Return an activation function given a string

img: [C,H,W] float, assumed in [0,1] (or at least bounded) sigma: std of Gaussian noise in same scale as img

Run inference on a pre-loaded numpy array (T, H, W, 3). Returns (ranges, intra_labels, inter_labels).

Inputs: [N, H, W, 3] in range [0, 255] Returns: float tensor [N, C, H, W] in [0, 1]

segments: list of (start, end, intra_label, inter_label) in global frame indices, [start, end) returns: ab: list of (start2, end2) in

Save list[dict] as JSON. Convert pred_intra_labels / pred_inter_labels from int to string.

This function disables printing when not in master process

Returns normalized float tensor [T, C, H, W]

Initializes the model. Parameters: backbone: torch module of the backbone to be used. See backbone.py transformer: to

The forward expects a NestedTensor, which consists of: - samples.tensor: batched images, of shape [batch_size x 3 x H x W]

Run shot cut detection on a video. Args: video: str file path | np.ndarray (T,H,W,3) uint8 RGB | torch.Tensor (T,H,W,3)

Load model weights and return an OmniShotCutModel instance. checkpoint_path can be: - a local file path → load directly (filename ignored)