A 4-week course project that fine-tunes published FER models on RAF-DB and ships a Gradio demo.
| Role | Choice | Justification (from research report) |
|---|---|---|
| Primary dataset | RAF-DB | 29,672 in-the-wild images, 7 classes, every model in the report reports on it (report.md:2186) |
| Backup dataset | FER-2013 | Free Kaggle download, no EULA, same 7 classes (report.md:1906) |
| Baseline model | DAN (ResNet-18) | 89.70% on RAF-DB, public weights, runs on a single T4 (report.md:252) |
| Stretch model | POSTER++ | 92.21% on RAF-DB SOTA, public weights, dual-stream CNN+Transformer (report.md:905) |
Full SOTA report: research/facial-emotion-recognition/report.md.
- Open
notebooks/01_colab_setup.ipynbin Google Colab. - Run all cells: mounts Drive, clones the DAN and POSTER_V2 reference repos into
third_party/, installsrequirements.txt, unpacks the dataset zip from your Drive into/content/data/. - For training:
notebooks/02_train_dan.ipynb(Week 2) ornotebooks/03_train_poster.ipynb(Week 3). - For analysis:
notebooks/04_analysis.ipynb(Week 4).
pip install -r requirements.txt
python app/gradio_app.pyOpens a Gradio interface on http://localhost:7860/ with image upload + webcam tabs and a model selector.
src/ library code (datasets, models, train/eval, inference)
app/ Gradio demo
configs/ YAML hyperparameter configs
notebooks/ Colab entrypoints
scripts/ dataset preparation
tests/ pytest smoke tests
research/ SOTA report (read-only)
This repo follows a strict change-log + commit-per-change discipline. See CLAUDE.md for the rules and CHANGES.md for the rolling log.