Deep Learning Optimisation — Quantisation, Pruning & ONNX Export

I will optimise your trained deep learning model for deployment — applying post-training quantisation (INT8 or FP16), structured or unstructured pruning, knowledge distillation into a smaller student model, and ONNX export for cross-platform deployment — delivering a model that is significantly smaller and faster than the original with minimal accuracy loss. Trained deep learning models are frequently impractical for production deployment in their raw form — a 500MB PyTorch model that takes 200ms per inference is not suitable for a mobile app or an API serving 1000 requests per second; quantisation, pruning, and ONNX conversion together typically reduce model size by 4-8x and inference latency by 2-4x.

The optimisation covers post-training quantisation (dynamic or static), magnitude-based or structured pruning, accuracy evaluation after each compression step, knowledge distillation training into a compact student model where required, ONNX export, and a compression report showing size and latency before/after each step.