Upgrade to Python 3.12

[calderjo]

we're bumping our image to Python 3.12

which required the following:
remove numpy-mkl: unfortunately we were not able to find/install a compatible version, we opted to remove it based on previous conv we had on this topic. we will instead use the default installed in colab.

remove cuml installation hack:
thankfully are able to use the pre-installed base image version without build errors.

unpinned package due to learn:
Learn is no longer dependent on this build, we can freely unpinned many packages
-seaborn, scikit-learn, matplotlib, geopandas, TPOT, shapely, tfdf, ydf, etc

remove incompatible packages:
Some of these are no longer support and cause build issues
-pydegensac, pymc3, eli5, etc

remove preinstalled package:
where applicable we removed packages that are already installed in colab base image

https://b.corp.google.com/issues/468103319

[calderjo]

placeholder: we'll add bugs in one go once we stabilize changes.

[rosbo]

Before or after merging this PR?

[calderjo]

i use to wait until we have a passing sgd (we use to test learn) to file all bugs
since we could end up needing to un/re-pinned packages

given we don't have a learn dependency anymore, i'll probably add them before merging.

i'm still looking into removing/unpinning more packages aside from these in a different PR, mostly I like waiting a bit before filing a bunch of bugs.

[djherbis]

Why do we need the massive assets added?

[calderjo]

@djherbis our "tests/test_keras_nlp.py" test requires them.
it loads the "bert_tiny_en_uncased" model and does a few things with it.

given our test are offline, we use a mock server and provide the files.

the older version uses bert_tiny_en_uncased version 2, this new one is requesting version 3.

we'll need to remove version 2 in a follow up.

---

if we want to remove the test due to size, we can, otherwise this is needed

[rosbo]

Any reason for not removing the version 2 of bert_tiny_en_uncased now that we are using version 3 in this PR?

https://github.com/Kaggle/docker-python/tree/main/tests/data/kagglehub/models/keras/bert/keras/bert_tiny_en_uncased/2

[rosbo]

In the PR description, you mention we agreed on removing mkl numpy, do you have a link to this discussion?

Mostly for future reference and my own curiosity.

[calderjo]

created a high level bug for the python 3.12 and posted it in their. i tend to avoid posting internal links here (other than bugs): https://b.corp.google.com/issues/468103319

removing mkl numpy is an option for unblocking numpy 2.0 given that we been stuck on 1.26.4 for a while.
Now the new base image (py3.12) requires we have 2.2.2.

we currently install mkl from this index, but doesn't provide anything pass 1.26.4: https://pypi.anaconda.org/intel/simple/numpy/

newer versions can be found in this repo, but unfortunately doesn't have a 2.2.2 version which this new base image really needs (as in many package break): https://software.repos.intel.com/python/pypi

[rosbo]

Thanks for the clarification and good work!

[djherbis]

@djherbis our "tests/test_keras_nlp.py" test requires them. it loads the "bert_tiny_en_uncased" model and does a few things with it.

given our test are offline, we use a mock server and provide the files.

the older version uses bert_tiny_en_uncased version 2, this new one is requesting version 3.

we'll need to remove version 2 in a follow up.

if we want to remove the test due to size, we can, otherwise this is needed

Its fine for now I guess, just generally doesn't feel great to commit a lot of extra data into the repo.
Maybe eventually our test can pull in the extra data another way.