Confirm tensorflow_gnn has explicit dependency on tensorflow_cpu?

Q: why does tensorflow_gnn download tensorflow_cpu when tensorflow is already installed? I thought tensorflow included both CPU and GPU support these days?

pip install tensorflow_gnn produced this…

Collecting tensorflow-cpu>=2.7.0
  Downloading tensorflow_cpu-2.9.1-cp310-cp310-win_amd64.whl (256.3 MB)

but the environment already had tf installed.

(testenv) C:\Users\Julian>conda list tensorflow
# packages in environment at C:\Users\Julian\anaconda3\envs\testenv:
# Name                    Version                   Build  Channel
tensorflow                2.9.1                    pypi_0    pypi
tensorflow-addons         0.17.0                   pypi_0    pypi
tensorflow-datasets       4.5.2                    pypi_0    pypi
tensorflow-estimator      2.9.0                    pypi_0    pypi
tensorflow-gan            2.1.0                    pypi_0    pypi
tensorflow-hub            0.12.0                   pypi_0    pypi
tensorflow-io-gcs-filesystem 0.26.0                   pypi_0    pypi
tensorflow-metadata       1.8.0                    pypi_0    pypi
tensorflow-probability    0.16.0                   pypi_0    pypi
tensorflow-text           2.9.0                    pypi_0    pypi

Setup: Win 10 home 64-bit; Anaconda; Python 3.10.4; CUDA

maybe @Jan can help here

Hi, I’m currently working on updating the CI, open-source process, and packaging for tensorflow_gnn.

The short answer is that TF-GNN plans to support the same hardware as TF, but that dependency was from a highly experimental OSS release. We have since updated and corrected the required dependencies. If you want to use the updated package, you can run pip install tensorflow_gnn --pre which will install the 0.2.0.dev1 package (full 0.2.0 release coming soon).

Thanks for your feedback!

Works fine, thx. :+1:

NB there are a lot of novel dependencies there, esp. to google-cloud-* and quite a few downgrades from packages that were brought in with the base TF insall (IIRC).

Also - ignorance speaking here? - surprising to see google_api_core downloaded 10 versions 2.3.2 to 2.8.0 (as a result of Collecting google-api-core[grpc,grpcgcp]!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.0,<3.0.0dev,>=1.31.5)

Wondering whether these will also be smoothed out in the full release?


The last few releases have removed the apache-beam[gcp] dependencies in favor of apache-beam so you should see fewer of those novel dependencies. We’re still actively working on smoothing out the packaging+release, but you should have a better experience if you pip install tensorflow_gnn now.