Train and evaluate Neural Frame Rate Upscaling models using Model Gym
Introduction
Understand neural graphics and Model Gym
Set up your environment
Launch the training notebook
Fine-tune and export a quantized model
Visualize your model with Model Explorer
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Fine-tune with quantization-aware training

The training and evaluation notebooks create and inspect PyTorch checkpoints, but they do not export a VGF model. To create a model for deployment, use the QAT notebook from the dedicated NFRU examples tag. This is the notebook that includes the export step.

Quantization-aware training (QAT) simulates lower-precision inference during fine-tuning. This helps preserve model accuracy when the model is quantized to INT8 for efficient deployment.

Post-training quantization (PTQ) is usually faster to try because it calibrates an already-trained model without another training phase. QAT adds fine-tuning so the model can adapt to quantization effects, making it useful when PTQ causes unacceptable accuracy or visual-quality regressions. To explore that tradeoff and the lower-level TorchAO and ExecuTorch export workflow, see Quantize neural upscaling models with ExecuTorch .

In the notebook, you will:

  • Load the NFRU model configuration
  • Fine-tune pretrained FP32 weights using QAT INT8 mode
  • Export the quantized model to VGF
GPU memory

QAT uses additional GPU memory. If you encounter an out-of-memory error, reduce the training batch size in the notebook.

Run the QAT notebook

If Jupyter Lab is not already running, start it from the neural-graphics-model-gym-examples directory:

    

        
        
jupyter lab

In Jupyter Lab, open:

    

        
        neural-graphics-model-gym-examples/tutorials/nfru/<nfru-qat-notebook>.ipynb

Run the notebook cells in order. The notebook starts from pretrained FP32 NFRU weights, applies QAT fine-tuning, and exports the resulting INT8 model.

When the notebook completes, the exported .vgf file is available in:

    

        
        neural-graphics-model-gym-examples/tutorials/output/vgf

What you’ve accomplished and what’s next

You’ve fine-tuned the NFRU model with QAT and exported an INT8 VGF model for deployment.

Next, you’ll use Model Explorer to inspect the exported model’s graph.

Back
Next