Description
A major issue we have with Addons is that it is compatible with PyPi TensorFlow, but has ABI incompatibilities with TF that was built from source or installed with conda depending on the compiler and settings:
#676
#753
#574
The ultimate fix for this will be the ABI stable kernel and ops:
tensorflow/community#133
However, as a mitigation it may be worthwhile to lazy load custom-ops when they are used:
https://github.com/tensorflow/addons/blob/master/tensorflow_addons/activations/gelu.py#L23
At the moment all custom-ops are loaded during init. Seeing as >80% of addons does not use a custom-op it is likely that many users are experiencing this issue even when they don't plan on using a custom op. This is certainly a hacky "fix" and doesn't solve the issue but I thought it was worth discussing as I feel this is a major issue regarding adoption.