janzuiderveld / continuous-audio-representations

Official PyTorch implementation for "Towards Lightweight Controllable Audio Synthesis with Conditional Implicit Neural Representations".

Towards Lightweight Controllable Audio Synthesis with Conditional Implicit Neural Representations

Official PyTorch implementation for Towards Lightweight Controllable Audio Synthesis with Conditional Implicit Neural Representations. Accepted to "Deep Generative Models and Downstream Applications" (Oral) and "Machine Learning for Creativity and Design" (Poster) workshops at NeurIPS 2021.

See INR-collection for more details on the used implicit neural representation implementations.

Several (implicit) decoder architectures, ablations and latent embedding inference methods are implemented.

Quickstart

git clone https://github.com/janzuiderveld/continuous-audio-representations
cd continuous-audio-representations

pip3 install -r requirements.txt

Running Train.py automatically downloads datasets and sets them in place.

Usage:

The Following architectures are readily available (these can be trained by supplying these tags as --architecture argument for Train.py):

• wavegan (TCNN)
• im-net
• pi-gan (PCINR)
  • pi-gan_prog
  • pi-gan_sine_first
  • pi-gan_sine_last
  • pi-gan_relu
  • pi-gan_concat_middle
  • pi-gan_concat_all
  • pi-gan_min_mapping
  • pi-gan_five_mapping
  • pi-gan_shrinking
  • pi-gan_deep
  • pi-gan_wide (PCINR Wide)

Latent embedding inference methdods (--meta_architecture):

• autoencoder
• autodecoder

For datasets, the following are automatically downloaded when supplied as --dataset_name:

• SPEECHCOMMANDS
• NSYNTH.diverse_baseline
• NSYNTH.keyboard_baseline

Alternatively;

• "Path/to/your/folder/containing/.wav/files"

Extensive list of parameters:

long	default	help
--help		show this help message and exit
--wandb	0	Enable wandb logging
--wandb_project_name	default	Name of wandb project
--dataset_name	SPEECHCOMMANDS	Which dataset to train on.
--dataset_size	128	Number of samples to train on. Maximum is 1024 in given datasets
--audio_length	16000	Audio length
--autoconfig	0	Enable autoconfig. Overrides omega_0 values depending on dataset and architecture for tested setups.
--lr	1e-05	Learning rate
--batch_size	32	Batch_size
--num_epochs	5001	Number of epochs
--use_gpu	1	Enable GPU
--use_multi_gpu	0	Enable multiple GPUs
--architecture	pi-gan	What architecture to use as the decoder.
--meta_architecture	autodecoder	What latent embedding inference method to use.
--num_latent	256	Number of latent dimensions
--double	0	Enable double precision throughout training
--weight_norm	0	Enable weight norm
--first_omega_0	155	First layer input scaling for sinusoidal architectures
--hidden_omega_0	390	Hidden layer input scaling for sinusoidal architectures
--coord_multi	1	Input scaling for any architecture
--latent_init_std	0.001	Latent embedding initialization std
--latent_descent_steps	1	Number of gradient descent steps per iteration for latent embedding optimization
--latent_lr	0.3	Learning rate for latent optimization.
--samples_per_datapoint	8000	Number of samples per
--sample_even	1	Sample coordinates with equal spacing.
--per_sample	1	MSE per sample multiplier for objective function.
--deriv_per_sample	0	MSE per sample of derivative of functions multiplier for objective function.
--cdpam	1	CDPAM multiplier for objective function
--multiscale_STFT	1	Multi STFT multiplier for objective function
--weight_decay	1	L2 weight decay amount.
--eval_every	500	Evaluate every n iterations
--save_audio_plots	1	Save audio plots at every evaluation
--save_latents	1	Save latent embeddings
--save_audio	1	Save generated audio at end of training.
--save_model	1	Save model at end of training.
--eval_samples	1	Number of samples to evaluate on.
--eval_upscale_ratio	1	Upscale ratio for evaluation of generations.
--save_path	auto	Path to save output. 'auto' creates directories based on setup.
--max_high_res_batch_size	16	Maximum batch size for high resolution evaluations.
--note_general	default	Note to add to general output directory name
--note	default	Note to filter wandb results.
--num_groups	0	Number of groups to use for progressive activation scaling in pi-gan_prog
--prog_weight_decay_factor	0	Weight decay reduction factor for progressive weight decay.
--prog_weight_decay_every	0	Number of iterations after which to reduce weight decay.