How do you write an input layer with Tensorflow's Functional API that expects a Dataset object?

Question

I am trying to create an Input Layer using tf.keras.Input using the type_spec argument to specify an input of DatasetSpec using Tensorflow's Functional API so that I can iterate over it later. If I try to define an input layer by specifying shape, I get error messages complaining that iterating over tf.tensor is not allowed.

X = np.random.uniform(size=(1000,75))
Y = np.random.uniform(size=(1000))

data = tf.data.Dataset.from_tensor_slices((X, Y))
data = data.batch(batch_size=100, drop_remainder=True)


input = tf.keras.Input(type_spec = tf.data.DatasetSpec.from_value(data))

I got the following error: ValueError: KerasTensor only supports TypeSpecs that have a shape field; got DatasetSpec, which does not have a shape.

score 0 · Accepted Answer · answered Oct 10 '22 at 18:35

The data object (output of tf.data) is the tuple of two item (X and Y). In order to create specification layer (keras.Input) with tf.data.DatasetSpec.from_value(data), you need to get the shape of first element.

tf.data.DatasetSpec.from_value(data)

DatasetSpec((TensorSpec(shape=(100, 75), dtype=tf.float64, name=None),
TensorSpec(shape=(100,), dtype=tf.float64, name=None)), TensorShape([]))

Here is a dummy model with your approach,

# element_spec[0] <- shape of the first element.
input = tf.keras.Input(
    type_spec = tf.data.DatasetSpec.from_value(data).element_spec[0]
)
output = tf.keras.layers.Dense(1)(input)
model = tf.keras.Model(input, output)
model.compile(loss='mse')

model.fit(data, epochs=3)
Epoch 1/3
10/10 [==============================] - 0s 2ms/step - loss: 0.2506
Epoch 2/3
10/10 [==============================] - 0s 2ms/step - loss: 0.2441
Epoch 3/3
10/10 [==============================] - 0s 2ms/step - loss: 0.2378

How do you write an input layer with Tensorflow's Functional API that expects a Dataset object?

1 Answers1