Chapter-2 Deep Learning for Coders with FastAI & Pytorch

Keypoints

• For deploying models into production we need data, obviously a trained model, API’s around Model allowing it to be accessed by the users, nice UI/UX experience if the model will be served directly as a service from the browser, good infrastructure, best coding practices etc.
• There are 4 main categories in a deep learning project before production.
  • Data Preparation
  • Labelling data
  • Model Training
  • Production
• The better way would be to allocate same time for each task.
• Underestimating the constraints and overestimating the capabilities of deep learning may lead to frustratingly poor results. So be keen on understanding what is needed.
• Conversely, overestimating the constraints and underestimating the capabilities of deep learning may mean you do not attempt a solvable problem because you talk yourself out of it. So don’t stop yourself from trying a model. Iterate your learnings.
• It’s better to iterate the project end-to-end rather than just fine-tuning the model or making some fancy GUI.
• It’s only by practicing (and failing) a lot that you will get an intuition of how to train a model.
• Start learning with the existing examples and the existing domains where deep learning is already applied and then look for more branches.
• There are many accurate models that are of no use to anyone, and many inaccurate models that are highly useful.
• A Drivetrain approach of how to use data not just to generate data but to produce actionable results is shown in the below picture:

• Below is the cool little gist which shows right from how to make our datasets to training & inference. Gist here¹

• Some of the problems to understand while building data centric products with deep learning involved:

  • Understanding and testing the behavior of a deep learning model is much more difficult than with most other code we write.
  • The neural network’s behavior emerges from the model’s attempt to match the training data, rather than being exactly defined. So this could be a disaster.
  • Out-of-domain data and domain shifts are another problem to be considered.
  • One possible approach outlined to understand the problems would be best described by below Image.

  

  Appendix