有人成功跑通了yolov2吗?

[复制链接] · 发表于 2019-3-26 14:31:22

我用的caffe版本试的，缺少reorg层的支持。我试着用我们剪掉的版本，发现无法标出物体。有人跑通过吗？谢谢

只看该作者 · 发表于 2019-3-27 02:37:21

I tried the Yolov2 from darknet. It works great on rknn. Any specific reason do you want to use the caffe version of Yolov2?

只看该作者 · 发表于 2019-3-27 09:29:19

yes， we train it in caffe, it can't support reorg layer. can we add the custom layer for rknn?

只看该作者 · 发表于 2019-3-27 13:08:07

guigen80 发表于 2019-3-27 09:29
yes， we train it in caffe, it can't support reorg layer. can we add the custom layer for rknn?

I'll leave RK people to comment how to add custom OP support for rknn. Personally I won't think it will be easy even if it is supported.

Why do you think rknn doesn't support reorg layer? I don't think I had any issue when converting yolov2 (which contains reorg layer) from darknet.

只看该作者 · 发表于 2019-3-27 14:21:09

T google.protobuf.text_format.ParseError: 915:5 : Message type "caffe.LayerParameter" has no field named "reorg_param".

when to transform , it report the error, so I think it can't support reorg layer. I use the model from yolo9000