Takes a (batch, chunk_size, action_dim) sequence of actions, update the temporal ensemble for all time steps, and pop/return the next

Update EMA parameters in place with ema_param <- alpha * ema_param + (1 - alpha) * param.

Write a csv file containg timeseries data of an episode (e.g. state and action). This file will be loaded by Dygraph javascript to plot data in re

Args: log_dir: The directory to save all logs and training outputs to. job_name: The WandB job name.

Args: shapes (dict): A dictionary where keys are input modalities (e.g. "observation.image") and values are their sha

Args: cfg: Policy configuration class instance or None, in which case the default instantiation of the configura

Args: config: Policy configuration class instance or None, in which case the default instantiation of the configu

Args: input_shape (list): (C, H, W) input feature map shape. num_kp (int): number of keypoints in output. If None, ou

Args: config: Policy configuration class instance or None, in which case the default instantiation of the configu

Args: input_shape (list): (C, H, W) input feature map shape. num_kp (int): number of keypoints in output. If None, ou

VQBeTHead takes output of GPT layers, and pass the feature through bin prediction head (`self.map_to_cbet_preds_bin`), and offset prediction

VQ-VAE is composed of three parts: encoder, vq_layer, and decoder. Encoder and decoder are MLPs consisting of an input, output layer,

Args: config: Policy configuration class instance or None, in which case the default instantiation of the configu

Creates encoders for pixel and/or state modalities. TODO(alexander-soare): The original work allows for multiple images by concatenat

Args: config: Policy configuration class instance or None, in which case the default instantiation of the con

Temporal ensembling as described in Algorithm 2 of https://arxiv.org/abs/2304.13705. The weights are calculated as wᵢ = exp(-temporal_ensembl

Convenience module for running multiple decoder layers followed by normalization.

Args: dimension: The desired dimension of the embeddings.

Sampler that optionally incorporates episode boundary information. Args: episode_data_index: Dictionary with keys 'from' and 'to'

The online buffer can be provided from scratch or you can load an existing online buffer by passing a `write_dir` associated with an

Dummy constructor. Use copy_from* and create_from* class methods instead.

Return JPEG XL image from numpy array. Float must be in nominal range 0..1. Currently L, LA, RGB, RGBA images are supported

Input validation (not exhaustive).

Input validation (not exhaustive).

Input validation (not exhaustive).

Input validation (not exhaustive).

Return numpy array as contiguous view of bytes if possible.