对于不同平台下实现的训练模型,复现或者使用TensorRT加速对应网络的时候,需要了解网络结构。但不同框架之间可视化网络层方法不一样,这样造成了一些麻烦。
网上有可视化工具挺好,但是没有标注对应层的序号。
Tool Netron: https://github.com/lutzroeder/Netron
MXNet model file: (download in mxnet net)
MXNet_Yolo3-symbol.json
You can see that it is different from print(net) of pyhton print the model.
There are some "elemwise_add" layers.
But there are some info not useful for layers structure.
{
"op": "null",
"name": "data",
"inputs": []
},
{
"op": "null",
"name": "darknetv30_conv0_weight",
"attrs": {
"__dtype__": "0",
"__lr_mult__": "1.0",
"__shape__": "(32, 3, 3, 3)",
"__storage_type__": "0",
"__wd_mult__": "1.0"
},
"inputs": []
},
{
"op": "Convolution",
"name": "darknetv30_conv0_fwd",
"attrs": {
"dilate": "(1, 1)",
"kernel": "(3, 3)",
"layout": "NCHW",
"no_bias": "True",
"num_filter": "32",
"num_group": "1",
"pad": "(1, 1)",
"stride": "(1, 1)"
},
"inputs": [[0, 0, 0], [1, 0, 0]]
},
{
"op": "null",
"name": "darknetv30_batchnorm0_gamma",
"attrs": {
"__dtype__": "0",
"__init__": "ones",
"__lr_mult__": "1.0",
"__shape__": "(32,)",
"__storage_type__":