Hang

you are right, the PyTorch version effects. please use torch 1.1.0 for training

It seems you need to manually download the pretrained hrnet18. please refer to this issue https://github.com/brandleyzhou/DIFFNet/issues/7

you can refer to [monodepth2](https://github.com/nianticlabs/monodepth2) for environment configuration. please note ours is implemented using torch 1.0, instead of torch 0.4

I have used your command in my environment and it works well. My hardware setting is two GPUs on one node

my apologies for the late reply. I do not have time to retrain a DIFFNet on cityscapes, but it is correct if the cityscape pre-trained and finetuning on kitti model...

- you can delete the 'import hr_network', but I don't think it is the reason. - IF you deploy it when testing, you can let depth_decoder only output a full-scale...

I think so, I guess maybe you have to change the way of initializing a hrnet e.g. hrnet_config.py which uses a yacs lib