Use this method to setup (i.e. load & initialize) the model. Can improve the runtime of the benchmark. Note that this method is option

Abstraction from os.path.join() Builds absolute paths from relative path strings with SIM_FRAMEWORK/ as root. If args already contains an

Strips a string to the sequence first characters after the `sep_str`. Args: s (str): The string to strip. sep_str (str, optional)

Computes the spatial softmax of a convolutional feature map. Read more here: "Learning visual feature spaces for robotic mani

Computes the spatial softmax of a convolutional feature map. Read more here: "Learning visual feature spaces for robotic mani

Get the forward and backward pass function for the mLSTM kernels with chunkwise formulation. Args: autocast_kernel_dtype: The dtype

Get the forward and backward pass function for the mLSTM kernels with chunkwise formulation. Args: autocast_kernel_dtype: The dtype

Returns the kernel function to benchmark. The inputs to the kernel function are the tensors returned by `_get_input_tensors`.

Get the forward and backward pass function for the mLSTM parallel formulation. Args: autocast_kernel_dtype: The dtype to use for the

Get the forward and backward pass function for the mLSTM parallel formulation. Args: autocast_kernel_dtype: The dtype to use for the

Generate a forward and backward pass function for the mLSTM kernels with chunkwise formulation. Args: autocast_kernel_dtype: The dty

Generate a forward and backward pass function for the mLSTM kernels with chunkwise formulation. Args: autocast_kernel_dtype: The dty

This is the core mLSTM operation in parallel form computed in tiles. This version is stabilized. We control the range of exp() arguments by en

Generate a forward and backward pass function for the mLSTM parallel formulation. Args: autocast_kernel_dtype: The dtype to use for

Generate a forward and backward pass function for the mLSTM parallel formulation. Args: autocast_kernel_dtype: The dtype to use for

Computes the training loss of the diffusion model given a batch of data. At every training sample of the batch we generate noisy samp

Predicts the denoised x_{t-1} sample given the current diffusion model :param x: noisy input action :param t: batch of tim

Main loop for generating samples using the learned model and the inverse diffusion step :param state: the current state observation

Converts model output into prediction of x₀. If predict_epsilon=True: - model outputs predicted noise (ϵ) - w

Computes the posterior mean and variance of the diffusion step at timestep t q( x_{t-1} | x_t, x_0)

Main Method to sample the forward diffusion start with random noise and get the required values for the desired noisy sample at q(x_{

Return a (size, latent_size) latent sample, corresponding to a traversal of a latent variable indicated by idx. Parameters --

Ensure that array or tensor @x has size @size in dim @dim. Args: x (np.ndarray or torch.Tensor): input array or tensor size

Compute the required embeddings for the visual ones and the latent goal.

Modifies the state dict in place.

Modifies the state dict in place.

Convert applicable model parameters to fp16

Copy current parameters into given collection of parameters. Args: parameters: Iterable of `torch.nn.Parameter`; the parame

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