Reproducible results

[EvanDufraisse]

I'd like to thank you beforehand for sharing this implementation.

I've compiled the code and ran several tests using the train and test sample dataset in ./sample-data/images.tgz.
However, whatever value I set alpha and k, I always get a decrease in loglikelihood for 3-steps and a model that outputs a unique class value at inference time.

Would you have indications on the results one should expect in terms of accuracy over the testing set, and for which settings ?

Thanks for your help