Check that all keys are strings that don't contain '.'

Directly copy from https://github.com/MishaLaskin/curl/blob/master/curl_sac.py Uses logits trick for CURL: - compute (B,B) ma

Create a semi-unique experiment name that has a timestamp :param exp_prefix: :param exp_id: :return:

Creates and returns a unique log directory. :param exp_prefix: All experiments with this prefix will have log directories be under this

Convert each value in the dictionary into a JSON'able primitive. :param d: :return:

Evaluate the policy, e.g. save/print progress. :param epoch: :return:

Assumes a potentially nested dictionary where all keys are strings that do not contain a '.' Returns a flat dict with keys having format

Save things that shouldn't be saved every snapshot but rather overwritten every time. :param epoch: :return:

Should be implemented on a per algorithm basis taking into consideration the particular get_epoch_snapshot implementation for

:return: # of unique items that can be sampled.

Preprocessing image, see https://arxiv.org/abs/1807.03039.

Do anything before the main training phase.

inputs np array outputs tensor

random covolution in "network randomization" (imbs): B x (C x stack) x H x W, note: imgs should be normalized and torch tensor

args: imgs: np.array shape (B,C,H,W) out: output size (e.g. 84) returns np.array

args: imgs: np.array shape (B,C,H,W) min / max cut: int, min / max size of cutout returns np.array

args: imgs: shape (B,C,H,W) out: output size (e.g. 84)

args: imgs: torch.tensor shape (B,C,H,W) device: cpu or gpu, p: prob of applying aug, returns torch.tensor

args: imgs: torch.tensor shape (B,C,H,W) device: cpu or cuda returns torch.tensor

args: imgs: torch.tensor shape (B,C,H,W) device: str, cpu or gpu, p: float, prob of applying aug, returns torch.tensor

Return generator of regularizable parameters. Right now, all non-flat vectors are assumed to be regularizabled, presumably because on