Feature Request: GPUOptions for Go binding

[mattn]

Current implementation of Go binding can not specify options.

GPUOptions struct is in internal package. And go generate doesn't work for protobuf directory. So we can't specify GPUOptions for NewSession.

[tensorflowbutler]

Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks.
Have I written custom code
OS Platform and Distribution
TensorFlow installed from
TensorFlow version
Bazel version
CUDA/cuDNN version
GPU model and memory
Exact command to reproduce
Mobile device

[frreiss]

This problem appears to be broader than just specifying GPUOptions for TensorFlow sessions. There is no native Go API for passing any options when creating a session. The user must create the binary representation of a ConfigProto protocol buffer outside of the TensorFlow Go API.

See, for example, the test case TestSessionConfig in session_test.go:

func TestSessionConfig(t *testing.T) {
	// Exercise SessionOptions.
	// Arguably, a better API would be for SessionOptions.Config to be the
	// type generated by the protocol buffer compiler. But for now, the
	// tensorflow package continues to be independent of protocol buffers
	// and this test exercises the option since the implementation has a
	// nuanced conversion to C types.
	//
	// Till then, the []byte form of Config here was generated using a toy
	// tensorflow Python program:
	/*
	 import tensorflow
	 c = tensorflow.ConfigProto()
	 c.intra_op_parallelism_threads = 1
	 print c.SerializeToString()
	*/
	graph := NewGraph()
	c, err := Const(graph, "Const", int32(14))
	if err != nil {
		t.Fatal(err)
	}
	opts := SessionOptions{Config: []byte("(\x01")}
[...]

Would the TensorFlow maintainers accept a non-Google contribution that added the ability to specify session options using pure Go code? This functionality would require generating instances of the ConfigProto protocol buffer defined in config.proto. I can see two ways to generate these protocol buffers: Either add a build target to generate Go bindings for the files in tensorflow/core/protobuf; or add Go wrappers for the generated C++ code in tensorflow/core/protobuf/config.pb.h

[ymodak]

@asimshankar Can you please take a look? Thanks!

[alextp]

@frreiss we'll definitely accept a Go-only API to configure tensorflow.

[frreiss]

I am looking into this.

[frreiss]

Created PR #26682 with a Go API to create ConfigOptions protocol buffer messages. I made the changes as narrow in scope as I could.

[mattn]

Closed by #26682

@frreiss Thank you

[mattn]

#26682 is reverted 6e9cb40#diff-dcd28ad951bd17e9c512d3b564640bab

[mattn]

Yes please.

[imrahul361]

May I know how to reproduce the issue?

[mattn]

Sorry, I don't know why the changes was reverted.

#26682 (comment)

[koliasa]

Thanks for the clarification! This does seem like a limitation of the current Go binding implementation. Since the GPUOptions struct is in the internal package, we can't directly pass it when creating a session with NewSession.

Regarding go generate and protobuf: this is also an important issue, as we need access to the appropriate configurations in protobuf to properly specify GPU options.

One potential solution might be to make GPUOptions accessible through a public API or consider adding custom options to NewSession via another configuration mechanism (e.g., a configuration file or a specialized getter method).

[mihaimaruseac]

Locking this due to so many spammers. You should really be proud of yourselves.

As a reminder: in open source, you don't ask to be assigned some work. You deliver said work.