Save a training checkpoint containing model weights, optimizer/scheduler states, and metadata. Behavior: - In Distributed Data Parallel

Compute Signal-to-Distortion Ratio (SDR) for one or more audio tracks. SDR is a measure of how well the predicted source (estimate) matches

Trains the model based on the provided arguments, including data preparation, optimizer setup, and loss calculation. The model is trained for

Validate a trained model on a set of audio mixtures and compute metrics. This function performs validation by separating audio sources fr

Initializes SCNet with input parameters.

Load specific chunk by track path and offset

BigShifts wrapper for inference-time demixing.

Close input and output audio streams. Args: stream_input (pyaudio.Stream): Input audio stream. stream_output (pyaudio.Stream

Compute and log the metrics for the current epoch, and save model weights if the metric improves. Args: all_losses: all_metr

Main function to copy mixture files from valid directory to inference directory. Parameters: ---------- args : Optional[argparse.Nam

Generates random audio tracks for stems in two subdirectories within the specified directory. Parameters: ---------- directory : Pat

Reverse normalization on an audio signal. Applies the stored mean and standard deviation to restore the original scale of a previously n

Finds all .wav files located in folders that do not contain specified categories in their folder names, within the given src_folder directory

Generate a descriptive name for a Weights & Biases (wandb) run. Combines the model type, a dash-joined list of training instruments, and

Returns an model class from a string (case-insensitive). Args: identifier (str): the model name. Returns: :class:`torch.nn.M

Collect validation mixture file paths from one or more root directories. Scans each directory in `args.valid_path` for files matching the

Initialize input and output audio streams. Args: chunk_size (int): Size of each audio chunk. Returns: Tuple[pyaudio.PyA

Initialize environment for Distributed Data Parallel (DDP) training/validation. Sets up the DDP process group, seeds random number generator

Load a possibly incompatible state dict into `model` with best-effort matching. Accepts either a raw state_dict or a checkpoint dict with we

Robust, quantile-masked MSE (“trimmed” MSE). Computes an elementwise MSE, optionally averages spatial dims (“coarse”), then masks out th

Normalize an audio signal using mean and standard deviation. Computes the mean and standard deviation from the mono mix of the input sig

Parse command-line arguments for training configuration. This function constructs an argument parser for model, dataset, training, and loggi

Parse command-line arguments for validation configuration. Builds the CLI for model selection, configuration paths, validation data loca

Build the training DataLoader. If torch.distributed.is_initialized() is True, construct a DDP DataLoader with DistributedSampler; otherwise,

Update accumulated metrics and (optionally) a tqdm progress bar. In non-DDP runs, appends each metric value to `all_metrics[metric_name][

Calculate the AuraMRSTFT metric, which evaluates the spectral difference between the reference and estimated audio signals using Multi-Resolu

Calculate the AuraSTFT metric, which evaluates the spectral difference between the reference and estimated audio signals using Short-Time Fou

Compute the L1 Frequency Metric between the reference and estimated audio signals. This metric compares the magnitude spectrograms of the re

Calculate the Log-WMSE (Logarithmic Weighted Mean Squared Error) between the reference, estimate, and mixture signals. This metric evaluates

Calculate L1-SNR metric (higher is better). Returns negative loss value for scheduler compatibility (mode='max').