Keras: callback_tensorboard error when embeddings_freq is not zero

Created on 21 Jun 2018  路  8Comments  路  Source: rstudio/keras

I'm reading Deep Learning with R, in Chapter 7, there's an example of tensorboard callback, but when I run the code,
callbacks = list( callback_tensorboard( log_dir = "my_log_dir", histogram_freq = 1, embeddings_freq = 1 ) )

then fit, this error occured:

Error in py_call_impl(callable, dots$args, dots$keywords) :
ValueError: To visualize embeddings, embeddings_data must be provided.

In the python version of keras, the Tensorboard callback has a `embeddings_data' parameter, and when embeddings_freq > 0, this must be set, but it doesn't exist in the R version

Most helpful comment

All 8 comments

Thanks for reporting this! Fix is here: https://github.com/rstudio/keras/commit/d478ad6998266320c401f6cb5a6d200189fc3006

Could you re-install with devtools::install_github("rstudio/keras") and let us know if that fixes the issue on your end?

I tried but still casting the same error. At least at GPU version tensorflow backend.

I am no longer seeing this error so I think we need to back up and have you provide us with the full set of code you are running along with the output of the following:

devtools::session_info()
tensorflow::tf_config()
keras:::keras_version()

I am getting the same error in Python, also with a Tensorflow-GPU back end. This is while running the example from section 7.2.2 in the Deep Learning With Python book.

I think you need to simply remove the embeddings_freq parameter from the call as it now requires embeddings_data to be passed as well and I'm not sure what the correct semantics for that are (I can't seem to get it to work and there isn't documentation that I can find on how to use it).

Okay, but that doesn't exactly fix the problem, it just removes the part that was causing a problem. Without that argument nothing gets embedded.

That's just the callback, - embeddings themselves are not affected...

But it seems like there was a PR changing related behavior in Keras 24 days ago (before, you could just pass embeddings_freq standalone):

https://github.com/keras-team/keras/pull/7766

We'll provide an R example how to use that.
In the meantime, see

https://github.com/keras-team/keras/blob/master/examples/tensorboard_embeddings_mnist.py

Was this page helpful?
0 / 5 - 0 ratings