Remove LearnLogger?

[ericphanson]

After #60 we've defined an interface that folks can plug into for custom logging.

We could drop the Tensorboard dep by removing the LearnLogger. We could make a no-op logger that doesn't actually log anything, or maybe a Dict logger that logs things to a logged dict as the current one does, which could be good for testing. Or we could keep the current LearnLogger.

[hannahilea]

maybe a Dict logger that logs things to a logged dict as the current one does, which could be good for testing

I vote this one!

And then either drop [tb] LearnLogger entirely, or perhaps add it to the docs or tests as a (tested?) example? It feels bad to have to keep supporting it, though, since as far as we know no one uses it. Maybe dropping it is fine, esp when our current internal logger is public (soon?) and we can point at it as an example of how to implement the interface.

[ericphanson]

One other option could be to have LearnLogger wrap other loggers. So it would be like

struct LearnLogger{L}
    logger::L
    logged::Dict{String,Vector{Any}}
end

and it could log everything to logged and then forward the call onto logger... So like

function log_value!(logger::LearnLogger, field, value)
    values = get!(() -> Any[], logger.logged, field)
    push!(values, value)
    return log_value!(logger.logger, field, value)
end

Then a LearnLogger is just something that wraps a logger to store the info to a Dict as well. Not really sure if we would actually use that or not though. I think the Dict-based thing is kinda problematic for huge training jobs because you need to store everything in memory.

[hannahilea]

I think the Dict-based thing is kinda problematic for huge training jobs because you need to store everything in memory.

Yeah---and I think if someone does want to log to a local dict, they can easily make a logger that does that.