Guidance with SpaceInvaders

[ThomasRot]

Hi folks,

I trained MarioNette with a data set of 2^15 128x128 scenes of SpaceInvaders, but only the background is reconstructed. Do you have some ideas, about which kind of parameters I should modify to reach the results described in the paper?

Thanks a lot for your insights!

[dmsm]

Thanks for your interest! What parameters are you using to run the training script? In particular, you want anchor grid resolution to be such that each patch roughly fits a single sprite. Passing in --layer_size 16 should be a reasonable choice. Let me know if this helps.

[ThomasRot]

Thanks for your answer, I'll try around with that. Are the used hyperparameters for each experiment in the paper documented somewhere? I can't seem to find them.

[dmsm]

Yes, all hyperparameters are specified in the paper. In particular, Section 3.2 states that patch size k = 16 is used for Mario, whereas k = 32 is used for Space Invaders. Note that k = canvas_size / layer_size.

[ThomasRot]

Thanks a lot. I had some success with num_layers = 2, num_classes = 54 (following the image of the dictionary in the paper only) and layer_size = 16 (like written two comments above, though k{32} = canvas_size{128} / layer_size would imply layer_size = 4 if I'm not mistaken, but that did not work that well)