Output of the inpainter gives a gray boundary about generative-inpainting-pytorch HOT 8 CLOSED

Yes, the results are not very good when testing on the converted TF model.

The TF model was converted by semi-manually matching the weight/shapes/names between the TensorFlow variables and PyTorch model weights according to here. I don't know if there are some incompatible part. And there are also some operations that don't have weights like contextual attention, which may incompatible with the original one.

from generative-inpainting-pytorch.

Thanks for the information. In that case, do you also have the pre-trained weights for the Pytorch model?

from generative-inpainting-pytorch.

Yes, I just updated the PyTorch model download links in the README. Now you can download and test on it.

from generative-inpainting-pytorch.

Perfect, thanks. I will give it a try.

from generative-inpainting-pytorch.

The pre-trained model weights are giving size mismatch errors for different fine_generator layers. Any comments?

from generative-inpainting-pytorch.

Em... so embarrassing.

The pre-trained model weights are trained with a bug version --- some convolution layers' channel numbers are wrong in that version. The bugs were fixed by this pull request. So you need to change some channel numbers to load the model. Sorry about those problems.

After fixing the bugs, I haven't trained the model again. I prepare to train a model from scratch later but it will cost a long time.

from generative-inpainting-pytorch.

from generative-inpainting-pytorch.

@chirag126 New model released with commit b2d7de.

from generative-inpainting-pytorch.