What is the right way to manipulate the shape of a tensor when there are unknown elements in it?

Question

Let's say that I have a tensor of shape (None, None, None, 32) and I want to reshape this to (None, None, 32) where the middle dimension is the product of two middle dimensions of the original one. What is the right way to do so?

Answer 1

import keras.backend as K

def flatten_pixels(x):
    shape = K.shape(x)
    newShape = K.concatenate([
                                 shape[0:1], 
                                 shape[1:2] * shape[2:3],
                                 shape[3:4]
                             ])

    return K.reshape(x, newShape)

Use it in a Lambda layer:

from keras.layers import Lambda

model.add(Lambda(flatten_pixels))

---

A little knowledge:

• K.shape returns the "current" shape of the tensor, containing data - It's a Tensor containing int values for all dimensions. It only exists properly when running the model and can't be used in model definition, only in runtime calculations.
• K.int_shape returns the "definition" shape of the tensor as a tuple. This means the variable dimensions will come containing None values.