Train or re-train ONNX Model Zoo Models and modifying their weights

[Izan Catalán]

Ask a Question

Can an ONNX model from this Github repo (ONNX model ZOO) be trained and modify its weights during the process? In addition, is it possible to re-train an existing onnx model?

Question

I would like to know if ONNX models (from ONNX model Zoo) can be trained or re-trained from an onnx file. I only saw notebooks in some specific models like Mobilenet, Resnet or Vgg in Onnx Model Zoo Repo.

These notebooks, however, only show a Python code to train those models (https://github.com/onnx/models/blob/main/vision/classification/vgg/train_vgg.ipynb) but they don't show how to export them to onnx files or if it is possible to import an model from an existing onnx file. They just show it is due to an internal converter which is to be released.

I would like to re-train some models, to change some of their weights and, later, see the results in terms of accuracy. I wonder if it is possible to use Onnx Runtime or any of those notebooks from the Git Repo.

I also wonder if those frameworks are valid for those models which don't have one of their own, for example, Densenet, Shufflenet or Alexnet. How can I train those models? Is there any other way? I need help finding a further guide to do it.

Finally, I would like to export my train models to onnx, and, therefore, to know the tool or interval converter or any other method to do it if it is possible or if anyone knows how.

Thanks.

Izan.