Python feature_column.embedding_column函数代码示例

  def testFromScratchWithCustomRNNCellFn(self):
    def train_input_fn():
      return {
          'tokens':
              sparse_tensor.SparseTensor(
                  values=['the', 'cat', 'sat'],
                  indices=[[0, 0], [0, 1], [0, 2]],
                  dense_shape=[1, 3]),
      }, [[1]]

    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    input_units = 2
    cell_units = [4, 2]
    n_classes = 2

    def rnn_cell_fn(mode):
      del mode  # unused
      cells = [rnn_cell.BasicRNNCell(num_units=n) for n in cell_units]
      return rnn_cell.MultiRNNCell(cells)

    est = rnn.RNNClassifier(
        sequence_feature_columns=[embed],
        rnn_cell_fn=rnn_cell_fn,
        n_classes=n_classes,
        model_dir=self._model_dir)

    # Train for a few steps, and validate final checkpoint.
    num_steps = 10
    est.train(input_fn=train_input_fn, steps=num_steps)
    self._assert_checkpoint(n_classes, input_units, cell_units, num_steps)

  def _build_feature_columns(self):
    col = fc.categorical_column_with_identity(
        'int_ctx', num_buckets=100)
    ctx_cols = [
        fc.embedding_column(col, dimension=10),
        fc.numeric_column('float_ctx')]

    identity_col = sfc.sequence_categorical_column_with_identity(
        'int_list', num_buckets=10)
    bucket_col = sfc.sequence_categorical_column_with_hash_bucket(
        'bytes_list', hash_bucket_size=100)
    seq_cols = [
        fc.embedding_column(identity_col, dimension=10),
        fc.embedding_column(bucket_col, dimension=20)]

    return ctx_cols, seq_cols

  def _testExampleWeight(self, n_classes):
    def train_input_fn():
      return {
          'tokens':
              sparse_tensor.SparseTensor(
                  values=['the', 'cat', 'sat', 'dog', 'barked'],
                  indices=[[0, 0], [0, 1], [0, 2], [1, 0], [1, 1]],
                  dense_shape=[2, 3]),
          'w': [[1], [2]],
      }, [[1], [0]]

    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    input_units = 2

    cell_units = [4, 2]
    est = rnn.RNNClassifier(
        num_units=cell_units,
        sequence_feature_columns=[embed],
        n_classes=n_classes,
        weight_column='w',
        model_dir=self._model_dir)

    # Train for a few steps, and validate final checkpoint.
    num_steps = 10
    est.train(input_fn=train_input_fn, steps=num_steps)
    self._assert_checkpoint(n_classes, input_units, cell_units, num_steps)

  def test_dnn_classifier(self):
    embedding = feature_column_lib.embedding_column(
        feature_column_lib.categorical_column_with_vocabulary_list(
            'wire_cast', ['kima', 'omar', 'stringer']), 8)
    dnn = estimator_lib.DNNClassifier(
        feature_columns=[embedding], hidden_units=[3, 1])

    def train_input_fn():
      return dataset_ops.Dataset.from_tensors(({
          'wire_cast': [['omar'], ['kima']]
      }, [[0], [1]])).repeat(3)

    def eval_input_fn():
      return dataset_ops.Dataset.from_tensors(({
          'wire_cast': [['stringer'], ['kima']]
      }, [[0], [1]])).repeat(2)

    evaluator = hooks_lib.InMemoryEvaluatorHook(
        dnn, eval_input_fn, name='in-memory')
    dnn.train(train_input_fn, hooks=[evaluator])
    self.assertTrue(os.path.isdir(dnn.eval_dir('in-memory')))
    step_keyword_to_value = summary_step_keyword_to_value_mapping(
        dnn.eval_dir('in-memory'))

    final_metrics = dnn.evaluate(eval_input_fn)
    step = final_metrics[ops.GraphKeys.GLOBAL_STEP]
    for summary_tag in final_metrics:
      if summary_tag == ops.GraphKeys.GLOBAL_STEP:
        continue
      self.assertEqual(final_metrics[summary_tag],
                       step_keyword_to_value[step][summary_tag])

  def test_sequence_length_with_empty_rows(self):
    """Tests _sequence_length when some examples do not have ids."""
    vocabulary_size = 3
    sparse_input = sparse_tensor.SparseTensorValue(
        # example 0, ids []
        # example 1, ids [2]
        # example 2, ids [0, 1]
        # example 3, ids []
        # example 4, ids [1]
        # example 5, ids []
        indices=((1, 0), (2, 0), (2, 1), (4, 0)),
        values=(2, 0, 1, 1),
        dense_shape=(6, 2))
    expected_sequence_length = [0, 1, 2, 0, 1, 0]

    categorical_column = sfc.sequence_categorical_column_with_identity(
        key='aaa', num_buckets=vocabulary_size)
    embedding_column = fc.embedding_column(
        categorical_column, dimension=2)

    _, sequence_length = embedding_column._get_sequence_dense_tensor(
        _LazyBuilder({'aaa': sparse_input}))

    with monitored_session.MonitoredSession() as sess:
      self.assertAllEqual(
          expected_sequence_length, sequence_length.eval(session=sess))

  def testWarmStartInputLayerEmbeddingColumn(self):
    # Create old and new vocabs for embedding column "sc_vocab".
    prev_vocab_path = self._write_vocab(["apple", "banana", "guava", "orange"],
                                        "old_vocab")
    new_vocab_path = self._write_vocab(
        ["orange", "guava", "banana", "apple", "raspberry", "blueberry"],
        "new_vocab")

    # Save checkpoint from which to warm-start.
    with ops.Graph().as_default() as g:
      with self.test_session(graph=g) as sess:
        _ = variable_scope.get_variable(
            "input_layer/sc_vocab_embedding/embedding_weights",
            initializer=[[0.5, 0.4], [1., 1.1], [2., 2.2], [3., 3.3]])
        self._write_checkpoint(sess)

    def _partitioner(shape, dtype):  # pylint:disable=unused-argument
      # Partition each var into 2 equal slices.
      partitions = [1] * len(shape)
      partitions[0] = min(2, shape[0].value)
      return partitions

    # Create feature columns.
    sc_vocab = fc.categorical_column_with_vocabulary_file(
        "sc_vocab", vocabulary_file=new_vocab_path, vocabulary_size=6)
    emb_vocab = fc.embedding_column(
        categorical_column=sc_vocab,
        dimension=2,
        # Can't use constant_initializer with load_and_remap.  In practice,
        # use a truncated normal initializer.
        initializer=init_ops.random_uniform_initializer(
            minval=0.42, maxval=0.42))
    all_deep_cols = [emb_vocab]
    # New graph, new session with warmstarting.
    with ops.Graph().as_default() as g:
      with self.test_session(graph=g) as sess:
        cols_to_vars = {}
        with variable_scope.variable_scope("", partitioner=_partitioner):
          # Create the variables.
          fc.input_layer(
              features=self._create_dummy_inputs(),
              feature_columns=all_deep_cols,
              cols_to_vars=cols_to_vars)
        ws_settings = ws_util._WarmStartSettings(
            self.get_temp_dir(), col_to_prev_vocab={
                emb_vocab: prev_vocab_path
            })
        ws_util._warmstart_input_layer(cols_to_vars, ws_settings)
        sess.run(variables.global_variables_initializer())
        # Verify weights were correctly warmstarted. Var corresponding to
        # emb_vocab should be correctly warmstarted after vocab remapping.
        # Missing values are filled in with the EmbeddingColumn's initializer.
        self._assert_cols_to_vars(
            cols_to_vars, {
                emb_vocab: [
                    np.array([[3., 3.3], [2., 2.2], [1., 1.1]]),
                    np.array([[0.5, 0.4], [0.42, 0.42], [0.42, 0.42]])
                ]
            }, sess)

def _sequence_embedding_column(
    categorical_column, dimension, initializer=None, ckpt_to_load_from=None,
    tensor_name_in_ckpt=None, max_norm=None, trainable=True):
  """Returns a feature column that represents sequences of embeddings.

  Use this to convert sequence categorical data into dense representation for
  input to sequence NN, such as RNN.

  Example:

  ```python
  watches = sequence_categorical_column_with_identity(
      'watches', num_buckets=1000)
  watches_embedding = _sequence_embedding_column(watches, dimension=10)
  columns = [watches]

  features = tf.parse_example(..., features=make_parse_example_spec(columns))
  input_layer, sequence_length = sequence_input_layer(features, columns)

  rnn_cell = tf.nn.rnn_cell.BasicRNNCell(hidden_size)
  outputs, state = tf.nn.dynamic_rnn(
      rnn_cell, inputs=input_layer, sequence_length=sequence_length)
  ```

  Args:
    categorical_column: A `_SequenceCategoricalColumn` created with a
      `sequence_cateogrical_column_with_*` function.
    dimension: Integer dimension of the embedding.
    initializer: Initializer function used to initialize the embeddings.
    ckpt_to_load_from: String representing checkpoint name/pattern from which to
      restore column weights. Required if `tensor_name_in_ckpt` is not `None`.
    tensor_name_in_ckpt: Name of the `Tensor` in `ckpt_to_load_from` from
      which to restore the column weights. Required if `ckpt_to_load_from` is
      not `None`.
    max_norm: If not `None`, embedding values are l2-normalized to this value.
    trainable: Whether or not the embedding is trainable. Default is True.

  Returns:
    A `_SequenceCategoricalToDenseColumn`.

  Raises:
    ValueError: If `categorical_column` is not the right type.
  """
  if not isinstance(categorical_column, _SequenceCategoricalColumn):
    raise ValueError(
        'categorical_column must be of type _SequenceCategoricalColumn. '
        'Given (type {}): {}'.format(
            type(categorical_column), categorical_column))
  return _SequenceCategoricalToDenseColumn(
      fc.embedding_column(
          categorical_column,
          dimension=dimension,
          initializer=initializer,
          ckpt_to_load_from=ckpt_to_load_from,
          tensor_name_in_ckpt=tensor_name_in_ckpt,
          max_norm=max_norm,
          trainable=trainable))

  def testParseExampleInputFn(self):
    """Tests complete flow with input_fn constructed from parse_example."""
    n_classes = 3
    batch_size = 10
    words = [b'dog', b'cat', b'bird', b'the', b'a', b'sat', b'flew', b'slept']

    _, examples_file = tempfile.mkstemp()
    writer = python_io.TFRecordWriter(examples_file)
    for _ in range(batch_size):
      sequence_length = random.randint(1, len(words))
      sentence = random.sample(words, sequence_length)
      label = random.randint(0, n_classes - 1)
      example = example_pb2.Example(features=feature_pb2.Features(
          feature={
              'tokens':
                  feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
                      value=sentence)),
              'label':
                  feature_pb2.Feature(int64_list=feature_pb2.Int64List(
                      value=[label])),
          }))
      writer.write(example.SerializeToString())
    writer.close()

    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    feature_columns = [embed]
    feature_spec = parsing_utils.classifier_parse_example_spec(
        feature_columns,
        label_key='label',
        label_dtype=dtypes.int64)

    def _train_input_fn():
      dataset = readers.make_batched_features_dataset(
          examples_file, batch_size, feature_spec)
      return dataset.map(lambda features: (features, features.pop('label')))
    def _eval_input_fn():
      dataset = readers.make_batched_features_dataset(
          examples_file, batch_size, feature_spec, num_epochs=1)
      return dataset.map(lambda features: (features, features.pop('label')))
    def _predict_input_fn():
      dataset = readers.make_batched_features_dataset(
          examples_file, batch_size, feature_spec, num_epochs=1)
      def features_fn(features):
        features.pop('label')
        return features
      return dataset.map(features_fn)

    self._test_complete_flow(
        feature_columns=feature_columns,
        train_input_fn=_train_input_fn,
        eval_input_fn=_eval_input_fn,
        predict_input_fn=_predict_input_fn,
        n_classes=n_classes,
        batch_size=batch_size)

  def test_warm_starting_selective_variables(self):
    """Tests selecting variables to warm-start."""
    age = feature_column.numeric_column('age')
    city = feature_column.embedding_column(
        feature_column.categorical_column_with_vocabulary_list(
            'city', vocabulary_list=['Mountain View', 'Palo Alto']),
        dimension=5)

    # Create a DNNLinearCombinedClassifier and train to save a checkpoint.
    dnn_lc_classifier = dnn_linear_combined.DNNLinearCombinedClassifier(
        linear_feature_columns=[age],
        dnn_feature_columns=[city],
        dnn_hidden_units=[256, 128],
        model_dir=self._ckpt_and_vocab_dir,
        n_classes=4,
        linear_optimizer='SGD',
        dnn_optimizer='SGD')
    dnn_lc_classifier.train(input_fn=self._input_fn, max_steps=1)

    # Create a second DNNLinearCombinedClassifier, warm-started from the first.
    # Use a learning_rate = 0.0 optimizer to check values (use SGD so we don't
    # have accumulator values that change).
    warm_started_dnn_lc_classifier = (
        dnn_linear_combined.DNNLinearCombinedClassifier(
            linear_feature_columns=[age],
            dnn_feature_columns=[city],
            dnn_hidden_units=[256, 128],
            n_classes=4,
            linear_optimizer=gradient_descent.GradientDescentOptimizer(
                learning_rate=0.0),
            dnn_optimizer=gradient_descent.GradientDescentOptimizer(
                learning_rate=0.0),
            # The provided regular expression will only warm-start the deep
            # portion of the model.
            warm_start_from=estimator.WarmStartSettings(
                ckpt_to_initialize_from=dnn_lc_classifier.model_dir,
                vars_to_warm_start='.*(dnn).*')))

    warm_started_dnn_lc_classifier.train(input_fn=self._input_fn, max_steps=1)
    for variable_name in warm_started_dnn_lc_classifier.get_variable_names():
      if 'dnn' in variable_name:
        self.assertAllClose(
            dnn_lc_classifier.get_variable_value(variable_name),
            warm_started_dnn_lc_classifier.get_variable_value(variable_name))
      elif 'linear' in variable_name:
        linear_values = warm_started_dnn_lc_classifier.get_variable_value(
            variable_name)
        # Since they're not warm-started, the linear weights will be
        # zero-initialized.
        self.assertAllClose(np.zeros_like(linear_values), linear_values)

  def test_get_sequence_dense_tensor(self):
    vocabulary_size = 3
    sparse_input = sparse_tensor.SparseTensorValue(
        # example 0, ids [2]
        # example 1, ids [0, 1]
        # example 2, ids []
        # example 3, ids [1]
        indices=((0, 0), (1, 0), (1, 1), (3, 0)),
        values=(2, 0, 1, 1),
        dense_shape=(4, 2))

    embedding_dimension = 2
    embedding_values = (
        (1., 2.),  # id 0
        (3., 5.),  # id 1
        (7., 11.)  # id 2
    )
    def _initializer(shape, dtype, partition_info):
      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
      self.assertEqual(dtypes.float32, dtype)
      self.assertIsNone(partition_info)
      return embedding_values

    expected_lookups = [
        # example 0, ids [2]
        [[7., 11.], [0., 0.]],
        # example 1, ids [0, 1]
        [[1., 2.], [3., 5.]],
        # example 2, ids []
        [[0., 0.], [0., 0.]],
        # example 3, ids [1]
        [[3., 5.], [0., 0.]],
    ]

    categorical_column = sfc.sequence_categorical_column_with_identity(
        key='aaa', num_buckets=vocabulary_size)
    embedding_column = fc.embedding_column(
        categorical_column, dimension=embedding_dimension,
        initializer=_initializer)

    embedding_lookup, _ = embedding_column._get_sequence_dense_tensor(
        _LazyBuilder({'aaa': sparse_input}))

    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
    self.assertItemsEqual(
        ('embedding_weights:0',), tuple([v.name for v in global_vars]))
    with monitored_session.MonitoredSession() as sess:
      self.assertAllEqual(embedding_values, global_vars[0].eval(session=sess))
      self.assertAllEqual(expected_lookups, embedding_lookup.eval(session=sess))

def _sequence_embedding_column(
    categorical_column, dimension, initializer=None, ckpt_to_load_from=None,
    tensor_name_in_ckpt=None, max_norm=None, trainable=True):
  if not isinstance(categorical_column, _SequenceCategoricalColumn):
    raise ValueError(
        'categorical_column must be of type _SequenceCategoricalColumn. '
        'Given (type {}): {}'.format(
            type(categorical_column), categorical_column))
  return _SequenceEmbeddingColumn(
      fc.embedding_column(
          categorical_column,
          dimension=dimension,
          initializer=initializer,
          ckpt_to_load_from=ckpt_to_load_from,
          tensor_name_in_ckpt=tensor_name_in_ckpt,
          max_norm=max_norm,
          trainable=trainable))

  def testNumpyInputFn(self):
    """Tests complete flow with numpy_input_fn."""
    n_classes = 3
    batch_size = 10
    words = ['dog', 'cat', 'bird', 'the', 'a', 'sat', 'flew', 'slept']
    # Numpy only supports dense input, so all examples will have same length.
    # TODO(b/73160931): Update test when support for prepadded data exists.
    sequence_length = 3

    features = []
    for _ in range(batch_size):
      sentence = random.sample(words, sequence_length)
      features.append(sentence)

    x_data = np.array(features)
    y_data = np.random.randint(n_classes, size=batch_size)

    train_input_fn = numpy_io.numpy_input_fn(
        x={'tokens': x_data},
        y=y_data,
        batch_size=batch_size,
        num_epochs=None,
        shuffle=True)
    eval_input_fn = numpy_io.numpy_input_fn(
        x={'tokens': x_data},
        y=y_data,
        batch_size=batch_size,
        shuffle=False)
    predict_input_fn = numpy_io.numpy_input_fn(
        x={'tokens': x_data},
        batch_size=batch_size,
        shuffle=False)

    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    feature_columns = [embed]

    self._test_complete_flow(
        feature_columns=feature_columns,
        train_input_fn=train_input_fn,
        eval_input_fn=eval_input_fn,
        predict_input_fn=predict_input_fn,
        n_classes=n_classes,
        batch_size=batch_size)

  def _test_complete_flow(
      self, train_input_fn, eval_input_fn, predict_input_fn, n_classes,
      batch_size):
    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    feature_columns = [embed]

    cell_units = [4, 2]
    est = rnn.RNNClassifier(
        num_units=cell_units,
        sequence_feature_columns=feature_columns,
        n_classes=n_classes,
        model_dir=self._model_dir)

    # TRAIN
    num_steps = 10
    est.train(train_input_fn, steps=num_steps)

    # EVALUATE
    scores = est.evaluate(eval_input_fn)
    self.assertEqual(num_steps, scores[ops.GraphKeys.GLOBAL_STEP])
    self.assertIn('loss', six.iterkeys(scores))

    # PREDICT
    predicted_proba = np.array([
        x[prediction_keys.PredictionKeys.PROBABILITIES]
        for x in est.predict(predict_input_fn)
    ])
    self.assertAllEqual((batch_size, n_classes), predicted_proba.shape)

    # EXPORT
    feature_spec = {
        'tokens': parsing_ops.VarLenFeature(dtypes.string),
        'label': parsing_ops.FixedLenFeature([1], dtypes.int64),
    }
    serving_input_receiver_fn = export.build_parsing_serving_input_receiver_fn(
        feature_spec)
    export_dir = est.export_savedmodel(tempfile.mkdtemp(),
                                       serving_input_receiver_fn)
    self.assertTrue(gfile.Exists(export_dir))

  def testConflictingRNNCellFn(self):
    col = seq_fc.sequence_categorical_column_with_hash_bucket(
        'tokens', hash_bucket_size=10)
    embed = fc.embedding_column(col, dimension=2)
    cell_units = [4, 2]

    with self.assertRaisesRegexp(
        ValueError,
        'num_units and cell_type must not be specified when using rnn_cell_fn'):
      rnn.RNNClassifier(
          sequence_feature_columns=[embed],
          rnn_cell_fn=lambda x: x,
          num_units=cell_units)

    with self.assertRaisesRegexp(
        ValueError,
        'num_units and cell_type must not be specified when using rnn_cell_fn'):
      rnn.RNNClassifier(
          sequence_feature_columns=[embed],
          rnn_cell_fn=lambda x: x,
          cell_type='lstm')

  def test_sequence_length(self):
    vocabulary_size = 3
    sparse_input = sparse_tensor.SparseTensorValue(
        # example 0, ids [2]
        # example 1, ids [0, 1]
        indices=((0, 0), (1, 0), (1, 1)),
        values=(2, 0, 1),
        dense_shape=(2, 2))
    expected_sequence_length = [1, 2]

    categorical_column = sfc.sequence_categorical_column_with_identity(
        key='aaa', num_buckets=vocabulary_size)
    embedding_column = fc.embedding_column(
        categorical_column, dimension=2)

    _, sequence_length = embedding_column._get_sequence_dense_tensor(
        _LazyBuilder({'aaa': sparse_input}))

    with monitored_session.MonitoredSession() as sess:
      sequence_length = sess.run(sequence_length)
      self.assertAllEqual(expected_sequence_length, sequence_length)
      self.assertEqual(np.int64, sequence_length.dtype)