#epoch and validation

[Bonnie970]

Hi again,

Could you share how many epochs you trained to get result ./script/pred_alpha.png? My GPU is pretty slow, my model performs not as well as what mentioned in the paper after 10 epochs. FYI, the paper author said they only used 1 Titan X for training. Also, I notice that the validation code in matting_unpool.py was commented out. How did you tell the training results without validation?

Thank you!

[protossw512]

I forgot how many epochs I trained, but it should be much more than 10. I am not able to access my server for now and in next month, so I cannot get detailed information for you.

The main reason of slow training could be data loading and processing, since this code did not use tensorflow's parallel dataloader, and some intermediate process cannot be down in cpu. I also downsampled original images, some of they have very high resolution.

I did not use any metric to validate, I just eyeballed the results and found them very close to ground truth matting results, but still not as good as ones in the paper. I mainly used this for segmentation refinement, so I did not pay too much attention for image matting.

I modified some part of the code afterwards, but still, I am not able to commit them right now. I would suggest you really try to convert this model to pytorch, and train from there, it should be much more easy to play around. The core idea of this paper is its loss function, network architecture is pretty outdated.

Are you doing this for research or production?

[Bonnie970]

Hello, I am doing this for research. Thank you for your insights and suggestions, I am quite new to the field.

I did experiments with both original image and down-sampled ones, the difference is minor. I will update you if I found anything interesting in the future experiments!