"no value provided for label" error with TFX Keras + Evaluator component about model-analysis HOT 8 CLOSED

To add a bit of additional context, in case its important, I based the training for the Keras model based on the code from this ipynb example: https://github.com/tensorflow/tfx/blob/master/tfx/examples/chicago_taxi_pipeline/taxi_pipeline_native_keras.py

which is to say, this module file: https://github.com/tensorflow/tfx/blob/master/tfx/examples/chicago_taxi_pipeline/taxi_utils_native_keras.py

we have some slight modifications in terms of removing some of the transform features (specifically the bucket features), but its pretty much otherwise the same. If the problem may be in the construction of the model I could provide the exact code we have.

from model-analysis.

I can't be certain without more logs, but it looks like the 'tips' feature is not present in all of the input examples (or was removed later on in the pipeline). Note that because keras doesn't have support for label transformations yet, TFMA requires the label that is used to be present in the raw inputs. The taxi example was updated to add a 'big_tipper' label to account for this:

https://github.com/tensorflow/tfx/blob/1985b30b39d8833b6fe724828ce97441813cdb5a/tfx/examples/chicago_taxi_pipeline/taxi_utils_native_keras.py#L65

from model-analysis.

Thanks a lot @mdreves, we confirm it is now working for us! Do you have a timeline on when support for label transformations could be added to Keras?

from model-analysis.

It depends on KPL (keras preprocessing layer) support, but we are actively working on this quarter. If you are using TFT (tensorflow transform) with keras then that will likely land first.

from model-analysis.

We are using TFT, we'll keep an eye on the upcoming releases then. Thanks for your help!

from model-analysis.

@sngahane @mdreves which TFX version allows for Keras label transformations? I did not see it mentioned on the release notes. I am currently on 0.27.0, and I am hesitant to upgrade as I know there will be lots of things to change.

from model-analysis.

I see that @mdreves committed changes in the taxi example for Keras label transformations. A strong argument for me to upgrade to 1.0.0.

from model-analysis.

Hi @axeltidemann, those committed changes above have made it into the 0.28 release, so any version >= 0.28 should have this support.

from model-analysis.