TODO.md

2024.02.20

Inference / Engine

• Script to play against the NN
• UCI interface?
• Duels against Stockfish for Benchmarking?

Managing model data / training runs:

• Have a helper class to organize models and training data?

Training script:

• Add proper resume function
• Add arguments for mixed precision training:
  • enable=True/False
  • arguments for GradScaler
• Save GradScaler state
• Save more infos in model_dict, such as
  • batchsize for each checkpoint
  • learning rate for each checkpoint
  • save hyperparams (model params) for automated model loading
• Support to change batchsize and learning rate over time (+ resume)

Config Files?

• Training script will have lots of args. Might need YAML config or so.

DataLoader:

• Add support for old DataLoader

Some basic ease-of-use extensions to the scripts:

• In speedtest.py and hf_trainer(_multi).py, add:
  • argparse launch args for
    • CPU or CUDA + optional device IDs
    • batchsize
    • num workers
  • print timestamp of run and params used for easier eyeballing and separation of runs
  • optional logging of ^ and what not

Hyperparam search/comparisons for optimal training perf:

• Add code to find best combined values for batchsize, device ids (1-vs-m), num_workers
• Add code to compare and find best activation function wrt model accuracy, maybe also check best convergence rates?
• Test relation between width, depth and speed. Best combo for tree search?
• BatchNorm instead of LayerNorm?

Open problems

• Predict scores for all legal moves at once.
• Cleaner embedding style for convolutional architecture.
• Bitboards to get legal moves.
• Tree search (maybe slow one in Python first is ok?)
• Test standard fully-connected transformer on longer training run?

Other things to test: