Calculate the histogram for the whole video sequence Args: vectors (np.ndarray): (B, S, H, W, 2) cell_size (int): the height

Check if the configuration is compatible with VideoVAE. Currently, it is not compatible with many functionalities, due to complicated

cleanup the project by applying TTL policy to the model artifacts

cleanup the models that are not best or latest and have expired Returns: size of the deleted artifacts (in GB)

Count the histogram for the subcube Args: vector_cell (np.ndarray): (S, H, W, 2) angle_bins (int): the number of angle bins

Returns a video of shape [T, 3, H, W] showing the camera path being revealed over time. Each frame shows all cameras up to the current time s

Create a panorama rotation-only trajectory (camera rotating in place)

Cut the whole video sequence into subcubes Args: vectors (np.ndarray): (B, S, H, W, 2) cell_size (int): the height and width

Modify DiffuserImageVAE to be compatible with VAE abstract class

Compute the distance and angle between two sets of cameras, represented by their extrinsics.

Download dataset from the internet and build it in save_dir

r"""Download a file containing a pre-trained model from a URL. Supports automatic caching at torch.hub's cache directory. Adapted from `torch.hub.

Encode a batch of videos to a batch of DiagonalGaussianDistributions.

enforce zero terminal SNR following https://arxiv.org/abs/2305.08891 returns betas

Exclude videos that failed to preprocess

Exclude videos that are shorter than n_frames

Executing a certain task specified by string. Each task should be a stage of experiment. In most computer vision / nlp applications,

Extract image features from the videos. Args: videos: Videos of shape (B, T, C, H, W), uint8, range [0, 255]. Ret

Load pretrained model from path, with additional kwargs.

Generate a random 8-character alphanumeric string.

Generate a random 8-character alphanumeric string that does not exist in the project.

Args: embed_dim: Embedding dimension. pos: Position tensor of shape (...). Returns: Positional embeddings with shape

Get the corners of the image in normalized coordinates. Returns: np.ndarray: [4, 2] - [top_left, top_right, bottom_left,

Args: K: [B, T, 3, 3]; [fx, fy, cx, cy] R_rc: [B, T, 3, 3] T_rc: [B, T, 3] Returns: frustum: [B, T, 8, 3]

Get n-dimensional sinusoidal positional embeddings. Args: embed_dim: Embedding dimension. shape: Shape of the input tensor.

K: [..., 3, 3] or [..., 4] Camera intrinsic matrix. uv: [..., 2] Pixel coordinates. R_wc: [..., 3, 3] Rotation mat

This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. :param timesteps: a 1-D

:param observation_hat: predicted observation tensor of shape (frame, batch, channel, height, width) :param observation_gt: ground-truth obse

Check if a run has a committed model checkpoint.

Flow is represented as difference between two coordinate grids flow = coords1 - coords0

Convert camera-world (cw) to world-camera (wc) coordinates. Args: R: [..., 3, 3] Rotation matrix from camera to world frame

Check if a run exists.

frames: float32/float64 ndarray with shape (T, C, H, W) and values in [0, 1] returns: same shape & dtype, still in [0, 1], with frame

Load metadata from metadata_dir

A parallel implementation of multiple flow field warping comb_block: An nn.Seqential object. img shape: [b, c, h, w] flow shape: [b,

Normalizes the camera poses by the mean of all cameras, i.e. computes the relative poses w.r.t. the mean frame. The mean camera becom

Normalize the score (shape: (,)) with the min and max values.

Additional setup before preparing clips (e.g. excluding invalid videos)

Override the backbone with the specified name.

Compute pairwise squared distances for a batch of point sets. Args: points: (B, N, D) tensor of B point sets each with N points of d

Compute cumulative sizes for the dataset and update self.cumulative_sizes Shuffle the dataset with a fixed seed

Returns the rays for the camera poses. Args: resolution (int): The resolution of the image. Returns:

Register mean and std of data as tensor buffer. Args: mean: the mean of data. std: the std of data.

For each sequence in the batch, replaces the invalid camera poses (mask == True) by interpolating between the nearest valid c