Refactoring #19
Refactoring #19
Conversation
|
|
||
| set_seed(config.train_seed) | ||
|
|
||
| def env_fn(): |
There was a problem hiding this comment.
Let's move it to a standalone function out of main
There was a problem hiding this comment.
then we have to add character as an argument, it's not very nice
| tmp_env = env_fn() | ||
| eval_env = AsyncVectorEnv( | ||
| env_fns=[env_fn for _ in range(config.eval_processes)], | ||
| shared_memory=True, |
There was a problem hiding this comment.
Add a comment explaining why this is needed
| seed=config.train_seed, | ||
| add_next_step=False | ||
| ) | ||
| tp = ThreadPoolExecutor(max_workers=14) |
There was a problem hiding this comment.
this should either be a config value or something automatic based on the number of cores/processors we have
There was a problem hiding this comment.
fixed, now as config value
katakomba/env.py
Outdated
| Returns score normalized against AutoAscend bot scores achieved for this exact character. | ||
| """ | ||
| if self.character.count("-") != 2: | ||
| raise ValueError("Reference score not provided for this character.") |
There was a problem hiding this comment.
typo: Reference score is not provided...
katakomba/env.py
Outdated
|
|
||
| def get_dataset(self, scale: str = "small", **kwargs): | ||
| if self.character.count("-") != 2: | ||
| raise ValueError("Reference score not provided for this character.") |
|
|
||
|
|
||
| @torch.no_grad() | ||
| def filter_wd_params(model: nn.Module): |
There was a problem hiding this comment.
add type for return value
| return no_decay, decay | ||
|
|
||
|
|
||
| def dict_to_tensor(data, device): |
|
|
||
|
|
||
| @torch.no_grad() | ||
| def vec_evaluate(vec_env, actor, num_episodes, seed=0, device="cpu"): |
|
|
||
|
|
||
| class Actor(nn.Module): | ||
| def __init__(self, action_dim, rnn_hidden_dim=512, rnn_layers=1, rnn_dropout=0.0, use_prev_action=True): |
|
|
||
| pbar.close() | ||
| result = { | ||
| "reward_median": np.median(episode_rewards), |
There was a problem hiding this comment.
let's rename to returns to be consistent
There was a problem hiding this comment.
return_median, return_mean, etc
There was a problem hiding this comment.
to late, we have all logs in wandb in this format....
There was a problem hiding this comment.
it is consistent across algorithms tho
| align: Optional[Alignment] = None, | ||
| **kwargs | ||
| ) -> nld.TtyrecDataset: | ||
| if not nld.db.exists(db_path): |
There was a problem hiding this comment.
this original solution from DD is actually a bit problematic
if the db was not properly initiialized for some reason (i.e., a wrong path and then fixed) this will silently re-use db
i think it's better to initialize the DB each time as it does not take much time
| CACHE_PATH = os.environ.get('KATAKOMBA_CACHE_DIR', os.path.expanduser('~/.katakomba/cache')) | ||
|
|
||
|
|
||
| def _flush_to_memmap(filename: str, array: np.ndarray): |
| gameid = self.gameids[idx] | ||
| return dict(self.hdf5_file[gameid].attrs) | ||
|
|
||
| def close(self): |
There was a problem hiding this comment.
let's add a flag for cleaning the memmap
sometimes people would like to work with just one dataset and rebuilding it every time is not desirable
| def close(self): | ||
| self.hdf5_file.close() | ||
| # remove memmap files from the disk upon closing | ||
| if self.mode == "memmap": |
There was a problem hiding this comment.
let's add logging that this is happening
No description provided.