Tensorflow requirement issue

When I import alibi_detect, I get the following error:

AttributeError: module 'tensorflow_core.keras.activations' has no attribute 'swish'

which is related to this issue: https://github.com/huggingface/transformers/issues/7333

They say to update to TensorFlow 2.2, but tf 2.2 requires at least Python 3.7.

So alibi-detect is basically not compatible with Python 3.6.

[candalfigomoro]

@ascillitoe Yes.

As a side note, I had the need to use alibi-detect with tensorflow 2.1 (because of other constraints). Since the latest transformers version that works with tf 2.1 is transformers 3.1.0, I “downgraded” the alibi-detect requirement for transformers (>=4.0.0) to also accept transformers version 3.1.0. I don’t know if everything works correctly, but at least I’m able to load alibi-detect and use the TabularDrift class (which doesn’t rely on tensorflow/transformers).

[ascillitoe]

@candalfigomoro after running some tests with the tensorflow_probablity and tensorflow versions you mentioned (e.g. 0.12 and 2.0.0), it looks to me like tensorflow_probability raises quite a descriptive ImportError, so I’m not sure we would add much by raising our own errors on top of this. We think it would be undesirable to move to tensorflow>=2.4 at this point, as it is still relatively new and lots of users might still be on older tensorflow/tensorflow_probability versions.

Hopefully in the future the tensorflow version will be specified by tensorflow_probability and this issue will be circumvented entirely 🤞🏻

I’ll close this issue now, but please do feel free to reopen if you think it still isn’t resolved!