josephsdavid / N2D

Library implementation of https://arxiv.org/abs/1908.05968v5
https://n2d.readthedocs.io/en/latest/
MIT License
23 stars 6 forks source link

random seeds for auto encoder #9

Closed stellaywu closed 4 years ago

stellaywu commented 4 years ago

Hey @josephsdavid! Another question - Is there a random seeds setting for the auto-encoder? Trying to have reproducible results Thanks!

Also would it be possible to add automatic choice of number of clusters for non dbscan based clustering methods?

Stella

josephsdavid commented 4 years ago

https://www.tensorflow.org/api_docs/python/tf/random/set_seed for now, you can do it through tensorflow! Will add a more convenient way to do it to the next update! Yeah let me think how automatic clustering would work, longer term project for that :)

josephsdavid commented 4 years ago

Once trained, the autoencoders should be deterministic, but since they are optimized through SGD without setting a seed results of training might be slightly different. Ideally we would want to initialize the N2D object with a seed that is constant throughout all steps of the process

stellaywu commented 4 years ago

thanks!