# Variational Autoencoder (VAE) Variant

Experimented with following VAE, methods are explained in the notebook prior to modelling

• Basic (vanilaa) VAE
• Beta VAE
• Dirichlet VAE
• Vector Quantized VAE

From experimentation in basic_vae, we have decided upon these few:

1. num_epochs = 30, 2. learning_rate = 25e-4, 3.latent_dim = 20 and 4. batch_size = 128

Results

VAE Variant	Parameters	Total Loss	Reconstruction Loss	KL-D Loss	Training Time
General VAE	310, 504	97.72	73.21	24.51	20 min 26s
Beta VAE	236, 740	131.97	94.51	12.49	16 min 49s
Dirirchlet VAE	284, 405	0.92	0.77	0.15	21 min 20s
VQ-VAE	299,985	0.9079	0.0012	VQ-Loss: 0.9067	61 min 44s
VQ-VAE ( celebA)	303,187	0.0104	0.0079	VQ_Loss: 0.0025	80 min 41s (GPU)