brainjs / brain-cluster

Brain + Distributed Deep Learning

brain-cluster

Getting brain to train in parallel (on many machines or on many threads).

Getting Started

This uses klyng which is built on fibers.

Get klyng installed globally:

yarn global add klyng

klyng by default runs a beacon on port :2222. You can change the port and add a password in the config file. If you ran the above command it is found here by default ~/.config/yarn/global/node_modules/klyng/config.json.

There is also the machine.json which describes how communication on the cluster will perform.

Scripts

yarn down

Shuts down the klyng beacon. You shouldn't have to do this unless you are running a bunch of klyng commands at once and it things start breaking and it isn't handled nicely

yarn up

Starts the klyng beacon. If running a klyng script and the beacon isn't up you shouldn't need to do that. This might need to be done manually when running on secondary machines.

yarn start

Start cluster.js Currently configured in the packages.json to start 6 local processses. Currently hard coded to train a network in solving the likely problem. As this gets flushed out there will be better documentation and integration with brain.js directly.

yarn cluster

will start cluster.js but will also apply the machines.json This is just in reference to how to run on a cluster, but hasn't been flushed out yet (also the machines.json doesn't mean anything at this point)

yarn example

starts a trivial example, used for concept testing on klyng. This will get removed eventually but is still useful in testing out specific functionality

yarn test-sizes

Like example this is used for testing specifically how large of a data set we can send back and forth from klyng instances. The documentation on klyng isn't stellar so this was a semi-stress test. This will also be removed as brain-cluster matures

brain-cluster's Issues

Crap Don't work

So I thought that if we added the weights it would let us combine multiple networks into one.
It isn't working right now either:

1. I am dead wrong.
2. I am adding the weights wrong.
   Here is where I need you to double check.