mm1327 / CornerNet-Caffe

Use caffe to implement CornerNet inference

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CornerNet Caffe

This is just inference code of Cornernet in caffe. And I implement corner pool and unsample layer by myself. Please tell me if there is something wrong.

The original pytorch implementation repository is here

Getting Started

Please use pytorch0.4 because I found that using other version of pytorch to convert model will be wrong.

How to use

CC5.0

I use cc5.0 which is a good extended frame of caffe and support windows. You can watch the introduction by README

PytorchToCaffe

I convert the original pytorch model to caffemodel by PytorchToCaffe. And I add some new feature in the pytorch_to_caffe.py which can find in my pytorch_to_caffe code. However, the caffemodel I converted has some problems, so the output of the inference code is wrong, I'm still looking for the answer.

I have tested the converted model and code. And it's correct. The pytorch model I converted is the offical CornerNet_500000.pkl. The caffemodel and prototxt can be downloaded by BaiduNetDisk. The extract code is 34p9.(The public link is in maintenance).

Why use my own relu layer

The weights of one conv layer in pytorch is too big which is over the boundary of float32. So if inputs times this weights, the result will be NaN. Howerver when the output of this conv layer crosses the caffe relu layer, NaN is still NaN while crossing the pytorch relu layer NaN will be zero.

To do

  • Train code

About

Use caffe to implement CornerNet inference


Languages

Language:C++ 62.8%Language:Python 37.2%