Clean up any resources or hooks after decoding is finished

Convert the string with format "cuda:X" to integer X.

Decode locations from predictions using priors to undo the encoding we did for offset regression at train time. Args: loc (tensor):

Download a model from a URL and save it to the specified path.

Finalize search and return the final candidate sequences Parameters ---------- tokens : Tensor, shape = (n_audio, n_group

Returns a tensor of the video frames at the given index. Args: index: The index of the video frames to return.

Get all the stored features as NumPy Array. Returns: Concatenation of the stored features.

Get extension directory path, optionally with a subpath

Automatically detect the best config file for LatentSync version

Import a Python file as a module using its file path.

Initialize global temporary directory settings

The `MultiHeadAttention` module optionally accepts `kv_cache` which stores the key and value tensors calculated for the previous po

Open an audio file and read as mono waveform, resampling as necessary Parameters ---------- file: str The audio file t

Loads all the video frames under the dataroot and returns a list of all the video frames. Args: dataroot: The root d

Compute the log-Mel spectrogram of Parameters ---------- audio: Union[str, np.ndarray, torch.Tensor], shape = (*) The

load the mel filterbank matrix for projecting STFT into a Mel spectrogram. Allows decoupling librosa dependency; saved using: np

Apply non-maximum suppression at test time to avoid detecting too many overlapping bounding boxes for a given object. Args: boxes:

Courtesy of Ross Girshick [https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/nms/py_cpu_nms.py]

Compute number of time frames of spectrogram

Pad or trim the audio array to N_SAMPLES, as expected by the encoder.

Pre-download all required models and cache them properly.