tf.data with multiple inputs / outputs in Keras

For the application, such as pair text similarity, the input data is similar to: pair_1, pair_2. In these problems, we usually have multiple input data. Previously, I implemented my models successfully:

model.fit([pair_1, pair_2], labels, epochs=50)

I decided to replace my input pipeline with tf.data API. To this end, I create a Dataset similar to:

dataset = tf.data.Dataset.from_tensor_slices((pair_1, pair2, labels))

It compiles successfully but when start to train it throws the following exception:

AttributeError: 'tuple' object has no attribute 'ndim'

My Keras and Tensorflow version respectively are 2.1.6 and 1.11.0. I found a similar issue in Tensorflow repository: tf.keras multi-input models don't work when using tf.data.Dataset.

Does anyone know how to fix the issue?

Here is some main part of the code:

(q1_test, q2_test, label_test) = test
(q1_train, q2_train, label_train) = train

    def tfdata_generator(sent1, sent2, labels, is_training):
        '''Construct a data generator using tf.Dataset'''

        dataset = tf.data.Dataset.from_tensor_slices((sent1, sent2, labels))
        if is_training:
            dataset = dataset.shuffle(1000)  # depends on sample size

        dataset = dataset.repeat()
        dataset = dataset.prefetch(tf.contrib.data.AUTOTUNE)

        return dataset

train_dataset = tfdata_generator(q1_train, q2_train, label_train, is_training=True, batch_size=_BATCH_SIZE)
test_dataset = tfdata_generator(q1_test, q2_test, label_test, is_training=False, batch_size=_BATCH_SIZE)


inps1 = keras.layers.Input(shape=(50,))
inps2 = keras.layers.Input(shape=(50,))

embed = keras.layers.Embedding(input_dim=nb_vocab, output_dim=300, weights=[embedding], trainable=False)
embed1 = embed(inps1)
embed2 = embed(inps2)

gru = keras.layers.CuDNNGRU(256)
gru1 = gru(embed1)
gru2 = gru(embed2)

concat = keras.layers.concatenate([gru1, gru2])

preds = keras.layers.Dense(1, 'sigmoid')(concat)

model = keras.models.Model(inputs=[inps1, inps2], outputs=preds)
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
print(model.summary())

model.fit(
    train_dataset.make_one_shot_iterator(),
    steps_per_epoch=len(q1_train) // _BATCH_SIZE,
    epochs=50,
    validation_data=test_dataset.make_one_shot_iterator(),
    validation_steps=len(q1_test) // _BATCH_SIZE,
    verbose=1)

I'm not using Keras but I would go with an tf.data.Dataset.from_generator() - like:

def _input_fn():
  sent1 = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.int64)
  sent2 = np.array([20, 25, 35, 40, 600, 30, 20, 30], dtype=np.int64)
  sent1 = np.reshape(sent1, (8, 1, 1))
  sent2 = np.reshape(sent2, (8, 1, 1))

  labels = np.array([40, 30, 20, 10, 80, 70, 50, 60], dtype=np.int64)
  labels = np.reshape(labels, (8, 1))

  def generator():
    for s1, s2, l in zip(sent1, sent2, labels):
      yield {"input_1": s1, "input_2": s2}, l

  dataset = tf.data.Dataset.from_generator(generator, output_types=({"input_1": tf.int64, "input_2": tf.int64}, tf.int64))
  dataset = dataset.batch(2)
  return dataset

...

model.fit(_input_fn(), epochs=10, steps_per_epoch=4)

This generator can iterate over your e.g text-files / numpy arrays and yield on every call a example. In this example, I assume that the word of the sentences are already converted to the indices in the vocabulary.

Edit: Since OP asked, it should be also possible with Dataset.from_tensor_slices():

def _input_fn():
  sent1 = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.int64)
  sent2 = np.array([20, 25, 35, 40, 600, 30, 20, 30], dtype=np.int64)
  sent1 = np.reshape(sent1, (8, 1))
  sent2 = np.reshape(sent2, (8, 1))

  labels = np.array([40, 30, 20, 10, 80, 70, 50, 60], dtype=np.int64)
  labels = np.reshape(labels, (8))

  dataset = tf.data.Dataset.from_tensor_slices(({"input_1": sent1, "input_2": sent2}, labels))
  dataset = dataset.batch(2, drop_remainder=True)
  return dataset

One way to solve your issue could be to use the zip dataset to combine your various inputs:

sent1 = np.array([1, 2, 3, 4, 5, 6, 7, 8], dtype=np.float32)
sent2 = np.array([20, 25, 35, 40, 600, 30, 20, 30], dtype=np.float32)
sent1 = np.reshape(sent1, (8, 1, 1))
sent2 = np.reshape(sent2, (8, 1, 1))

labels = np.array([40, 30, 20, 10, 80, 70, 50, 60], dtype=np.float32)
labels = np.reshape(labels, (8, 1))

dataset_12 = tf.data.Dataset.from_tensor_slices((sent_1, sent_2))
dataset_label = tf.data.Dataset.from_tensor_slices(labels)

dataset = tf.data.Dataset.zip((dataset_12, dataset_label)).batch(2).repeat()
model.fit(dataset, epochs=10, steps_per_epoch=4)

will print: Epoch 1/10 4/4 [==============================] - 2s 503ms/step...