pin numpy version in run requirements for meta.yml

[pstjohn]

Currently, a basic conda-forge environment with python and tensorflow 2.4.1 will pull numpy 1.20.* by default, which isn't compatible.

Would be great if you could add numpy=1.19.* to the run requirements for the meta.yml

[izahn]

I'm not sure I understand.

Currently, a basic conda-forge environment with python and tensorflow 2.4.1 will pull numpy 1.20.* by default, which isn't compatible.

Isn't compatible with what?

[pstjohn]

with tensorflow 2.4.1. The TF team tends to pin numpy versions in their pip releases, and the current meta.yml builds against numpy 1.19.*, but doesn't specify that as a run requirement.

As anecdotal evidence, in the linked PR from my repo, I got a bunch of numpy warnings / errors without pinning numpy, but were fixed when pinning numpy to 1.19.*

[izahn]

Good point, maybe this does need a numpy matrix. I'm never really sure about that!

[jakirkham]

We don't build with NumPy matrices any more (we got rid of these a while back). We just build against the oldest supported NumPy version and let it float above that at runtime. Though this only matters for libraries that link against NumPy

[pstjohn]

Floating above 1.19.* might be problematic in this case, since using numpy 1.20 at runtime with tensorflow has some known issues tensorflow/tensorflow#47691

[jakirkham]

Again that comment only pertains to linking NumPy. Idk how TensorFlow uses NumPy so can't say whether that is relevant.

[pstjohn]

Ah, ok. I suppose to be concrete, I propose changing

tensorflow-feedstock/recipe/meta.yaml

Line 156 in f18fb00

- numpy >=1.19

to numpy =1.19

[miketrumpis]

+1 on this proposal. I just ran into this issue last week. Conda allowed numpy 1.20 to install alongside tensorflow 2.4.1.

Result was this runtime issue: tensorflow/models#9706 (similar to those referenced in tensorflow/tensorflow#47691)

[pstjohn]

Might be trickier now that they've released a 2.6.0 version.
Although I've been getting similar errors with that version, so I had to revert back to pip wheels for some of my CI builds

[h-vetinari]

The most recent tensorflow builds have all pinned numpy to 1.19 (see also #151 and the linked discussions about numpy compat). Thankfully, TF 2.7. should alleviate this.

In the meantime, is this still an issue for any conda-forge builds in the wild?

[miketrumpis]

fyi I just had another opportunity to set up a conda-forge tensorflow environment, which now results in a non-conflicting numpy version. Issue resolved here.

numpy                     1.19.5           py38h9894fe3_2    conda-forge
tensorflow                2.6.2           cuda112py38ha230376_1    conda-forge

[ngam]

Now the pinning seems to be >=1.20