A Deep Learning Project Checklist

Questions to ask

Is deep learning the right tool for this data and size? Is there a simple baseline? Are data, labels and splits checked, with no leakage? Is the initial loss sensible and can the model overfit one batch? Are training and validation curves logged? Are learning rate, schedule and regularisation tuned on validation data? Is the best checkpoint kept by validation score? Is inference done in eval mode without gradients? Are preprocessing, versions and metrics saved with the weights? Is there a monitoring plan?

The checklist

Use it in reviews.

[ ] right tool? baseline compared (e.g. gradient boosting)
[ ] data + labels inspected; splits without leakage
[ ] shapes asserted; initial loss sensible
[ ] overfits one batch
[ ] train/val curves logged every epoch
[ ] learning rate + schedule + regularisation tuned on validation
[ ] best checkpoint saved (early stopping)
[ ] inference: model.eval() + no_grad
[ ] weights saved with preprocessing, versions, labels, metrics
[ ] monitoring + retraining plan

Prefer pretrained models

For images, text and audio, fine-tuning a pretrained model is usually faster, cheaper and better than training from scratch.

Quick check: Which item belongs on a deep learning checklist?

• Run inference in training mode
• Skip validation curves
• Verify the model can overfit a single batch
• Tune on the test set