models

Autoregressive Pretrained Models

Autoregressive library comes with a set of pre-trained models. Neither of models was carefully tuned. In the following

• R refers to the models' receptive field,
• Q to the number of input/output quantization channels.

Model mnist_q256

WaveNet trained on MNIST. R=699, Q=256, conditioned on the digit class.

Model mnist_q2

WaveNet trained on binarized MNIST. R=699, Q=2, conditioned on the digit class.

Model fseries_q127

WaveNet trained on random Fourier series dataset. R=699, Q=127, conditioned on the periodicity of the signal. Training parameters allow periods (integer) between 5-10secs which maps to conditions [0..4].

For usage see main Readme.