search

allenai / deep_qa

A deep NLP library, based on Keras / tf, focused on question answering (but useful for other NLP too)
Apache License 2.0
404 stars 132 forks source link

Tensorboard logging #365

Closed DeNeutoy closed 7 years ago

DeNeutoy commented 7 years ago

This PR allows us to have tensorboard summaries from any tensor within a model by using:

tf.summary.scalar("name", scalar_tensor) tf.summary.histogram("name", tensor) etc

rather than the previous slightly restrictive inbuilt keras ones which only allowed loss summaries or summarising all weights in the network, which isn't very helpful. By default, it records the total loss from the model at each iteration. It's also a step in the right direction for having a completely separate training loop, eventually.

DeNeutoy commented 7 years ago
screen shot 2017-05-18 at 5 07 57 pm

This is what the default loss summary looks like.

At the moment, because we aren't using tensorflow to save weights, we don't have embedding visualisations, but that will be the next thing i'll try to switch over (possibly as part of getting mattp's multi-gpu model integrated) so then we'll be able to do the tsne plots in the browser and stuff.