How does the Flatten() Layer work in Tensorflow?

Question

Im currently working with tensorflow and neural networks and im quite new to the topic.

Im having a stack of 4 images passed to my conv network in the shape of (4,160,120,1) as the images are in grayscale.

After passing my images through the neural network i wanted to flatten the images into one long array that gets passed to dense layers.

But after using Flatten() on the output of my neural network i get a 2 dimensional array in the shape of (4, 2240) instead of a long one dimensional array.

Why is that and what would i need to do to flatten it to a 1D array? Just use Flatten() again?

Im asking because my goal is to concatenate the flattened with another input which is one dimensional but i get an array that concatenate wont work bc they dont have the same shape

Thanks for the help!

Answer 1

Keras layers don't affect batch dimension. You have to use tf.reshape() before feeding your data to the model:

input1 = tf.reshape(input1, (-1,))
output = model(input1, input2) # exclude Flatten layer from your model