Generate a windowing array with a linear fade-in at the beginning and a fade-out at the end. This function creates a window of size `window_

spec: complex (B, C, F, TT) -> audio: (B, C, T)

Precompute all chunks for dataset_type 5 with overlap 2 using multiprocessing

Precompute exact number of good chunks

x: (B, C, T) -> spec: complex (B, C, F, TT)

Asserts that the shapes of input features are equal.

:param pred_track: shape = (num, channels, length) :param weights: shape = (num, ) :param algorithm: One of avg_wave, median_wave, min_wa

Replaces specific layers in the model with LoRA-extended versions. Parameters: ---------- config : Dict[str, Any] Configurat

Calculate the 'bleed' and 'fullness' metrics between a reference and an estimated audio signal. The 'bleed' metric measures how much the est

Split a dataset of `num_tracks` items into `world_size` equal contiguous blocks and return the half-open interval [start, end) assigned to

Build a composite loss from CLI/config options. Returns a callable that sums enabled terms (with per-term coefficients): - `masked_loss`

Finalize and clean up a Distributed Data Parallel (DDP) process group. Calls `torch.distributed.destroy_process_group()` to release resource

Removes all cached test tracks. This function deletes the entire directory specified by the global `TEST_DIR` variable if it exists.

Compute the global compression ratio. Args: - subband_shapes (List[List[int]]): List of subband shapes. Returns: - float: Globa

Compute the shapes for the subband layers. Args: - input_shape (int): Input shape. - bandsplit_ratios (List[float]): Ratios for spli

Converts MoisesDB dataset to MUSDB18 format for a specified number of folders. Parameters: - src_root (str): Root directory of the Moises

Parallelly counts the number of subfolders in each folder inside src_folder. Parameters: - src_folder (str): Root folder containing subf

Create a test configuration file based on an existing configuration. Parameters: ---------- original_config_path : str Path

Generate and save separate images for spectrograms and waveforms for both estimated and original audio. Creates two separate images:

Generate and save a spectrogram image from an audio waveform. Converts the provided waveform into a mono signal, computes its Short-Time

Fake initialization function to simulate processing and model loading. Args: config (Any): Configuration used for demixi

Formats a filename from a template. e.g "{file_name}/{instr}" Using slashes ('/') in template will result in directories being created Re

Compute the output shape of a convolutional layer. Args: - input_shape (int): Input shape. - kernel_size (int, optional): Kernel siz

Compute the output padding for a convolution transpose operation. Args: - input_shape (int): Input shape. - output_shape (int): Desi

Detect available instrument files and their extensions.

Calculate a list of metrics to evaluate the performance of audio source separation models. The function computes the specified metrics based

Should return args to re-instantiate the class.

Create and configure an optimizer for training. Selects the optimizer type based on `config.training.optimizer` and applies the correspo

In case the state dict needs to be modified before sharing the model.