Call a method, or get a property, from each sub-environment. Parameters ---------- name : string Name of the meth

r"""Clean up the extra resources e.g. beyond what's in this base class.

The drift function of the reverse-time SDE.

Generate a sinusoidal embedding for a scalar value.

Convert Euler Angles to Quaternions. See rotation.py for notes

evaluate onestep policy or the multistep base policy.

Extract metrics from an online W&B run using wandb.Api.

Extract metrics from W&B run's .wandb file using wandb.Api.

Calls the MLP to compute the mean, scale, and logits of the GMM. Returns the torch.Distribution object.

Generate training targets for the velocity field. Args: x1: Real data tensor of shape (batch_size, horizon_steps, action_dim).

Calculating the logprobs of the entire chain of denoised actions. Args: cond: dict with key state/rgb; more rece

for training and validation on fixed dataset

get num_ind random indices from a set of size sz (used for getting critic targets)

get num_ind random indices from a set of size sz (used for getting critic targets)

Compute the behavior cloning flow loss by comparing predicted and target flow values. This is the same as doing pre-training. Args:

Convert Rotation Matrix to Euler Angles. See rotation.py for notes

Resets the environment.

Resets the environment.

simply return action via integration (Euler's method). the initial noise could be specified. when `save_chains` is True, also return

B: batchsize outputs: xt: torch.Tensor of shape `[batchsize, self.horizon_steps, self.action_dim]` log_prob:

for resume purpose

saves model and ema to disk;

Use [:, None, None] to add two dimensions (horizon and transition)

Apply padding

Gives a list of TD estimates for a given list of samples from an RL environment. The TD(λ) estimator is used for this computation. :

Construct an SDE. Args: N: number of discretization time steps.

Initialize the FQL model with behavior cloning flow, actor, critic, and target critic. Args: bc_flow (ReFlow): Behavior cloning f

Adds the MIT License to the specified Python file if it doesn't already contain it. Preserves the existing content of the file.

Calculate the average episode reward from the rewards array and done indices.

Call instance methods of vectorized environments.

Call instance methods of vectorized environments with args.

Compare two checkpoint dictionaries for equality. rtol: relative tolerance for floating-point comparison atol: absolute tolerance for flo

cosine schedule as proposed in https://openreview.net/forum?id=-NEXDKk8gZ

Crop a PDF by specified percentages from the left, right, top, and bottom of each page. Parameters: - input_pdf_path (str): Path to