# Image Trainer

> Train your own image classifier in the browser, with no code and no upload. Free; no sign-in.

URL: https://image-trainer.skillsafe.ai/

## What it does
- Make 2 to 10 classes and name them.
- Add samples per class: hold a button to record webcam frames, take single frames, or upload/drop photos (up to 1,000 per class).
- Each sample is a 224 x 224 centre crop. An image model running in the browser (onnxruntime-web, WebAssembly) turns it into a feature vector:
  - Accurate: CLIP ViT-B/32 image encoder, int8, 89 MB, 512 numbers (MIT licence).
  - Light: ResNet-50 ImageNet, int8, 26 MB, 1,000 numbers (Apache-2.0).
- A small classifier is trained on those vectors in the page: linear softmax regression or one hidden layer of 64 (ReLU), Adam, L2 weight decay, optional class balancing.
- A stratified share of every class (default 20%, classes with 4+ samples) is held out. The page shows held-out accuracy, a confusion matrix (rows = true class, columns = predicted) with per-class recall, the most-confused pair and the held-out samples it got wrong. By default it then retrains on every sample. After a retrain it also shows the previous training's held-out accuracy, so you can see whether new samples helped.
- Try it: live webcam predictions with confidence bars, or classify a batch of photos and download a CSV (file, predicted, confidence, one probability per class).
- Projects are saved in the browser (IndexedDB) and can be saved to / opened from a .json file.
- Export: a model JSON (format "image-trainer-head/1": labels, the image model's file, SHA-256, resize and normalisation recipe, standardisation, layer weights, held-out evaluation) plus JavaScript and Python usage code.

## Free example
"See an example" opens 64 shapes drawn by code (Circle, Square, Triangle, Star) with their image features bundled, so training and the confusion matrix need no model download. Switching to the light model shows a model that mixes some shapes up.

## Privacy
Samples, training and predictions stay on the device. Only the image model file is downloaded, once, from SkillSafe's model host, and kept in the browser cache.