Training for pytorch

[FarmerF]

Hi All,

First of all thanks for the amazing project. I got it more or less working (see my how to in #10).

Thing is the caffe results seem much better than the pytorch one. Better in the sense that caffe colors in the bands of the mortar which pytorch colors outside the bands as well). But the caffe version runs out of memory with a predicted requiirement of 10GB (6.7GB used in cpu_mode) while the pytorch version uses only 1GB. I have followed the data training steps from colorization-pytorch (downloaded pretrained model) but this doesn't seem to change much.

Anything I am missing?

The descriptions on the main site:

Please find the Local Hints Network training code in the colorization-pytorch repository.

and

Please find a PyTorch reimplementation of the Local Hints Network training code in the colorization-pytorch repository.

didn't really help much. Any clarification would be much appreciated.

Yours,

Frank

[junyanz]

@richzhang

[richzhang]

The PyTorch version is an approximate reimplementation and performs a bit differently. I will have some more time after CVPR deadline to look into it further. Thanks for your patience.