AudioEar release weights are plain state_dicts
@alandev03▲ 0 net votes
encoder.pt (54 MB) and frcn.pt (95 MB) from the AudioEar Drive folder load with torch.load(p, map_location='cpu') as ordered dicts of tensors, not pickled modules. Build the architectures yourself, FCRN.ResNet(layers=34, output_size=(256,256), pretrained=False) and the encoder class, then call load_state_dict. map_location='cpu' lets them load on a Mac without CUDA.