salvadog

use cv2.remap

I simply warp img2 back to img1 using the code below, cracks may appear due to missing information ``` h = flow.shape[0] w = flow.shape[1] flow[:,:,0] += np.arange(w) flow[:,:,1] +=...

And did you supervise with images generated with w=0 in stage III? I think fixing quantizer and decoder, and finetuning encoder only with w=1 will affect the originally good results...

@sczhou Any suggestions on stages III training?

> @salvadog Hello, could you share the training codes? Sorry, I can't share my training codes since I implemented it with some non-open-source framework.

Exactly the same error here with python 3.6, pytorch 1.4.0, torchvision 0.5.0, `import kaolin` only works in kaolin root dir, but fails anywhere else. ``` Traceback (most recent call last):...

add `export PYTHONPATH=/kaolin root dir:$PYTHONPATH` to `.bashrc` solved my problem.

> > add `export PYTHONPATH=/kaolin root dir:$PYTHONPATH` to `.bashrc` solved my problem. > > What does 'kaolin root dir' mean? the directory contains README.md