Bea07

I have problems too relating to this. May I ask exactly what are the reasons why this convergence may not happened? I am using this package for perform a DEM...

I got same problems and calling eval() is the reason.

Hi, is there a pretrained model aerial to map? there is the dataset but missing the model, is that correct?

The issue of accuracy loss is due to batch normalization layer. BN in inference mode performs differently than training. The test.py script is not launching an inference however you set...

I understood the issue is coming from a pytorch function model.eval() that causes a freeze of some operations in the model to inference mode that drop accuracy. Es. batch normalization.

> I am not familiar with onnx conversion. But the default CycleGAN uses instance normalization. yep indeed the pretrained cycleGAN es horse2zebra don't have accuracy loss ( for pix2pix GAN...

i am having same problem but using venv. I have them installed on the environment `cycler 0.11.0 debugpy 1.6.7 decorator 5.1.1 distlib 0.3.6 earthpy 0.9.4 eccodes 1.6.1 enpt-enmapboxapp 0.7.6 entrypoints...