layers now specifies the number of hidden layers in the encoders. (seems to have previously specified the number of dense connections, which was 2 by default with 1 hidden layer in between)
Users can now specify a linear encoder by using an MLP or NGAM with <1 layers, which defaults to a single matrix of trainable weights with no activations.
cnellington
commented
2 years ago
layersnow specifies the number of hidden layers in the encoders. (seems to have previously specified the number of dense connections, which was 2 by default with 1 hidden layer in between)Users can now specify a linear encoder by using an MLP or NGAM with <1 layers, which defaults to a single matrix of trainable weights with no activations.