Closed rizar closed 7 years ago
Hmm. As for me, it is better to crash sooner (if something is wrong with evaluation) rather than later.
You can then trigger DataStreamMonitoring on before_training. This way you will first do the monitoring and only then compile the training algorithm.
On Thu, 30 Mar 2017 at 11:12 dmitriy-serdyuk notifications@github.com wrote:
Hmm. As for me, it is better to crash sooner (if something is wrong with evaluation) rather than later.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mila-udem/blocks/pull/1185#issuecomment-290442233, or mute the thread https://github.com/notifications/unsubscribe-auth/AAn8YrhA6Gi83gorGUOc0LOMKIw9NFb3ks5rq8ZngaJpZM4Mucvr .
Yes, but evaluation might be very slow. It is always frustrating when model compiles for several minutes, then evaluates for half an hour and then crashes with a first batch of training. And if you interrupt it during evaluation you'll have to wait again for compilation.
I would explicitly compile the evaluator on before_training
callback.
Let's discuss this offline.
On Thu, 30 Mar 2017 at 11:24 dmitriy-serdyuk notifications@github.com wrote:
Yes, but evaluation might be very slow. It is always frustrating when model compiles for several minutes, then evaluates for half an hour and then crashes with a first batch of training. And if you interrupt it during evaluation you'll have to wait again for compilation.
I would explicitly compile the evaluator on before_training callback.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mila-udem/blocks/pull/1185#issuecomment-290446089, or mute the thread https://github.com/notifications/unsubscribe-auth/AAn8YgGDofc4RnlinPR5vSan3GGNe86dks5rq8lAgaJpZM4Mucvr .
LGTM then
For debugging it's much better if the compilation in
DatasetEvaluator
happens when it's first called, not when it's created.