Combine keras functional api with tensorflow

Question

It's possible to combine tensorflow with keras sequential models like this: (source)

from keras.models import Sequential, Model

model = Sequential()
model.add(Dense(32, activation='relu', input_dim=784))
model.add(Dense(10, activation='softmax'))

# this works! 
x = tf.placeholder(tf.float32, shape=(None, 784))
y = model(x)

However, I want to use the functional API like this:

x = tf.placeholder(tf.float32, shape=(None, 784))
y = Dense(10)(x)
model = Model(inputs=x, outputs=y)

but when I try to do this, I get these errors:

TypeError: Input tensors to a Model must be Keras tensors. Found: Tensor("Placeholder_2:0", shape=(?, 784), dtype=float32) (missing Keras metadata).

Answer 1

What you are looking for is the tensor argument of Input layer for the functional API.

tensor: Optional existing tensor to wrap into the Input layer. If set, the layer will not create a placeholder tensor.

Answer 2

The functional and seqential apis are two different ways to create a model object. But once that object you can treat them the same way. For example calling them whith tensorflow objects.

Here you can find documentation for the functional api

Here is a minimal example of using tensorflow tensors with the functional api.

import tensorflow as tf
from keras.layers import Input, Dense
from keras.models import Model

# Create the keras model.
inputs = Input(shape=(784,))
outputs = Dense(10)(inputs)
model = Model(inputs=inputs, outputs=outputs)

# Now that we have a model we can call it with x.
x = tf.placeholder(tf.float32, shape=(None, 784))
y = model(x)