contentsmaybehot / rivalsofaethersupervisedlearning Goto Github PK

Make it so the model fits on the GPU.

Seriously watch the graphs. You can see the exact periods when file I/O and unpack_sample take place. It makes big ol valleys.

lstm.py lines 80-84, maybe make a generator.

There's loads of identical code that could be extracted.

Write tests BEFORE extracting into functions.

We have all this useless RAM sitting around, we may as well use it to build a dataframe.

Each video can have a maximum of n frames. At a rate of 10 frames per second, this corresponds with up to n/10 seconds of gameplay. A good starting value is 10. If a video's length does not divide evenly by n, then the last video will be dropped. This will allow us to specify the timesteps in our input shape and is a prerequisite for making the LSTM stateful.

This must be factored into the len(sequence) calculation.

Stages are enumerated in replay_management.py. Some of the enumerations are missing, however. We need to generate replays on these missing stages so that we can enumerate them. I don't have enough coins to unlock all of the stages.