Strategy:

• Get dataset
• Build vocab (64x64 + promotions)
• Stream process PGN (with filter)
  • Downsample opening data
• Store as (board tensor - 3d tensor, move index)
  • (17, 8, 8) -> shows entire board for each of the pieces (6 per side) + castleing / en passant
• Augment data? (horizontal flip)
• Save each as big pytorch tensors
• Architecture -> CNN (layers takes 3D, flattened at the end, output 4672 possible moves)
• Train with PyTorch (dataloaders, dataset, adam optimizer, cross entropy loss)
• Integrate with modal
• Inference engine
  • Get legal moves, choose highest probability,