Clear instruction of training for custom data?

[innat-asj]

The performance of MI-GAN is good compared to its earlier works. I have tired off-the-shelf model and it worked well and thus promising. Now, I am looking for to fine-tune it on our custom data to make it more useful. My custom data consist of image and binary mask which contains the region of interest to be removed.

However, the instruction on the training section is somewhat unclear how should we adopt MI-GAN for such custom data training. Would you please spare some time to improve the documentation? For example, how to fine tune MI-GAN. What type of data is needed? How they should be organized? Some tips and tricks for better performance?

The performance of this model is promising and hopefully we can finally use it for our cause. Thank you for your hard work. You have done a great work here.