tf.keras.layers.DenseFeatures api as the candidate of hb.feature_column.DenseFeatures can not work with tf.feature_column.shared_embedding_columns

[taoyun951753]

Current behavior

HB version: HybridBackend 0.7.0-e277c15f3843f98901f0795bc9b7d0768056d5a3; tf1.15.5-v1.15.5+nv22.06-0-g55be3962f8; g++ 7.5.0; CUDA 11.4 (70,75,80,86)
new hb package removes the hb.feature_column.DenseFeatures api. However, tf.keras.layers.DenseFeatures api can not deal with
tf.feature_column.shared_embedding_columns.

Expected behavior

I want a new way to run the code successfully without hb.feature_column.DenseFeatures.

System information

• GPU model and memory:
• OS Platform:
• Docker version:
• GCC/CUDA/cuDNN version:
• Python/conda version:
• TensorFlow/PyTorch version:

Code to reproduce

Willing to contribute

Yes

[francktcheng]

We recommend you to use the tf.feature_column.input_layer instead of tf.keras.layers.DenseFeatures when the feature columns are generated by using tf.feature_column.shared_embedding_columns. Please refer to the code example as follows:

...
self._embedding_columns = tf.feature_column.shared_embedding_columns(
        [tf.feature_column.categorical_column_with_identity(
          key=f,
          num_buckets=self._args.data_spec.embedding_sizes[f],
          default_value=self._args.data_spec.defaults[f])
          for f in self._categorical_fields],
        dimension=self._args.embedding_dim,
        initializer=tf.random_uniform_initializer(-1e-3, 1e-3)
      )
cols_to_output_tensors = {}
      tf.feature_column.input_layer(
        {f: features[f] for f in self._categorical_fields},
        self._embedding_columns,
        cols_to_output_tensors=cols_to_output_tensors)
 ...

Internally, the tf.feature_column.input_layer API does accept feature columns with a type of _FeatureColumn while the API of tf.keras.layers.DenseFeatures only accepts a type derived from FeatureColumn which is not compatible with that of a shared embedding.