Move to numpy >=1.20 because <1.20 is difficult to build on Apple Silicon

[sradc]

model-analysis/setup.py

Line 297 in b4f2e9f

'numpy>=1.16,<2',

Numpy >=1.20 is easy to install on Apple silicon, because it's precompiled, but 1.19 fails unless openblas is set up on the system (which it will not be for most users). (See this NumPy issue.)

See also this issue in scipy where they moved away from 1.19 because it fails to build on Apple Silicon. scipy/oldest-supported-numpy#19

[gaikwadrahul8]

Hi, @sradc

Apologies for the delay and as we mentioned in TFMA setup file numpy>=1.16,<2 is not specific to particular OS so it seems like numpy==1.19 needs openblas set up to install numpy==1.19 on Apple Silicon so as you mentioned numpy==1.20 working fine so I would suggest you to go with numpy>=1.20 and as of now we haven't noticed more issues specific to numpy==1.19 w.r.t Apple Silicon

Please let me know if I'm missing something here or you need any further assistance ?

Thank you!

[sradc]

Hey @gaikwadrahul8,

Thank you for the response.

Does TFMA support Apple Silicon? I'm trying to install it alongside tensorflow-metal.

• When I run pip install tensorflow-model-analysis, it's been trying to install numpy=1.19.*, even though numpy=1.24.* is installed. (I checked, and it is trying to install tensorflow-model-analysis==0.43.0).
• When I run pip install 'tensorflow-model-analysis==0.43.0', it doesn't try to downgrade numpy, but the install fails with tfx-bsl, saying ERROR: Could not find a version that satisfies the requirement tfx-bsl (from versions: none).

(I'm a bit confused about what's going on.) I'm thinking this could be user error on my part, so will keep trying things, but letting you know on the off chance you have any information that might help.

[gaikwadrahul8]

Hi, @sradc

You can install Tensorflow with the help of tensorflow-metal plugin and use it but at the moment it seems like we do not support TFX and supported libraries like TFMA, TFDV and TFT for Apple Silicon(M1) you can check similar issue and same discussion going on our Tensorlflow Forum , In the meantime, some users have reported success with Rosetta. Other options include using a VM. We understand that neither of those is ideal.

If I'm missing something here, please let me know ? Thank you!

[sradc]

Thanks @gaikwadrahul8, appreciate the info. 👍