Returns a list of full S3 paths to files in a given S3 bucket and directory path.

Get codebook coordinates in the layout that corresponds to the specified timestep t and optionally to the codebook q. Coordinates are returned

Return a window function. Args: win_type (str): Window type. Can either be one of the window function provided in PyTorch ['h

Torch version of int16_to_float32 for GPU tensors.

Reconstruct the original signal from the frequency bands.

Perform RF-Inversion on audio to obtain the latent noise. Technique based on SEMANTIC IMAGE INVERSION AND EDITING USING RECTIFIED STOCHASTIC

very fast `glob` alternative. from https://stackoverflow.com/a/59803793/4259243

Load LoRA checkpoints from disk and attach them to `model`. - Uses a two-pass approach: first resolves each LoRA's adapter type, then loads

Load OpenL3 model for audio embeddings. Args: device: Device to load model on content_type: 'music' or 'env' (environment so

Load PANNs CNN14 model.

Load VGGish model via torchaudio or torchvggish. Returns (model, input_processor, backend_type).

calculates data array for mel spectrogram (in however many channels)

Compute the MMD loss between two flat sets of vectors. Args: x: Tensor of shape (B, D) y: Tensor of shape (B, D) band

Pads the signal to make its length divisible by the given number of bands. Parameters ---------- signal : torch.Tensor The i

Pads the input tensor 'x' on both sides such that its last dimension becomes the nearest larger power of two. Parameters: -----

returns a 3D point cloud of the tokens

Applies the polyphase method to efficiently analyze the signal using a filter bank. Parameters: ----------- signal : torch.Tensor

Polyphase Inverse: Apply polyphase filter bank synthesis to reconstruct a signal. Parameters ---------- signal : torch.Tensor

Precompute varlen attention metadata once to avoid recomputation in every attention layer. Args: padding_mask: Boolean tensor of sha

Rearrange signal into Batch x Channels x Length. Parameters ---------- signal : torch.Tensor or numpy.ndarray The input

Preprocess a [list] of audio (Channels x Length) into a batch tensor to be compatable with the encoder. The audio in that list can b

Remap LoRA state dict keys from parametrization index 0 to target_index. Checkpoints always save keys with index 0 (e.g. *.parametrizations.weigh

Removes silence longer than max_silence_duration and replaces it with a short silence. :param audio: torch tensor of shape [1, T] :param

Revert model logits obtained on a sequence built from the pattern back to a tensor matching the original sequence. This method is sim

Revert a sequence built from the pattern back to the original multi-codebook sequence without interleaving. The sequence is reverted using up

Draws samples from a model given starting noise. Euler method Args: sigmas: Pre-computed schedule tensor. Shape (steps+1,) for global sch

Draws samples from a model given starting noise. DPM-Solver++ for RF models Args: sigmas: Pre-computed schedule tensor. Shape (steps+1,)

Draws samples from a model given starting noise. Ping-pong sampling for distilled models Args: sigmas: Pre-computed schedule tensor. Shap

Draws samples from a model given starting noise. 4th-order Runge-Kutta Args: sigmas: Pre-computed schedule tensor of shape (steps+1,).

Sample timesteps for diffusion training by sampling logSNR values and converting to t. Args: batch_size (int): Number of timesteps t

Sample timesteps for diffusion training by sampling logSNR values and converting to t. Args: batch_size (int): Number of timesteps t

Sample next token from top K values along the last dimension of the input probs tensor. Args: probs (torch.Tensor): Input probabilities w

Sample next token from top P probabilities along the last dimension of the input probs tensor. Args: probs (torch.Tensor): Input probabil

Modified from PyTorch tutorial https://pytorch.org/tutorials/beginner/audio_feature_extractions_tutorial.html

Broadcast non-FSDP trainable parameters from rank 0 to all other ranks. Call this after optimizer.step() to ensure parameter synchronization

Test and visualize the noise spectrum

Tight, one-sided, *demodulated* complex STFT for real signals, suitable for MRSTFT losses. Args ---- x : [B, T] real tensor n_ff

Detect if state_dict is from a wrapped training checkpoint and unwrap it. Wrapped checkpoints have keys like 'diffusion.model.xxx' or 'd

Wrap model components for FSDP distributed training. Uses selective wrapping: only transformer layers are wrapped with FSDP for

Zero pad the shorter of two tensors to the length of the longer one.

signal: torch.Tensor [channels, time]