Upgrade tensorflow to version to 2.6.x

[ophirfrish]

Upgrade the following versions:
keras - from 2.4.3 to 2.6
tensorflow - from 2.4.2 to 2.6.2
h5py - from version < 3.0 to version 3.1.0

cc @areusch@leandron@Mousius@gromero for reviews

Co-authored-by: Chris Sidebottom Chris.Sidebottom@arm.com

[lhutton1]

Thanks @ophirfrish! Just a small thing below

[lhutton1]

LGTM!

[leandron]

I'm using this PR to generate a set of temporary Docker images so that we can validate the proposed change: https://ci.tlcpack.ai/blue/organizations/jenkins/docker-images-ci%2Fdaily-docker-image-rebuild/detail/daily-docker-image-rebuild/204/pipeline

[leandron]

New attempt to generate docker images https://ci.tlcpack.ai/blue/organizations/jenkins/docker-images-ci%2Fdaily-docker-image-rebuild/detail/daily-docker-image-rebuild/208/pipeline, after fixing tlc-pack/tlcpack#87.

[leandron]

@jtuyls@anilmartha, in the process of updating TF to 2.6 we (@ophirfrish@Mousius and me) bumped into an issue that cpptests were crashing, which turns out to be some conflicting h5py (versions 2.x and 3.x) installations in some Vitis related scripts. We are proposing to unify the h5py installation to be done only in the TF script.

Can you please review and in case you disagree, propose some options to those changes?

[jtuyls]

@jtuyls@anilmartha, in the process of updating TF to 2.6 we (@ophirfrish@Mousius and me) bumped into an issue that cpptests were crashing, which turns out to be some conflicting h5py (versions 2.x and 3.x) installations in some Vitis related scripts. We are proposing to unify the h5py installation to be done only in the TF script.

Can you please review and in case you disagree, propose some options to those changes?

@leandron Could we keep the h5py being installed separately in the vitis_ai_core.sh script: https://github.com/apache/tvm/pull/10084/files#diff-7773869aa5e34502bbaa476fe3339c2e3956199b0be759084feee12fe8bcdfe1L37?
In our Dockerfile we don't use the ubuntu_install_tensorflow.sh script so h5py won't be installed otherwise. This shouldn't affect the CI as this script is not used there.
The changes in vitis_ai_packages_ci script are fine for us.

[leandron]

@leandron Could we keep the h5py being installed separately in the vitis_ai_core.sh ?

Sure, and with regards to versions, do you want to keep it as is, with version 2.x or update to be coherent with the one installed in TVM, which means h5py 3.x ?

[jtuyls]

@leandron Could we keep the h5py being installed separately in the vitis_ai_core.sh ?

Sure, and with regards to versions, do you want to keep it as is, with version 2.x or update to be coherent with the one installed in TVM, which means h5py 3.x ?

Yeah, I think it's best to keep it as is (2.10.0) for now and I can work on updating it later to 3.x. I think 3.x works but I would like to completely verify that first.

[leandron]

After the last changes, I'm rebuilding the Docker images again to re-run tests - https://ci.tlcpack.ai/blue/organizations/jenkins/docker-images-ci%2Fdaily-docker-image-rebuild/detail/daily-docker-image-rebuild/230/pipeline/

[leandron]

With the latest version of this PR, rebased on top of main at bd14a4d36e0d364ef9bd34b2ee96cc09ce64d4b3, and with #10406 applied, I generated a new set of Docker images for testing at https://ci.tlcpack.ai/job/docker-images-ci/job/daily-docker-image-rebuild/238/, with label tlcpackstaging/ci_****:20220301-163838-977bdd3d2.

[leandron]

This is validated in https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/ci-docker-staging/232/pipeline/, covering up to 4797850.

@ophirfrish can you trigger CI again, so that I can merge this?

[leandron]

LGTM, thanks @ophirfrish!

[leandron]

This is merged now, as it is running in production CI starting on tlcpack/ci-gpu:v0.82 and tlcpack/ci-cpu:v0.82 in #10654.

Thanks @ophirfrish@lhutton1 and @jtuyls!