[Feature,Refactor] Chess improvements: fen, pgn, pixels, san #2702
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/2702
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
vmoens
added
enhancement
torchrl/envs/custom/chess.py
Outdated
| # Generate the PGN string | ||
| pgn_string = str(game) | ||
| return pgn_string | ||
|
|
||
| @classmethod | ||
| def _get_fen(cls, tensordict): |
There was a problem hiding this comment.
Maybe this function can be removed, in favor of a direct TensorDict.get call?
There was a problem hiding this comment.
I was thinking that one day we might let people choose the fen / pgn key for themselves but it's true that as of now it's not incredibly useful
torchrl/envs/custom/chess.py
Outdated
| fen = self._get_fen(tensordict).data | ||
| dest = tensordict.empty() | ||
| if self.include_pgn: | ||
| fen = self._get_pgn(tensordict).data |
There was a problem hiding this comment.
| fen = self._get_pgn(tensordict).data | |
| pgn = self._get_pgn(tensordict).data |
Also _get_pgn doesn't exist, but I figure you realize that
| @classmethod | ||
| def _get_tensor_image(cls, board): | ||
| try: | ||
| svg = board._repr_svg_() |
There was a problem hiding this comment.
Maybe we could have another option that has a more minimal representation as well, if we want to train without vision. For instance, OpenSpiel's chess env has a representation where each square of the board is given a size 20 array of ones and zeros, which is somehow used to represent which piece is in that square and which color it is
There was a problem hiding this comment.
Sure, I was using this because it's builtin - but a more minimal representation could also be cool!
Stack from ghstack (oldest at bottom):