Nix build

[sheepforce]

Currently the Haskell bindings are a little bit difficult to access with Nix. This is some work to build the tensorflow bindings with Nix and provide an overlay via a Nix flake that allows to use the Haskell packages of this repo via an overlay in other Nix projects.

• [ ] Build all modules
  • [x] tensorflow
  • [ ] tensorflow-core-ops
  • [ ] tensorflow-logging
  • [x] tensorflow-opgen
  • [ ] tensorflow-ops
  • [x] tensorflow-proto
  • [x] tensorflow-records
  • [x] tensorflow-records-conduit
  • [x] tensorflow-test
• [ ] shell.nix
• [ ] default.nix

In the tensorflow-core-ops build I am getting:

tensorflow-core-ops> [1 of 1] Compiling TensorFlow.GenOps.Core ( dist/build/global-autogen/TensorFlow/GenOps/Core.hs, dist/build/TensorFlow/GenOps/Core.o, dist/build/TensorFlow/GenOps/Core.dyn_o )
tensorflow-core-ops> dist/build/global-autogen/TensorFlow/GenOps/Core.hs:69046:13: error:
tensorflow-core-ops>     Variable not in scope: num_side_inputs :: Int64
tensorflow-core-ops>       |
tensorflow-core-ops> 69046 |     pureOp [num_side_inputs] $ do

and I assume this is caused by having tensorflow-2.9.1 in nixpkgs but using some 2.3ish version as the submodule from which c_api.h is imported. Using tensorflow-2.9.1 as a submodule makes things worse, unfortunately.

[google-cla[bot]]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[mikesperber]

@sheepforce While I support this work, note that the Tensorflow Haskell bindings have had regular Nix packages for a long time.

The current version has a blocking issue though:

https://github.com/tensorflow/haskell/issues/274

[sheepforce]

While I support this work, note that the Tensorflow Haskell bindings have had regular Nix packages for a long time.

Yes, I am aware :) However, it is marked broken since a long time and does not actually build. I saw your PR attempting to fix it https://github.com/NixOS/nixpkgs/pull/119411 The nixpkgs pulls the Hackage version from 2019 which misses all the fixes that have made it to this repo. I was hoping to get a more recent version directly from this repo. Having the Nix expressions here also gives the option to get a development environment with Nix.

The current version has a blocking issue though. https://github.com/tensorflow/haskell/issues/274

Could this also be caused by the Tensorflow version mismatch between nixpkgs and the tensorflow submodule here with c_api.h? I would love to help but I am not so fluent with C and how to call it via the Haskell FFI.

[mikesperber]

However, it is marked broken since a long time and does not actually build. I saw your PR attempting to fix it NixOS/nixpkgs#119411 The nixpkgs pulls the Hackage version from 2019

That's incorrect. It refers to github revision 568c9b6f03e5d66a25685a776386e2ff50b61aa9, which is pretty close to HEAD. Updating to HEAD is "just" a matter of resolving the above PR.

Could this also be caused by the Tensorflow version mismatch between nixpkgs and the tensorflow submodule here with c_api.h? I would love to help but I am not so fluent with C and how to call it via the Haskell FFI.

It's not out of the question, and that's what I looked for, too. However in testing I was pretty careful making sure the versions match.

[sheepforce]

That's incorrect. It refers to github revision https://github.com/tensorflow/haskell/commit/568c9b6f03e5d66a25685a776386e2ff50b61aa9, which is pretty close to HEAD.

You are completely right, saw it in the PR. I've got confused by the version in nixpkgs being labelled 0.1.0.0

[Minnozz]

@sheepforce The num_side_inputs error seems to be a bug in the either the code generation or the protobuf definition of the op. You can add it to the blacklist to exclude it from generating bindings.