richzhang/colorization-pytorch

Model training Loss graph

Ugness opened this issue · 6 comments

Ugness commented

newplot

Plot.ly Link: https://plot.ly/~Ugness/1/

I am training the model with ILSVRC2012 training set and same options as your implementation and my loss graph looks like above.
I am afraid that my model's loss reduces correctly. Can you check this loss graph or share your loss graph?
Thanks.

richzhang commented

Thanks for the message. I'll look into it in the next week or so.

  • 👍1
richzhang commented

My apologies for the delay. I'm caught up with a few things and will have to look into this more after CVPR deadline (mid-Nov).

Ugness commented

Thanks.

Samir55 commented

Hello author,
I have the same problem above, what is the solution?
Thanks in advance.

richzhang commented

I believe I did not do the initialization properly. I am looking into it. Thanks for your patience!

richzhang commented

The problem was with the initialization. In the original paper, I used k-means initialization, but didn't have the code for pytorch.

I just pushed a change. If you rerun python make_ilsvrc_dataset.py, it will make a directory dataset/ilsvrc2012/train_small/, which contains 10 imagenet classes. Then rerun bash scripts/train_siggraph.sh again. It will first run 100 epochs on this small training set as "initialization", before training on the full training set.

It does still train quite slowly. Better networks/loss functions would be a good update to the project.

Thanks for your patience.

  • 👍2