DDA4210 Course Project Code

This is the only folder intended for submission and GitHub upload.

Structure

Code/
  baseline/       # unified baseline: ResNet18 and frozen CLIP
  new_model/      # causal FIND model
  data/           # dataset download/preparation scripts and manifests
  README.md
  requirements.txt

The project uses Tiny-GenImage for training/validation and MNW for final testing.

Data Layout

Place datasets under this folder:

Code/
  data/
    tiny-genimage/
      train/
      val/
    MNW/
      AI_Images/
    manifests/
      train.csv
      val.csv
      test.csv

Label convention:

real = 0
fake = 1

Build CSV manifests after placing or downloading data:

python data/scripts/build_manifests.py \
  --tiny-genimage-root data/tiny-genimage \
  --mnw-root data/MNW/AI_Images \
  --output-dir data/manifests

Training and testing code read from data/manifests/*.csv first. The committed test.csv is intentionally empty except for its header.

Check data readiness with:

python data/scripts/check_data.py

Install

cd Code
pip install -r requirements.txt

Baseline

Run from inside Code/:

python baseline/train.py --model resnet18
python baseline/train.py --model clip

Device selection defaults to CUDA when available, then MPS on macOS, then CPU. You can override it with --device cuda, --device mps, or --device cpu.

MNW evaluation:

export HF_TOKEN="your_huggingface_token"

python baseline/evaluate_mnw.py \
  --checkpoint baseline/outputs/resnet18_seed4210/best.pt

See baseline/README.md for more options.

New Model: Causal FIND

Run from inside Code/:

python -m new_model.main train-causal \
  --data-dir data/tiny-genimage \
  --output-dir runs/causal

MNW evaluation:

python -m new_model.main test-causal --output-dir runs/causal

See new_model/README.md for causal-model options and threshold settings.