This pretrained GAN system consists of a ResNet-style generator and discriminator. During training, stability is controlled by adapting the strength of gradient penalty regularization on-the-fly. The gradient penalty weighting term is contained in switch.npy. The generator is conditioned on a latent coordinate system to produce samples of arbitrary length. The latent representations produced by the generator are then passed to a decoder which converts them into waveform audio. The generator has a context window of about 12 seconds of audio.
This pretrained Misc GAN system is automatically downloaded at the first execution of the system. Try Musika here!
The Misc GAN system was trained on the SXSW music dataset (17000 songs with diverse genres).
- Downloads last month