Deployment for Computer Vision Engineers
Deployment is the work of moving a trained vision model into a product that runs fast and reliably. It covers exporting to formats like ONNX, shrinking models through quantisation and pruning, serving with batching on a GPU or running on a phone or edge device, and monitoring accuracy once real-world images start arriving.
Stage: Generative and Production · Part of the Computer Vision Engineer roadmap · Last reviewed 19 November 2026
What to learn in Deployment
- Export a PyTorch model to ONNX and run it with ONNX Runtime
- Apply post-training quantisation and measure accuracy and latency changes
- Profile inference time and find the preprocessing step that dominates
- Serve a model behind an HTTP API with request batching
- Monitor for data drift when production images differ from training data
Deployment interview questions
What is quantisation and what is the trade-off?
Quantisation stores weights and sometimes activations in lower precision, such as 8-bit integers instead of 32-bit floats. This shrinks the model and speeds inference on supported hardware, usually at a small accuracy cost that you should measure on your own validation data.
Why can a model perform worse in production than in testing?
Real images often differ from the training set in lighting, camera, resolution or content, a problem called distribution shift. Preprocessing mismatches between training and serving code also cause silent errors. Monitoring inputs and sampling predictions for review helps catch this early.
Where Deployment fits in the Computer Vision Engineer roadmap
Deployment belongs to the Generative and Production stage (advanced level), which takes about 38 hours. Create and ship.
Other units in this stage
Back to the full Computer Vision Engineer roadmap