Segment Anything full ONNX export

This is a source code for the Export Segment Anything neural network to ONNX: the missing parts article.

Both the article and this repository aimed to show how to solve the ONNX export issue in the Segment Anything model. Official documentation explains how to export only the mask decoder model, but not the image encoder. So, even with exported ONNX model, developer still needs to use original PyTorch model and official API. It makes most benefits of ONNX useless and greatly limits the possible use of this model in production. Due to this, the Segment Anything network can be deployed only with Python and only if PyTorch and official API installed.

This repo fills this gap, here you can find how to export both SAM image encoder and mask decoder and how to use them for inference with only ONNX.

• sam_onnx_export.ipynb - Shows how to export both image encoder and masks decoder models to ONNX.
• sam_onnx_inference.ipynb - Shows how to use exported models for image segmentation.

Read the article for thorough explanation of this code.

Local execution

1. create and install a virtual environment, dependencies installation (this should works also with future dependencies, you can try removing also the version constrains in requirements_no_versions.txt)

   python -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt   

2. create your own .env file copying the template.env file, change the variables as needed

3. export the encoder and decoder onnx models

   python sam_onnx_export.py

4. try an interference example

   python sam_onnx_inference.py

5. if you want, you can download these onnx models from https://huggingface.co/aletrn/sam-quantized and use them in an onnx project like https://huggingface.co/spaces/aletrn/samgis

6. enjoy!