Protobuf version conflict between tf2onnx and tensorflow

[marqueurs404]

Ask a Question

Question

The release logs of version 1.15 of tf2onnx mentioned supporting TensorFlow 2.12/2.13. However, when I tried to pip install tf2onnx 1.15 and tensorflow 2.12 or 2.13, there is clearly a version conflict preventing the pip install.

Further information

[marqueurs404]

Closing as the conflict was happening with another related library not, not tf2onnx

[Zahlii]

I would like to re-open this issue, while the comment above is true (in the sense that you can manage to resolve tensorflow and tf2onnx in a clean virtualenv), it will lead to protobuf ~=3.20 to be installed.

This was only changed from tf2onnx 1.15 onwards (see #2171)

If I want to install this together with some GRPC dependencies, this will resolve to the following setup - note that this is using the most recent protobuf version, but "downgrade" tf2onnx. Will this lead to issues with tf2onnx 1.14? What was the original reason on why to include the protobuf pin?

    #7 2.483 protobuf==4.24.4
    #7 2.483 ├── grpcio-reflection==1.59.0 [requires: protobuf>=4.21.6]
    #7 2.483 ├── grpcio-tools==1.59.0 [requires: protobuf>=4.21.6,<5.0dev]
    #7 2.483 ├── onnx==1.14.1 [requires: protobuf>=3.20.2]
    #7 2.483 │   └── tf2onnx==1.14.0 [requires: onnx>=1.4.1]
    #7 2.483 ├── onnxruntime==1.16.1 [requires: protobuf]
    #7 2.483 ├── tensorboard==2.12.3 [requires: protobuf>=3.19.6]
    #7 2.483 │   └── tensorflow==2.12.1 [requires: tensorboard>=2.12,<2.13]
    #7 2.483 └── tensorflow==2.12.1 [requires: protobuf>=3.20.3, <5.0.0dev,!=4.21.5,!=4.21.4,!=4.21.3,!=4.21.2,!=4.21.1,!=4.21.0]