This is the implementation of VAE for own research purpose
Unlike traditional auto encoders it used variational inference.
Model architecture
encoder 1784===>1400 1400===>120 (Bottle Neck layer)
decoder 120====>1400 1400===>1784
Please note the loss function consists of two terms 1) reconstruction loss 2) Regularization model is trying to minimize the both i.e it is minimiizing the KL divergence of the distribution coming out of the encoder and Multivarient Normal Distribution( Mean=0, in covariance matrix only diagonal element are 1 other are fixed to zero.), so we are assuming some dimension of latent variable(int this model 20) and trying the predict the parameters accordingly.
Result
