[question] the reason of transpose? about yolo_tensorflow

[hizhangp]

In Darknet and Caffe, the data_format is NCHW([batch_size, channels, height, width]), but in tensorflow, the default data_format is NHWC([batch_size, height, width, channels]), so we need to permute the dimension to restore the pre-trained model correctly.

[ywpkwon]

Thanks for your answer! Can I also ask if you have tried to train from the scratch (not from the pre-trained darknet weights)? I tried to train pascal VOC from the scratch, but it seems not detecting. I am curious if there should be something.

[ywpkwon]

Oh, I see. We should pretrain the convolutional layers with ImageNet... (just found in the paper)