Segmentation fault when using tfa.rotate in tf-nightly

[bw0248]

System information

• OS Platform: Linux Ubuntu 18.04.3 LTS
• TensorFlow version: tf-nightly (2.2.0-dev20200310) installed with pip
• TensorFlow-Addons version: 0.8.3 (pip) and tfa-nightly (0.9.0-dev; pip)
• Python version: 3.6.9
• Is GPU used?: yes

Describe the bug
I currently have to use tf-nightly because of a separate issue. When I use tfa.rotate I get a segmentation fault. Behavior is the same with tfa 0.8.3 and tfa-nightly.

Code to reproduce the issue

import tensorflow as tf
import tensorflow_addons as tfa

img = tf.io.read_file(IMG_PATH)
img = tf.image.decode_png(img)
print(img.shape)
img = tfa.image.rotate(img, 90)

Output:

(128, 128, 3)
Segmentation fault (core dumped)

I also tried interpolation="BILINEAR". Problem is the same.

[gabrieldemarmiesse]

Segfaults are normal when using tfa with tensorflow nightly. Tensorflow Addons is built against tensorflow 2.1.0 stable. Since the ABI is not stable from version from version to version, people get segfaults when using it with other tensorflow versions than the 2.1.0.

Can you install tensorflow 2.1.0 and tell us if the error still occurs?

[bw0248]

Ah ok, thanks for your quick reply.
I can confirm that it works without issues with tensorflow 2.1.0.

[gabrieldemarmiesse]

I understand that the fact that we're targeting TF stable is not obvious, you're not the first person raising an issue like that. So there is a problem on our side.

We need to do a better job at checking the tensorflow version and tell users that something is wrong.

Just leaving them with a segmentation is not good enough. I'll open a pull request for that. Sorry about the trouble.