code-tour.md

Code Tour: Run First, Then Read

简体中文 | Documentation

Trace one image from a manifest row to a prediction instead of reading src/ alphabetically.

Step	Run	Read	Focus
1	example 01	data/dataset.py::map_mask	raw values to target
2	example 02	data/transforms.py	paired geometry
3	example 03	data/dataset.py	sample to batch
4	example 04	models/unet.py, training/losses.py	logits and loss
5	training dry run	training/train.py, trainer.py	production assembly
6	example 05	inference/predictor.py	checkpoint to mask

manifest.py validates and records pairs. Dataset opens one row, applies paired transforms, and maps labels. U-Net only converts image tensors to same-resolution logits. train.py assembles components; Trainer runs epochs and saves best/last.

uv run segment train --config configs/learning_minimal.yaml --dry-run

Evaluation metrics count pixels, evaluate writes reports and selected visuals, and Predictor handles unlabeled images. On a first read, skip strict config helpers, Oxford audit details, RNG persistence, argparse declarations, and pixel- level rendering internals. Return when a concrete question requires them.

Use configuration flow for overrides and how it works for formulas and design reasons.