Lambda layer for smooth image

TensorFlow General Discussion
1

Hi, I need to create a lambda layer to smooth an image. First I created a function that only calls tensor flow functions to smooth a 1D tensor.
To smooth an image this function is applied to all rows and then to all columns of all image colors layers.

What I can’t implement is the lambda layer that uses to smooth the image by rows and by columns of all colors layers. Could someone help me thanks?

3

We have already some smoothing functions in Tensorflow Addons:

We are duplicating these also in Keras:

4

I did the implementation as follow:

import tensorflow as tf
def bMatrix(m, lmbd = 1600, d = 2, dtype=tf.float32):
    E = tf.linalg.diag(tf.repeat(tf.constant(1), m))
    D = tf.experimental.numpy.diff(E,n=d,axis=0)
    B = E + (lmbd * (tf.linalg.matmul(a=D,b=D,transpose_a=True)))
    B = tf.linalg.inv(tf.cast(B,dtype=dtype))
    return B

def whittaker1D(m, lmbd = 5, d = 2, dtype=tf.float32):
    B = bMatrix(m, lmbd, d, dtype)
    def W(y):
        return tf.linalg.matvec(B,y)
    return tf.function(W,input_signature=(tf.TensorSpec(shape=(m), dtype=dtype),))

def whittaker2D(m, lmbd = 5, d = 2, transpose=False, dtype=tf.float32):
    B = bMatrix(m, lmbd, d, dtype)
    if (not transpose):
        def W(y):
            return tf.linalg.matvec(B,y)
        return tf.function(W,input_signature=(tf.TensorSpec(shape=(None,m), dtype=dtype),))
    else:
        def W(y):
            return tf.linalg.matvec(B,tf.transpose(y))
        return tf.function(W,input_signature=(tf.TensorSpec(shape=(m,None), dtype=dtype),))

def whittakerImage(shape, lmbd = 5, d = 2, dtype=tf.float32):
    assert len(shape) == 4
    fn = whittaker2D(shape[1], lmbd = lmbd, d = d, dtype=dtype)
    fn2 = whittaker2D(shape[2],lmbd = lmbd, d = d, transpose=True, dtype=dtype)
    def W(z):
        z = tf.vectorized_map(lambda z : tf.vectorized_map(fn2, tf.vectorized_map(fn,  tf.reshape(z,(z.shape[2],z.shape[0],z.shape[1])))),z)
        z = tf.transpose(z, perm=[0,3,2,1])
        return z
    return tf.function(W,input_signature=(tf.TensorSpec(shape=shape, dtype=dtype),))
smoother = whittakerImage((None,28,28,1))
layer = tf.keras.layers.Lambda(function=smoother)(layer)
5

What is your problem?