ViT-ADE20K-Segmentation

ViT Semantic Segmentation on ADE20K

A Vision Transformer built from first principles (patch embedding, multi-head self-attention, MLP, transformer encoder blocks) adapted for semantic segmentation on the ADE20K dataset (150 classes), with training, evaluation, benchmarking against published ViT segmentation models, and an evolutionary hyperparameter search.

Contents

• vit-ade20k-segmentation.ipynb — the full notebook
• config.json — model/training/search configuration
• checkpoints/vit_seg_best.pt — best model checkpoint (created after training)

Configuration

See config.json for the exact data, model, training, and evolutionary-search settings used to produce the results below.

Results

Training curves

Train/val loss, validation mIoU, validation pixel & mean-class accuracy, and the learning-rate schedule.

Inference samples

Image / ground truth / prediction, side by side, for a few validation images.

How to reproduce

1. Install dependencies (see the first cell of the notebook).
2. Run all cells top to bottom.
3. Images (training_curves.png, inference_samples.png, benchmark_comparison.png, evolution_fitness.png) and config.json are written to the working directory as you go.

Notes

• Published benchmark numbers are commonly-cited approximations from the original papers — verify exact figures before citing them elsewhere.
• Demo defaults (small embed_dim/depth, few epochs) are set to run quickly; scale up for real training.