Do you accept tf.keras.callbacks?

[beangoben]

Seeing as keras is going to become a first citizen inside of tensorflow, are you also accepting submission for callbacks? Are there guidelines for this (writing tests/docs)?.

[seanpmorgan]

My first thought is that it seems like it would be in scope as there is a pre-defined structure. I'll add it to the monthly meeting this Friday where we can discuss adding this as a new standardized API.

I'll update this issue after the meeting... but can you give some example callbacks you'd like to see added and/or would be willing to contribute?

[seanpmorgan]

Also, just a reminder... the monthly meeting is at 1pm PST (join our google group and you'll get calendar invites) and we encourage all to come!

[beangoben]

Would have to check, but potentially some callbacks would be:

• LearningRateScheduler compatible functions like SGDR (https://www.tensorflow.org/api_docs/python/tf/train/cosine_decay_restarts).
• In-memory best weights restore (like ModelCheckpoint but in-memory).
• Evaluation of metrics (r^22,auroc,f1,etc.) on external datasets (test set) each epoch, batch.
• Saving 'snapshots' of your embeddings during training.

[facaiy]

Do you mean to implement them as subclass of tf.keras.callbacks.Callback?

[beangoben]

Yes, with tests..but that all depends on the guidelines that will be recommended/set.

[seanpmorgan]

Short answer to this issue is Yes. See #61