A collection of custom objects to use with keras and tensorflow
Research goes faster if implementation doesn't suck up weeks
A Keras Model that collects updates from the loss it's compiled with and passe them to ops.
A Loss object based on BEGAN adaptative architecture by Berthelot et al [1]. Main point is having parameters that adapt to current performance. This is to be used with a discriminator autoencoder. Current implementation assumes that, during the training, even-numbered rows are issued from actual training set while odd-numbered rows are generated by the adversarial model.
no metrics yet
[1] BEGAN : https://arxiv.org/abs/1703.10717