Add support for keras layer DenseFeatures

[blidblid]

TensorFlow.js version

1.3.2

Browser version

Chome 78.0.3904.108

Describe the feature request

Add support for the keras layer DenseFeatures.

Code to reproduce the bug / link to feature request

Step 1:

Create a model in Python using the DenseFeatures layer.

       feature_columns = [tf.feature_column.numeric_column('foo')]
       feature_layer = tf.keras.layers.DenseFeatures(feature_columns)
       model = tf.keras.Sequential([
            feature_layer,
            tf.keras.layers.Dense(8),
            tf.keras.layers.Dense(1)
      ])

Step 2:

Train, fit and then save the model using tfjs.converters.save_keras_model(model, 'path_to_model')

Step 3:

Load the model into tfjs: tf.loadLayersModel('./model.json');
It will throw: Error: Uncaught (in promise): Error: Unknown layer: DenseFeatures.

[blidblid]

Will this be part of any upcoming tfjs release?

[tafsiri]

@caisq I don't think anyone is actively working on this. Mind if I add a contributions welcome tag here?

[caisq]

@tafsiri Please go ahead.

[blidblid]

For any interested, a workaround is to freeze the model. That way you can use DenseFeatures and give predictions in tfjs:

model.save(temp_path, overwrite=True)
tfjs.converters.convert_tf_saved_model(temp_path, path)

and then

const model = await tf.loadGraphModel(path);

[djemeljanovs]

@blidblid, For some reasons suggested workaround does not apply for my model with the error:
AssertionError: Identity is not in graph

Meanwhile I have started porting DenseFeatures layer to tfjs in this branch:
https://github.com/djemeljanovs/tfjs/tree/feature/dense-feature-layer-support

It seems to be excessive amount of work, since not only DenseFeatures, but whole FeatureColumn API has not been ported to tfjs yet.

@tafsiri, I tried to create branch for this on tfjs, but git did not allow me that, so I created a fork. Could you please check out if this is not the right place to do, I am quite new to contributions. Thank you!

[inguelberth]

For any interested, a workaround is to freeze the model. That way you can use DenseFeatures and give predictions in tfjs:

model.save(temp_path, overwrite=True)
tfjs.converters.convert_tf_saved_model(temp_path, path)

and then

const model = await tf.loadGraphModel(path);

hey @blidblid in which version of tensorflow have you tested this trick?

[blidblid]

Python                   3.7.5
tensorflow               2.3.0
tensorflowjs             1.5.2

Not all feature columns seem to work with this method. I had to use categorical_column_with_identity over categorical_column_with_vocabulary_list.

[gaikwadrahul8]

Hi, @blidblid

Apologize for the delayed response and It seems like we haven't implemented this feature request for DenseFeatures Op as of now and May I know still are you looking for this DenseFeatures Op ?

If someone wants to contribute for DenseFeatures Op, you're always welcome and please feel free to do it and refer this link. Thank you!

[google-ml-butler]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you.

[google-ml-butler]

Closing as stale. Please @mention us if this needs more attention.