Ambiguity in KL Divergence Implementation

[amasawa]

I would like to extend my gratitude for your valuable work. I have a few inquiries and would appreciate your insights.

1. The paper refers to three types of losses (Equations 20, 21, and 22). However, these are not implemented in the code. The attached image illustrates that both the Mean Squared Error (MSE) and Negative Log-Likelihood (NLL) are traditional VAE reconstruction losses. Beta 1 and Beta 2 represent two types of Beta-divergence implementations. Therefore, there is no implementation for the Alpha and Gamma divergence. Could you provide some clarification on this?

2. I have been unable to replicate the results in the table from your paper (VQRAE in GD with beta-1 divergence). Could you share the hyperparameters for each of the datasets you used in your work? 4. There is a noticeable performance difference between the comparison method (OMI KDD19) and ICDE22[1] on the same dataset. Can you share the implementation details in OMI KDD19 If possible.

[1] Robust and Explainable Autoencoders for Unsupervised Time Series Outlier Detection.