{lib}[fosscuda/2019b] TensorFlow v2.2.0 w/ Python 3.7.4

[Flamefire]

(created using eb --new-pr)

Can be used as a starting point for fosscuda/2020a once CUDA is out. Not sure if another TF version is acceptable for fosscuda/2019b but it has multiple already and it seems fosscuda/2020a will take a while...

[Flamefire]

Test report by @Flamefire
SUCCESS
Build succeeded for 1 out of 1 (1 easyconfigs in this PR)
taurusml30 - Linux RHEL 7.6, POWER, 8335-GTX, Python 2.7.5
See https://gist.github.com/4b97825babefb2309a358ac8b46c082d for a full test report.

[Flamefire]

Test report by @Flamefire
SUCCESS
Build succeeded for 2 out of 2 (2 easyconfigs in this PR)
taurusa6 - Linux centos linux 7.7.1908, x86_64, Intel(R) Xeon(R) CPU E5-2603 v4 @ 1.70GHz, Python 2.7.5
See https://gist.github.com/90fd5878352f360eeeeef71c3706dc8b for a full test report.

[Micket]

Test report by @Micket
SUCCESS
Build succeeded for 3 out of 3 (2 easyconfigs in this PR)
vera-c1 - Linux centos linux 7.7.1908, Intel Xeon Processor (Skylake), Python 2.7.5
See https://gist.github.com/7392c8578c72373d0dd82c5246117d3a for a full test report.

[JackPerdue]

Might I suggest you provide an example of how to provide cuda_compute_capabilites.

1. at present it uses a default value
2. it isn't clear at all that the TensorFlow extension is using the tensorflow.py easyblock
3. it isn't clear at all (without a long search of readthedocs) on how to add it

Here, I just did a test by adding the setting to the TensorFlow extension block (at the bottom):
'cuda_compute_capabilities': ['3.5', '3.7'], # for Tesla K20 (ada) and K80 (terra)

For your easyconfig, I'd recommend the usual EB default of something like:
'cuda_compute_capabilities': ['3.0', '3.2', '3.5', '3.7', '5.0', '5.2', '6.0', '6.1', '7.0']

Now then... if I had my druthers, I'd explicitly indicate the easyblock it is using. Of course, I prefer that in all easyconfigs. Alas, the EB team disagrees and won't accept easyconfigs that explicitly list the easyblock used instead some "magic" that always infers the easyblock from whatever is being built. But maybe they will it in things like PythonBundle (I just assumed it was using PythonPackage for extensions).

In any case, providing a cuda_compute_capabilites would be a good idea.

[Flamefire]

@JackPerdue

1. Using a default value is fine
2. It is clear as no EasyBlock is explicitely specified meaning "tensorflow" uses the EasyBlock "TensorFlow"
3. eb --help

For your easyconfig, I'd recommend the usual EB default of something like:
'cuda_compute_capabilities': ['3.0', '3.2', '3.5', '3.7', '5.0', '5.2', '6.0', '6.1', '7.0']

That would be extremely wasteful. Each site has to configure this to their own hardware just like any other site-specific config variable. This is not for the EasyConfig to decide

Alas, the EB team disagrees and won't accept easyconfigs that explicitly list the easyblock used instead some "magic" that always infers the easyblock from whatever is being built.

What's wrong with that? If nothing is specified the name is used. Couldn't get simpler

In any case, providing a cuda_compute_capabilites would be a good idea.

If you don't overwrite it, a default is used by TF which isn't that bad. Not sure what it was exactly but IIRC it is something like 3.5, 7.0.

[Micket]

Regarding the cuda_compute_capabilities, it should print a verbose warning to the terminal, as well as the log with instructions on what one should do if one hasn't set the cuda_compute_capabilities.
Though, there was a confusing typo in the warning, which will be fixed in next release
easybuilders/easybuild-easyblocks#2057

[Micket]

Test report by @Micket
SUCCESS
Build succeeded for 3 out of 3 (2 easyconfigs in this PR)
hebbe-c1 - Linux centos linux 7.7.1908, Intel Core Processor (Haswell, no TSX), Python 2.7.5
See https://gist.github.com/63712b8fd24eb871ac805bd9e26af3c8 for a full test report.

[Micket]

lgtm

[Micket]

Going in, thanks @Flamefire!