Skip to content

Latest commit

 

History

History
73 lines (51 loc) · 7.43 KB

CHANGELOG.md

File metadata and controls

73 lines (51 loc) · 7.43 KB
Raw

Change Log

v0.3.1

Added

  • A pyproject.toml file has been added in compliance with PEP 660. While the inclusion of a setup.py file is not deprecated, its use as a command-line tool, such as in the legacy setup.py develop method for performing an editable installation is now deprecated.

Changed

  • chess_transformers.train.datasets.ChessDataset was optimized for large datasets. A list of indices for the data split is no longer maintained or indexed in the dataset.
  • The TRAINING_CHECKPOINT parameter in each of chess_transformers.configs.models was set to None to reflect the correct conditions for beginning training of a model.
  • Dynamic shape tracing is disabled for the compilation of CT-ED-45 to prevent memory leaks as seen in #16.
  • References to torch.cuda.amp.GradScaler(...) have been replaced by torch.amp.GradScaler(device="cuda", ...) following its deprecation.

v0.3.0

Added

  • There are 3 new datasets: ML23c, GC22c, and ML23d.
  • A new naming convention for datasets is used. Datasets are now named in the format "[PGN Fileset][Filters]". For example, LE1222 is now called LE22ct, where LE22 is the name of the PGN fileset from which this dataset was derived, and "c", "t" are filters for games that ended in checkmates and games that used a specific time control respectively.
  • CT-EFT-85 is a new trained model with about 85 million parameters.
  • chess_transformers.train.utils.get_lr() now accepts new arguments, schedule and decay, to accomodate a new learning rate schedule: exponential decay after warmup.
  • chess_transformers.data.prepare_data() now handles errors where there is a mismatch between the number of moves and the number of FENs, or when the recorded result in the PGN file was incorrect. Such games are now reported and excluded during dataset creation.

Changed

  • The LE1222 and LE1222x datasets are now renamed to LE22ct and LE22c respectively.
  • All calls to chess_transformers.train.utils.get_lr() now use the schedule and decay arguments, even in cases where a user-defined decay is not required.
  • chess_transformers.train.datasets.ChessDatasetFT was optimized for large datasets. A list of indices for the data split is no longer maintained or indexed in the dataset.
  • Dependencies in setup.py have been updated to newer versions.
  • Fixed an error in chess_transformers.play.model_v_model() where a move would be attempted by the model playing black even after white won the game with a checkmate.
  • Fixed the EVAL_GAMES_FOLDER parameter in the model configuration files pointing to the incorrect folder name chess_transformers/eval instead of chess_transformers/evaluate.
  • Fixed an error in chess_transformers.evaluate.metrics.elo_delta_margin() where the upper limit of the win ratio for the confidence interval was not capped at a value of 100%.
  • All calls to torch.load() now use weights_only=True in compliance with its updated API.

v0.2.1

Changed

  • All model checkpoints, datasets, and logs have now been moved to Microsoft Azure Storage for more reliable access.

v0.2.0

Added

  • ChessTransformerEncoderFT is an encoder-only transformer that predicts source (From) and destination squares (To) squares for the next half-move, instead of the half-move in UCI notation.
  • CT-EFT-20 is a new trained model of this type with about 20 million parameters.
  • ChessDatasetFT is a PyTorch dataset class for this model type.
  • chess_transformers.data.levels provides a standardized vocabulary (with indices) for oft-used categorical variables. All models and datasets will hereon use this standard vocabulary instead of a dataset-specific vocabulary.

Changed

  • The LE1222 and LE1222x datasets no longer have their own vocabularies or vocabulary files. Instead, they use the standard vocabulary from chess_transformer.data.levels.
  • The LE1222 and LE1222x datasets have been re-encoded with indices corresponding to the standard vocabulary. Earlier versions or downloads of these datasets are no longer valid for use with this library.
  • The row index at which the validation split begins in each dataset is now stored as an attribute of the encoded_data table in the corresponding H5 file, instead of in a separate JSON file.
  • Models CT-E-20 and CT-ED-45 already trained with a non-standard, dataset-specific vocabulary have been refactored for use with the standard vocabulary. Earlier versions or downloads of these models are no longer valid for use with this library.
  • The field move_sequence in the H5 tables has now been renamed to moves.
  • The field move_sequence_length in the H5 tables has now been renamed to length.
  • The load_assets() function has been renamed to load_model() and it no longer returns a vocabulary — only the model.
  • The chess_transformers/eval folder has been renamed to chess_transformers/evaluate.
  • The Python notebook lichess_eval.ipynb has been converted to a Python script evaluate.py, which runs much faster for evaluation.
  • Fairy Stockfish is now run on 8 threads and with a hash table of size 8 GB during evaluation instead of 1 thread and 16 MB respectively, which makes it a more challenging opponent.
  • Evaluation results have been recomputed for CT-E-20 and CT-ED-45 against this stronger Fairy Stockfish — they naturally fare worse.

Removed

  • The environment variable CT_LOGS_FOLDER no longer needs to be set before training a model. Training logs will now always be saved to chess_transformers/training/logs.
  • The environment variable CT_CHECKPOINT_FOLDER no longer needs to be set before training a model. Checkpoints will now always be saved to chess_transformers/checkpoints.
  • The environment variable CT_EVAL_GAMES_FOLDER no longer needs to be set before evaluating a model. Evaluation games will now always be saved to chess_transformers/evaluate/games.