Chapter Summary

Chapter Summary

Key Points

• 1.

  Pre-trained models are the starting point. Use torchvision or HuggingFace models. Feature extraction for tiny datasets, fine-tuning for medium, LoRA for large models with limited compute.

• 2.

  PnP denoisers solve inverse problems without task-specific training. Pre-trained DRUNet acts as a universal prior. Alternate between data fidelity and denoising. Controls noise level via the noise map input.

• 3.

  Domain adaptation bridges simulation and reality. Train on cheap simulated data, adapt to real measurements. Domain-adversarial training learns domain-invariant features.

• 4.

  Export models for deployment. TorchScript for PyTorch inference, ONNX for cross-framework, TensorRT for maximum GPU speed. Quantisation (INT8) reduces latency by 2-4x.

• 5.

  Always use the correct preprocessing. Pre-trained models expect specific normalisation. Wrong preprocessing silently degrades performance.