Model development is the practical workflow for solving a deep learning problem: explore and transform available data, choose effective training algorithms, train models, and write inference code that tests unseen data. It differs from building a deep learning system, which supports this workflow and makes it more efficient and scalable. section 1.3
A production-oriented workflow typically includes:
- Translate a business problem into a machine learning problem and apply research techniques to production data. section 1.1.2
- Explore data, transform it for training, select training algorithms, train models, and validate inference against unseen data. section 1.3
- Use interactive code development, data exploration, and visualization during research and prototyping; notebook environments are common for this stage. section 10.4
- Track experimental training data with dataset management, then package stable training code, inference code, and source models for production services. section 10.4
From prototype to product
A team explores data and trains candidate models in notebooks, tracks the datasets used by experiments, packages the training and inference code once results stabilize, and integrates a satisfactory inference response with the end-user product. If inference cannot be interrupted, it selects a deployment strategy that can also support production experimentation. section 10.4
