Write to TensorBoard every x samples. #11152
Conversation
…d` method to the TensorBoard class
…rboard-callback-modifications. Resolved conflicts and tested callbacks # Conflicts: # keras/callbacks.py # tests/keras/test_callbacks.py
…om/sameermanek/keras into batch_tensorboard
gabrieldemarmiesse
changed the title
gabrieldemarmiesse
changed the title
the metrics and losses.
time it wrote to the logs.
gabrieldemarmiesse
changed the title
There was a problem hiding this comment.
Seems like a useful feature, thanks for the PR. @omalleyt12 would you mind taking a look? Thank you.
keras/callbacks.py
Outdated
| @@ -715,6 +715,12 @@ class TensorBoard(Callback): | |||
| `embeddings_layer_names`. Numpy array (if the model has a single | |||
| input) or list of Numpy arrays (if the model has multiple inputs). | |||
| Learn [more about embeddings](https://www.tensorflow.org/programmers_guide/embedding) | |||
| write_step: `'batch'` or `'epoch'` or integer. When using `'batch'`, writes | |||
There was a problem hiding this comment.
There's already a embeddings_freq argument, and this new argument write_step is not consistent with it. Maybe write_freq or update_freq would be better. Thoughts?
There was a problem hiding this comment.
Naming things is as hard as ever.
I like update_freq since we "abstract" the writing to the file in this case. The user doesn't really care (take this with a grain of salt) what happens to the file. He/she wants to see the plot being updated. That's what matters.
|
|
|
ping @omalleyt12 |
|
LGTM, a useful feature and a nice improvement over the pr it is based on. The new param names + options are much better. |
|
Thanks. I don't think I'll make the PR in the Tensorflow repo because of lack of time. I'll let someone else do it if you don't mind. |
|
I just recently switched from Keras to tf.keras, and noticed that TensorBoard summaries looked a little bit different there: With tensorflow-gpu==1.11.0 and tensorboard==1.11.0, I no longer have data in val_loss, but in epoch_val_loss; and in addition to loss, there are epoch_loss and batch_loss. Is that a difference between the two Keras version? Will it impact porting of this feature here to tf.keras? |
Summary
Closes #7617 .
This PR is an update of #7617. Credit goes to @sameermanek for his work.
The goal of the PR is to be able to see the metrics and losses in a per-batch basis instead of a per-epoch basis.
Related Issues
#6692
PR Overview
This PR is backward compatible
A new argument has been added to the constructor of the TensorBoard callback. Here is the description that I added in the docs:
Feedback welcome. Especially from @fchollet since the API has been changed.