Training model with ONNX runtime

[tiberiusferreira]

Hello, thanks for building the ONNX runtime!

I'm trying to create a simple model and train it using onnx runtime, but I can't find an equivalent of sess = onnxruntime.InferenceSession(path_save) for training, is it something planned? I saw mentions of plans on making the training more seamless with Pytorch, which is great, but what I personally like about ONNX is how framework agnostic it is.

The tutorials seems to be Pytorch focused, but I'm more interested in using a "standalone" ONNX model, such as shown here with the appropriate training information embedded in it, is this use case going to be supported?

Nevertheless I tried the getting started tutorial, but it used
from onnxruntime.training import ORTTrainer, optim which raised a ModuleNotFoundError: in both the latest stable and nightly build installed using pip3 install -i https://test.pypi.org/simple/ ort-nightly.

[tianleiwu]

@edgchen1, could you help on this topic?

[tiberiusferreira]

Thanks, I ran into a problem. 😓 I tried to build with "./build.sh --enable_training --build_wheel" but then it errors

[ 97%] Building CXX object CMakeFiles/onnxruntime_pybind11_state.dir/Users/tiberiodarferreira/onnx-git/onnxruntime/python/onnxruntime_pybind_mlvalue.cc.o
In file included from /Users/tiberiodarferreira/onnx-git/onnxruntime/python/onnxruntime_pybind_mlvalue.cc:4:
In file included from /Users/tiberiodarferreira/onnx-git/onnxruntime/python/onnxruntime_pybind_mlvalue.h:7:
In file included from /Users/tiberiodarferreira/onnx-git/build/Linux/Debug/pybind11/src/pybind11/include/pybind11/iostream.h:12:
In file included from /Users/tiberiodarferreira/onnx-git/build/Linux/Debug/pybind11/src/pybind11/include/pybind11/pybind11.h:44:
In file included from /Users/tiberiodarferreira/onnx-git/build/Linux/Debug/pybind11/src/pybind11/include/pybind11/attr.h:13:
/Users/tiberiodarferreira/onnx-git/build/Linux/Debug/pybind11/src/pybind11/include/pybind11/cast.h:2011:26: error: loop variable 'a' is
      always a copy because the range of type 'detail::args_proxy' does not return a reference [-Werror,-Wrange-loop-analysis]
        for (const auto &a : ap)

I'm on MacOS.

[tiberiusferreira]

Seems to be same as google/filament#3066 introduced by clang 12

[edgchen1]

Hm, unfortunately I don't have experience with building on Mac. It looks like our Mac CI build uses clang 11 (AppleClang 11.0.0.11000033).
You can file an issue - perhaps someone else will be able to help more.
Is using a different version of clang an option? Or building within a Docker container?

[tiberiusferreira]

I'll try downgrading clang, thanks!

[tiberiusferreira]

It worked, thanks for the help!

[SherlockNoMad]

We don't have plan to have a mirrored interface like onnxruntime.TrainingSession(). We usually recommend defining the model in pytorch nn.module and use ORTTrainer to interact with it. But if you already have a .onnx model together with loss function exported, you can also use ORTTrainer to load it.
see https://github.com/microsoft/onnxruntime/blob/master/orttraining/orttraining/python/training/orttrainer.py#L68

[tiberiusferreira]

Thank you. Just to make sure I understood, it will be possible to user whatever tool to export a model.onnx file and load it up on the ORTTrainer and train it without having to go through Pytorch or any other framework, right?

Some context: having a .onnx file containing everything needed for training and inference is very good for deploying to devices and doing on-device learning, but if one has to bundle Pytorch or something similar to do the training, it quickly becomes unfeasible.

So my vision was being able to have ONNX runtimes in general not only do inference, but also training.

[SherlockNoMad]

We share the same vision!

Could you please share a bit more for on-device learning? What's the scenario, what's the underlaying device?

[PTMMachineLearning]

@tiberiusferreira Hello I am on a Facebook Pytorch team reaching out to understand the status and if everything is now working?

[tiberiusferreira]

Yep, it is. Thank you for reaching out.

[PTMMachineLearning]

@tiberiusferreira Thank you for the update. If there is interest in accelerated ML workflow, training or inference, on MacOS I would be interested in knowing more about the project. Thanks again.