Load video list and prompt information based on a specified dimension from a JSON file. Parameters: - json_dir (str): The directory

Returns a `logging.Logger` for `name` that can handle multiprocessing.

Sample a float from a provided range description.

q: [B, Lq, Nq, C1]. k: [B, Lk, Nk, C1]. v: [B, Lk, Nk, C2]. Nq must be divisible by Nk. q_lens

Load joints from either .npy or .pt file format. Args: evaluation_file: Path to the evaluation file (.npy or .pt) device

Convert 276-dim global (aligned) motion representation back to SMPL params + joints. The representation layout follows collect_motion_rep_DAR

Return mean/std/count summary for a list of scalar values.

Compute focal length based on FOV.

Delegates logger call after checking if we should log. Accepts a new kwarg of `main_process_only`, which will dictate whether it will be logg

q: [B, Lq, Nq, C1]. k: [B, Lk, Nk, C1]. v: [B, Lk, Nk, C2]. Nq must be divisible by Nk. q_lens

Get mapping of dimensions to their categories Returns: mapping category name

Load data from JSON file Args: filepath: Path to the JSON file Returns: Loaded data

process original height width, sample a (height, width) in bucket_config with the closed aspect ratio.

Convert 6d rotation representation to 3d vector of axis-angle rotation. Shape: - Input: :Torch:`(N, 6)` - Output: :Torch:`(N, 3)`

Convert 6d rotation representation to 3x3 rotation matrix. Shape: - Input: :Torch:`(N, 6)` - Output: :Torch:`(N, 3, 3)`

Save data to JSON file Args: data: Data to save filepath: Path to save the JSON file

Locate an already rendered video for a given category/id.

Return the directory that stores converted assets (same as evaluation path).

Update SMPL parameters based on the rotation matrix R. Args: smpl_params (dict): Dictionary containing SMPL parameters ('global_orie

Convert 3d vector of axis-angle rotation to 3x3 rotation matrix. Shape: - Input: :Torch:`(N, 3)` - Output: :Torch:`(N, 3, 3)`

Convert 3d vector of axis-angle rotation to 4d quaternion. Shape: - Input: :Torch:`(..., 3)` - Output: :Torch:`(..., 4)`

Convert 3d vector of axis-angle rotation to 6d rotation representation. Shape: - Input: :Torch:`(N, 3)` - Output: :Torch:`(N, 6)`

Call Gemini API with video and text prompt. Args: video_bytes: Video file content as bytes prompt: Text prompt for the m

Run evaluation on motion data by first rendering to videos and then evaluating Args: evaluation_path: Path conta

Detect contact from foot velocities and height.

Load pose data from evaluation file. Args: evaluation_file: Path to the evaluation file (.npy or .pt) device: Target dev

Convert 3x3 rotation matrix to 3d vector of axis-angle rotation. Shape: - Input: :Torch:`(N, 3, 3)` - Output: :Torch:`(N, 3)`

Visualize motion by rendering a depth video. This mirrors depth_render.motion_checker.motion_vis with a narrowed dependency surface.

Remove global translation by subtracting the root joint's position.

Render depth maps for all frames in batches.

different from DistributedSampler, you don't need to call this function at the start of each epoch, because the return Iterator of __i

Return mean/std/count summary for scalar values.

Normalize depth map to uint8 images (white is closer).

Resolve SMPL/SMPLX family model path with sensible defaults. Env overrides: - SMPLX_MODEL_PATH for smplx - SMPLH_MODEL_PATH for smplh

Locate the raw motion file for a given category/id.

Load motion data from a supported file into a numpy array.

Return ordered candidate motion file paths to probe.

Check if log should be performed.