Implementation of convolutional VAE in pytorch
MNIST dataset which consists of images of shape 1x28x28
Encoder-Decoder network consisting of conv and convtranspose layers respectively, activation function used is LeakyReLU instead of sigmoid as suggested in original paper to tackle vanishing gradient problem.
Learning Rate = 2*e-05
Batch Size = 16
Epochs = 50
Optimizer = Adam with betas-(0.5,0.999)
Reconstructed Images from original images:
Images constructed from sampled noise:
Loss Curve:-
