[conversion issue] Convert tiny-yolov3 from darknet format to Caffe format

[Rahn80643]

Hi,
I'm trying to convert a tiny-yolov3 model from .weight to .Caffemodel, and the models are from the author's website: https://pjreddie.com/darknet/yolo/

I followed the tutorial in https://github.com/eric612/MobileNet-YOLO/tree/master/models/darknet_yolov3, and darknet53 YOLOv3 can be converted succefully. But tiny-yolov3 cannot be converted, and I got the following error message:

Check failed: top_shape[j] == bottom[i]->shape(j) (24 vs. 26) All inputs must have the same shape, except at concat_axis.

I've also tested the model in darknet53 project, and tiny-yolov3 model can be execute succesfully. Should the settings in darknet2caffe.py be modified ?

[pra-dan]

Hey @Rahn80643 Did you get past this problem ? Also, have you tried it on v2-tiny models ?

[Rahn80643]

@PrashantDandriyal,
sorry I have not solved the problem yet, and I only tried to convert v3-tiny model

[troyliu0105]

This is problem of downsampling or stride in conv. Did you check the maxpooling roundmode or conv stride?