Autoencoder Architechture about keras

[kenterao]

I have an implementation of Autoencoder with tied weights. It's basically two-layers-in-one. After pre-training you can extract the weights and initialize two Dense layers.

def get_output(self, train):
    X = self.get_input(train)
    encode = self.activation1(T.dot(X, self.W) + self.b)
    decode = self.activation2(T.dot(encode, self.W.T) + self.b_p)
    return decode

The trick for fitting multi-layer autoencoders is how to not backprop through the previously fit weights. The method I use is to create a regularizer function

def zero_grad(g, p):
    return 0.

Then I pass that into the pre-trained Dense layer to prevent backprop from changing the weights.

[jramapuram]

Implemented #180