Tensorflow Federated | tff.learning.from_keras_model() with a model with DenseFeature layer and multiple inputs

Question

I am trying to federate a keras model which has multiple inputs.
These some of these inputs are categorical and some of them are numerical, so I have some DenseFeature layers to embed the values.

The problem is that using tff.learning.from_keras_model() to expect as input_spec a dictionary with just 2 elements (x,y) but I have multiple inputs which then I have to distinguish in the model to perform the Embedding correctly with the feature_columns functions and the DenseFeature layers.

How can I handle the single feature columns if the model accepts just an ‘x’ as input without proper columns names?

Here is the code and the error:

def create_keras_model():
  l = tf.keras.layers

  # handling numerical columns 
  for header in numerical_column_names:
    feature_columns.append(feature_column.numeric_column(header))

  # handling the categorical feature  
  pickup = feature_column.categorical_column_with_vocabulary_list(
      'pickup_location_id', [i for i in range(number_of_locations)])
  #pickup_one_hot = feature_column.indicator_column(pickup)
  #feature_columns.append(pickup_one_hot)

  pickup_embedding = feature_column.embedding_column(pickup, dimension=64)
  #feature_columns.append(pickup_embedding)


  feature_inputs = {
    'pickup_week_day_sin': tf.keras.Input((1,), name='pickup_week_day_sin'),
    'pickup_week_day_cos': tf.keras.Input((1,), name='pickup_week_day_cos'),
    'pickup_hour_sin': tf.keras.Input((1,), name='pickup_hour_sin'),
    'pickup_hour_cos': tf.keras.Input((1,), name='pickup_hour_cos'),
    'pickup_month_sin': tf.keras.Input((1,), name='pickup_month_sin'),
    'pickup_month_cos': tf.keras.Input((1,), name='pickup_month_cos'),
  }
  numerical_features = l.DenseFeatures(feature_columns)(feature_inputs)#{'x': a}

  location_input = {
      'pickup_location_id': tf.keras.Input((1,), dtype=tf.dtypes.int32, name='pickup_location_id'),
  }
  categorical_features = l.DenseFeatures(pickup_embedding)(location_input)#{'x': a}
  #i = l.Input(shape=(64+6,))

  #embedded_lookup_feature = tf.feature_column.numeric_column('x', shape=(784))
  conca = l.Concatenate()([categorical_features, numerical_features])

  dense = l.Dense(128, activation='relu')(conca)
  dense_1 = l.Dense(128, activation='relu')(dense)
  dense_2 = layers.Dense(number_of_locations, kernel_initializer='zeros')(dense_1)
  output = l.Softmax()(dense_2)

  inputs = list(feature_inputs.values()) + list(location_input.values())
  return tf.keras.Model(inputs=inputs, outputs=output)

input_spec = preprocessed_example_dataset.element_spec
def model_fn():
  # We _must_ create a new model here, and _not_ capture it from an external
  # scope. TFF will call this within different graph contexts.
  keras_model = create_keras_model()
  return tff.learning.from_keras_model(
      keras_model,
      input_spec=input_spec,
      loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]
      )

error when called:

ValueError: The top-level structure in `dummy_batch` or `input_spec` must contain exactly two elements, as it must contain type information for both inputs to and predictions from the model.

preprocessed_example_dataset.element_spec:

OrderedDict([('pickup_location_id',
              TensorSpec(shape=(None,), dtype=tf.int32, name=None)),
             ('pickup_hour_sin',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('pickup_hour_cos',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('pickup_week_day_sin',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('pickup_week_day_cos',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('pickup_month_sin',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('pickup_month_cos',
              TensorSpec(shape=(None,), dtype=tf.float32, name=None)),
             ('y', TensorSpec(shape=(None,), dtype=tf.int32, name=None))])

Asked By: Alex

||

Source

Answer 1

I was able to find the answer looking at the Federate Learning repository on GitHub:

The way to do it is to make the ‘x’ value of the orderedDict an orderedDict itself using as keys the name of the columns we want as input.

A concrete example is given here: https://github.com/tensorflow/federated/blob/3b5a551c46e7eab61e40c943390868fca6422e21/tensorflow_federated/python/learning/keras_utils_test.py#L283

Where it define the input spec:

 input_spec = collections.OrderedDict(
        x=collections.OrderedDict(
            a=tf.TensorSpec(shape=[None, 1], dtype=tf.float32),
            b=tf.TensorSpec(shape=[1, 1], dtype=tf.float32)),
        y=tf.TensorSpec(shape=[None, 1], dtype=tf.float32))
    model = model_examples.build_multiple_inputs_keras_model()

To be used in the model defined as:

def build_multiple_inputs_keras_model():
  """Builds a test model with two inputs."""
  l = tf.keras.layers
  a = l.Input((1,), name='a')
  b = l.Input((1,), name='b')
  # Each input has a single, independent dense layer, which are combined into
  # a final dense layer.
  output = l.Dense(1)(
      l.concatenate([
          l.Dense(1)(a),
          l.Dense(1)(b),
      ]))
  return tf.keras.Model(inputs={'a': a, 'b': b}, outputs=[output])

Answered By: Alex

Tensorflow Federated | tff.learning.from_keras_model() with a model with DenseFeature layer and multiple inputs

Question:

Answers: