Add support for loading models in `ONNX` format

Pale-Blue-Dot-97 / Minerva

Minerva project includes the minerva package that aids in the fitting and testing of neural network models. Includes pre and post-processing of land cover data. Designed for use with torchgeo datasets.

MIT License

20 stars 1 forks source link

Add support for loading models in `ONNX` format #67

Closed Pale-Blue-Dot-97 closed 1 year ago

Pale-Blue-Dot-97 commented 1 year ago

Issue

Support for saving models to ONNX format has been added to minerva but no support is currently provided for loading ONNX format models.

Solution

[x] Add a new function to Trainer to load a model directly from ONNX format given a path to the file. ~- [ ] Add support to Trainer.make_model to load model weights from ONNX format and graft them onto the model object.~

Pale-Blue-Dot-97 commented 1 year ago

Doing some reading on this, turns out that pytorch does not officially support loading onnx models despite the ability to export them. It was a long running feature request before being recently rejected. However, there are third-party implementations such as onnx2torch which I'm going to look into using.