sxsphinxstack

Projects / An image classifier for your own photos

An image classifier for your own photos

Level 3: a month or more, real users. Use with: train-a-model, build-web-app. Resume line when done: Built {project_name}, an image classifier trained on {n} of my own photos with a live demo — at {url}

Brief

Train an image classifier on categories from your own camera roll: your dog vs other dogs, sketchbook pages vs lecture notes, the climbing gym vs the skatepark. Curate a few hundred of your own photos per class, fine-tune a small pretrained model, and evaluate it on photos it could not have memorized. The demo is a live page where anyone can upload an image and see the prediction, the confidence, and a gallery of the model's real mistakes.

Personalize

Milestones

  1. Define the classes and audit your camera roll: enough photos per class, exported at a consistent size, consent noted for any photo showing another person.
  2. Build the dataset with the split done first: hold out whole events or months as the test set, and freeze it.
  3. Label a quick sample of the test set twice, a week apart. If you disagree with yourself, tighten the class definitions and note it.
  4. Compute the majority-class baseline, then build the cheap-rule baseline and score it on the test set.
  5. Fine-tune a small pretrained model on the training set only. Track training without peeking at test.
  6. Evaluate once on the frozen test set: accuracy per class next to both baselines, plus a confusion matrix.
  7. Collect the failures: every misclassified test photo goes into a gallery with the predicted and true labels.
  8. Stress-test with photos from outside the dataset entirely (a friend's phone, with permission) and record how it does on truly foreign images.
  9. Ship the live demo: upload an image, get prediction and confidence, with the failure gallery and evaluation table on the same page.
  10. Write the limitations section: your lighting, your camera, your places, and what that means for anyone else's photos.

Done means