Returns the 4x4 camera-to-world matrix for the camera poses.

Compute final MET3R score.

Decode latent tensor z into original space.

Download a pretrained model from the DFoT Hugging Face model hub. Set is_full to True to download the full model (including optimizer sta

Freeze the torch model

Check if a string is a run ID.

Load video from video_idx with given start_frame and end_frame (exclusive) if end_frame is None, load until the end of the video

Load video and conditions and bev from video_idx with given start_frame and end_frame (exclusive)

Run the algorithm.

Warmup optimizer learning rate.

Compute image centers by extending along optical axis Args: poses: Camera poses Returns: centers: I

Evaluate the denoising performance during training.

Extract I3D features. Requires a batch of videos of shape (B, T, C, H, W) and range [0, 1].

Load the pretrained VAE model.

Process the frequency ranges, replacing ALL with (0.0, 1.0), and following rules #1, #2, #3.

Helper function to run the VAE, either for encoding or decoding. - Requires shape to be a permutation of b, t, c, h, w. - Res

The unified sampling method, with length up to maximum token size. context of length can be provided along with a mask to achieve con

Generate a tensor of the given size, where each element is sampled from a Bernoulli distribution with probability `p`.

Wrapper for grid_sample, uses pixel coordinates

History guidance scheme equivalent to: typical conditional sampling, i.e. sampling with only a conditional score (on the enti

Encode input tensor x into latent distribution.

Create a cam2world trajectory depicting an orbital motion - camera moves in a circular path in the xz plane while always looking at the cente

Load video and conditions from video_idx with given start_frame and end_frame (exclusive)

Compute the mean absolute error (MAE) between videos1 and videos2. `score = (255 - MAE) / 255`. Args: videos1: Videos of shape (

Normalizes the camera poses by the first camera, i.e. computes the relative poses w.r.t. the first camera. After normalization, the f

Parse load into run_id and download option. (for load=xxxxxxxx in configurations) - If load_id is a run_id, return the run_id and None.

A history segment that contains a constant frequency range for all tokens.

Rescale and crop the video to the specified resolution. Used for preprocessing. Args: video (torch.Tensor): video tensor of shape (t,

Safely save a torch object to disk. If the path does not exist, create folders as needed.

Scales the camera locations, so that they are within the boundary box [-bounds, bounds]^3. Each scene is scaled independently, while

Returns the rays represented as a tensor. Args: use_plucker (bool): Whether to use Plücker coordinates or not. Re

Args: x: patchified tensor of shape (B, num_patches, patch_size**2 * C) Returns: unpatchified tensor of shape

check that input is a valid image to the network.

Constructor Args: cfg: configuration file that contains everything about the experiment logger: a pytorch-li