HEPESU

Maybe you run your prediction too early. You can bind "initProgress" event to see whether the model is initialized or not. If the model is not fully initialized, it would...

Two methods to use AMP, I prefer 2 in my case. 1. Add `custom_fwd` and `custom_bwd` to use gradient checkpoint with AMP, which speeds up 1.2x (A5000) and costs same...

Nice idea, you can try combining HED like networks and some semantic segmentation networks (FCN?) to do the work which can outputs edge with semantic tags.

You can ignore InputLayer in kerasify.py with sample code: ```python model_layers = [l for l in model.layers if type(l).__name__ not in ['Dropout','InputLayer']] ``` If you are using VGG16 with top,...