torch-ort 0.0.8.dev20210322

training Pytorch models with onnxruntime

training pytorch models with onnxruntime

torch-ort can be used to train pytorch models with onnxruntime backend. Before using torch-ort, one need to have a working pytorch gpu environment.

to build (you need to update version number in version.txt in order to be able to upload python whl):

(run 'pip install setuptools', if it is not already installed) 
rm dist/*
python setup.py bdist_wheel

to publish (it will ask for user name and password):

twine upload dist/*

to install:

stable:

(uninstall onnxruntime version that do not support gpu and training)
pip install torch-ort onnxruntime=1.9.0+cu111_training

nightly:

(make sure you are not in the ort repo folder - otherwise torch-ort is taken as already installed)
pip install --pre torch-ort ort-gpu-nightly-training

(eventually we are aiming at: pip install --pre torch-ort onnxruntime=1.9.0+cu111_training)

to test:

python ./ort/tests/bert_for_sequence_classification.py

to use torch-ort within PyTorch training scripts:

import onnxruntime
from torch_ort import ORTModule
model = ORTModule(model)
# normal PyTorch training script follows