TensorFlow 2.0: Function with @tf.function decorator doesn't take numpy functions

Question

I am writing a function to implement a model in TensorFlow 2.0. It takes image_batch (a batch of image data in numpy RGB format) and performs some specific data augmentation task that I need. The line that is causing me problems is:

@tf.function
def augment_data(image_batch, labels):
    import numpy as np
    from tensorflow.image import flip_left_right

    image_batch = np.append(image_batch, flip_left_right(image_batch), axis=0)

    [ ... ]

numpy's .append() function doesn't work anymore when I put the @tf.function decorator on top of it. It returns:

ValueError: zero-dimensional arrays cannot be concatenated

When I use the np.append() command outside of the function, or without the @tf.function on top, the code runs without problems.

Is this normal? Am I forced to remove the decorator to make it work? Or is this a bug, due to the fact that TensorFlow 2.0 is still a beta version? In that case, how can I solve this?

Sharky · Accepted Answer · 2019-04-14 20:27:19Z

3

Just wrap numpy ops into tf.py_function

def append(image_batch, tf_func):
    return np.append(image_batch, tf_func, axis=0)

@tf.function
def augment_data(image_batch):
    image = tf.py_function(append, inp=[image_batch, tf.image.flip_left_right(image_batch)], Tout=[tf.float32])
    return image

answered Apr 14, 2019 at 20:27

Sharky

4,5412 gold badges21 silver badges27 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

Leevo Over a year ago

Hi, I have tried your code. Only one thing returned error: you must write np.append instead of just append. Apart from that eveything works, thank you for the reply.

Sharky Over a year ago

In this case append is a user defined function, but it's can use np.append as well.

Leevo Over a year ago

Shall I use tf.py_function even for more "basic" ones, like int() ? I'm now forcing a number to be integer, and I got: TypeError: 'int' object is not callable. Is it the same problem? Thanks

Sharky Over a year ago

I think with regard to int() it's another problem.

Collectives™ on Stack Overflow

TensorFlow 2.0: Function with @tf.function decorator doesn't take numpy functions

1 Answer 1

4 Comments

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Related