Providing Open Source pre-trained models using 35K+ images

[Fmstrat]

Hi,

I used to run a photography business and have ownership of over 35,000 high resolution raw (uncompressed CR2/RAW) images taken from DSLRs mostly of individuals or groups of people. I'm considering creating pre-trained models to provide for this project via open source, but would like to know the best way/method to train (I'm new to all this).

I would be running training through Docker, similar to the PR I recently added (#81). Any help or recommendations on training methods would be appreciated.

Thanks.

[krasserm]

Hi @Fmstrat, thanks for #81 and this initiative, looks great! I have limited connectivity this week but will review your pull request next week.

[krasserm]

Reviewed ... awesome PR, thanks a lot! Maybe a later addition could use argparse in the training and inference scripts to set some of the hyper-parameters but your initial PR is an excellent starting point.

Would you like me to host your pre-trained models or would you like to host them yourself? Do you plan making the dataset public (or is it already public)? We could add a new data provider class then.

[Fmstrat]

Oh yea, that's a good idea (argparse). I can find somewhere to host them, mainly I just need to figure out the best way to create them (which you're helping with in the PR). As for the data set, no I won't be making that public, as while I own rights to the images, they were not all public events.

[Fmstrat]

Now that I think about it, in addition to providing a link to the models here, I will also look to provide them through https://github.com/kritiksoman/GIMP-ML via @kritiksoman once that project is updated to support Python 3. Once I'm done here I'll work over there to figure out how to get them in. It would be nice to be able to use the models directly in GIMP.

[krasserm]

This sounds great. Using EDSR as generator might improve results a bit ... According to the paper of https://github.com/kritiksoman/GIMP-ML, they are still using the original SRGAN generator.