6272
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) c
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 6272
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) c
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 6272
2: n->outputs[0]->ndata = 1568
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) c
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 1568
2: n->outputs[0]->ndata = 3136
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 3136
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 3136
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 3136
2: n->outputs[0]->ndata = 256
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 4
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 256
2: n->outputs[0]->ndata = 256
3: n->inputs[0]->ndim = 4
4: n->outputs[0]->ndim = 2
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 256
2: n->outputs[0]->ndata = 10
3: n->inputs[0]->ndim = 2
4: n->outputs[0]->ndim = 2
(gdb) 
Continuing.
Breakpoint 3, onnx_run (ctx=0x5555559ff250) at onnx.c:2124
2124					n->operator(n);
1: n->inputs[0]->ndata = 10
2: n->outputs[0]->ndata = 10
3: n->inputs[0]->ndim = 2
4: n->outputs[0]->ndim = 2
(gdb) 
Continuing.

然后再以ndim为上限，索引dims，还是以reshape为例：