Open JasOlean opened 5 years ago
This paper has an excellent overview of what the beta
parameter is doing: https://arxiv.org/abs/1804.03599
To summarize, larger beta
will result in a more disentangled latent representation but lower-fidelity reconstructions. Smaller beta
will not impose disentangling as much, allowing for higher-fidelity reconstructions. At beta = 1
, the B-VAE is equivalent to a plain VAE, so it should is usually set to a value greater than one.
Determining the proper beta
depends on the problem and your goals. You can try several values for beta with your data, and you can create a custom training regimen that changes beta over time. This implementation assumes a constant beta
, but you can rebuild the model with a different beta during training.
In your case, you use beta = 100. So, how to choose proper beta value (not constant)? And large or small beta value is good or not?