Export Swin2SR for Arm VGF

Export the pretrained model

You don’t need to train Swin2SR. The exporter downloads the pretrained ×2 checkpoint — the saved model weights — and converts it into an ExecuTorch .pte program. The pinned checkpoint revision keeps the model weights consistent between runs.

Run the exporter from your ExecuTorch directory:

    

        
        
python examples/arm/super_resolution_example_vgf/model_export/export_super_resolution.py \
  --model-name swin2sr \
  --checkpoint caidas/swin2SR-classical-sr-x2-64 \
  --checkpoint-revision cee1c923c6a37361c6e5650b65dcf4be821e5d52 \
  --input-height 64 \
  --input-width 64 \
  --quantization-mode none \
  --output-path swin2sr-work/swin2sr.pte

    

The first run downloads the model weights. --quantization-mode none keeps floating-point calculations, so you don’t need calibration images.

The 64 dimensions fix the input size for this export. The model produces a 128 × 128 image because the checkpoint upscales by two.

Keep the program and metadata together

After export finishes, check the two files that the runner needs:

    

        
        
ls -lh swin2sr-work/swin2sr.pte swin2sr-work/swin2sr.json

    

swin2sr.pte contains the executable model, including its Vulkan Graph Format (VGF) graphs. swin2sr.json tells the image helper how to read the input and reconstruct the output. Keep both files in the same directory with the same base name.

The exporter also saves swin2sr_delegation.txt. It records which operations run through VGF and which remain in ExecuTorch. You don’t need to change this report to run the example.

What you’ve accomplished and what’s next

You have a floating-point Swin2SR program configured for one 64 × 64 image with red, green, and blue (RGB) color channels.

Next, you’ll build the host runner and use it to upscale your image.