I just went thru all the layers and activations that are present in the Darknet framework and the proposed h-sigmoid and h-swish activations and Adaptive Avg Pooling layer are not present. The only way to train a YOLOV4 with MobileNetV3 backbone would be to write them in this framework... Then it would be way easier to make a cfg file.
I just went thru all the layers and activations that are present in the Darknet framework and the proposed h-sigmoid and h-swish activations and Adaptive Avg Pooling layer are not present. The only way to train a YOLOV4 with MobileNetV3 backbone would be to write them in this framework... Then it would be way easier to make a cfg file.