DiffusionDVAE

This is a frozen, old version of one of Zach Evans’ diffusion-based autoencoders, for use with audio-algebra

See Zach Evans’ official audio-diffusion for updates.

Other parts of audio-algebra may use Scott Hawley’s “Python Packaging Fork” of this: https://github.com/drscotthawley/audio-diffusion. That fork is also “out of date” w.r.t. @zqevans’ research; we’ll sync back up someday! See “LICENSE(S)” at the bottom of the notebook.

source

get_alphas_sigmas

 get_alphas_sigmas (t)

Returns the scaling factors for the clean image (alpha) and for the noise (sigma), given a timestep.

source

get_crash_schedule

 get_crash_schedule (t)

source

alpha_sigma_to_t

 alpha_sigma_to_t (alpha, sigma)

Returns a timestep, given the scaling factors for the clean image and for the noise.

source

sample

 sample (model, x, steps, eta, logits)

Draws samples from a model given starting noise.

source

DiffusionDVAE

 DiffusionDVAE (global_args)

Hooks to be used in LightningModule.

Testing that:

args_dict = {'num_quantizers':0, 'sample_size': 65536, 'sample_rate':48000, 'latent_dim': 64, 'pqmf_bands':1, 'ema_decay':0.995, 'num_quantizers':0}
global_args = namedtuple("global_args", args_dict.keys())(*args_dict.values())
model = DiffusionDVAE(global_args=global_args)