Github user avulanov commented on the pull request:
https://github.com/apache/spark/pull/1290#issuecomment-82484515
@debasish83 Sounds interesting!
1. Gradient calculation is distributed by means of data, i.e. each worker
calculates delta gradient on its part of the data. Although, it would be
interesting to distribute the model for the cases when it does not fit into the
workers memory.
2. If you have a different solver you can try using it for artificial
neural networks by replacing the current one (which is LBFGS)
3. If you wish to select features, do you plan to use L1 on the bottom
layer? In this case you might need to add it in the code by yourself.
This branch is relatively stable, I've done a lot of testing (see earlier
posts) so you can experiment with it. There is a separate branch with stacked
autoencoder (based on the code from this branch), which I've also tested with a
lot: https://github.com/avulanov/spark/tree/autoencoder
The main goal of the PR you mentioned is DBN and RBM. Our plan with @witgo
and @mengxr is to develop a more generic and scalable interface for artificial
neural networks and the port existing code.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
