distributed training #12
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 4 commits
July 19, 2020 01:50
Trying to see if this is the cause of serialization errors
added 25 commits
July 21, 2020 21:27
bugfix attempt
Much better. Less bigs/workarounds required with the newer version :)
added 18 commits
July 28, 2020 19:06
`rmprocs()` can hang beyond timeout.
…eFlux.jl into ks/big_train
this matters because BatchNorm and other similar layers need to maintain state that is not learned by gradient descent
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is the companion PR to a beacon-internal project.
It defines a
DistributedClassifierwith an impl oftrain!that sends batch specs (whatever info workers need to know what their next batch is) to workers, where losses and gradients get computed and sent back. The driver node performs parameter updates after summing the gradients from all workers. This is a purely synchronous distributed training loop: the workers are always running with the latest version of the model, which ensures that the model converges and performs the same as it would if it were trained locally (there are no such guarantees for asynchronous training schemes where workers are often running with stale model params.)It also defines a
DistributedLoggerallowing workers to send back logs to the driver node. This hasn't been tested, beyond it not barfing.It also defines some utilities in
distributed/that are not specific toLighthouseorFlux, and which should be moved somewhere else eventually.