Topologies
[training topology]
@beam/fedavgFederated averaging
Rounds of local training on every worker, a tree all-reduce merge of their models and a stop on the training loss.
Static membershipBuiltinActivedistributed run
[topology identity]
Details
What version 1.0.1 declares.
- Package
@beam/fedavg- Membership
- Static
- Engines
- distributed run
- Tags
- all-reduce, federated-averaging, rounds, synthetic
- Step kinds
- local, all-reduce
- Latest tag
- latest -> 1.0.1
- Publisher
- registry
- Visibility
- Public
- Checksum
sha256:51ad0cd46bb63589c0c9ce0a5bf975ae6afd4e4d3e355452b944534b573c75fd
[roles]
Roles
Who takes part, and how many members each role holds.
Every eligible room member trains and takes part in the merge.
[slots]
Slots
The actions a run binds: what each slot consumes and produces.
model?, neighbor? -> model, report
Initializes a worker's model (a training run without compute or failures).
update* -> model, report?
Averages the workers' models; associative, commutative and closed (runs as a tree).
model?, neighbor? -> update, report
Trains the received model on the worker's data; its report holds the training loss.
[round]
Step kinds
The steps of one round, in order.
- localrole workers · slot train
- all-reducerole workers · object update · slot merge