Tip for satisfying the GANS with Hands grader

I picked out my best guesses and submitted but failed the grader. However, the grader feedback shows the scores and some of them were really low. I submitted a second time just selecting the first 16, and passed. My tip is that it doesn’t matter whether the pictures look like hands to you, it only matters whether they are classified as hands by the grader model. Don’t spend too much time trying to compare them yourself…just submit and adjust based on the feedback. HTH.

Ps 60 epochs was enough for me. I might have tried more but I was in colab GPU jail and 60 takes 4 1/2 hours with no acceleration. If you have the GPU budget it might help to try more epochs.

ps: in another thread @balaji.ambresh proposed collecting the discriminator predictions which could be used both to implement an early stopping rule and to ensure the best 16 images were submitted to the grader. Much more efficient than the ‘eyeball’ approach suggested in the notebook.


The tip is posted in this thread.


FYI, CPU took around 3.5 hours to complete.