DINOv2 Real/Fake Classifier

This project fine-tunes a DINOv2 backbone on a binary classification task: distinguishing real images from AI-generated (fake) images.

Paths to adjust

• Training data default: /mnt/c/Users/warre/workspace/train (expects fake/ and real/ subfolders). Override with --train-dir <path>.
• Test data default: /mnt/c/Users/warre/workspace/test. Override with --test-dir <path>.
• Checkpoint output default: checkpoints/best.pt (created automatically). Override with --checkpoint <path>.
• Submission output default: submission.csv. Override with --submission <path>.

Environment

• Python 3.10+ recommended.
• PyTorch + torchvision required; the script auto-downloads the DINOv2 backbone via torch.hub (facebookresearch/dinov2).
• CUDA is used if available; mixed precision is on by default for speed.

Train

cd /home/warren/deeplearning/final_project
python team_3_DINOv2.py --mode train \
  --train-dir /mnt/c/Users/warre/workspace/train \
  --val-ratio 0.1 \
  --epochs 5 \
  --batch-size 32 \
  --img-size 256

Key outputs:

• Best model: checkpoints/best.pt
• Training log: console JSON per epoch
• Metrics history: checkpoints/history.json

Common tweaks

• Memory tight: lower --img-size or raise --accumulation-steps.
• Longer training: increase --epochs.
• Different backbone size: --backbone dinov2_vitb14 (or other names from the hub repo).

Evaluation (on held-out validation split)

Validation runs automatically each epoch during training using --val-ratio. The best checkpoint is chosen by highest validation accuracy. Results live in checkpoints/history.json and console output.

To re-evaluate a saved checkpoint on a specific folder split:

python - <<'PY'
from pathlib import Path
import torch
from team_3_DINOv2 import DinoV2Classifier, RealFakeDataset, build_transforms, evaluate
from torch.utils.data import DataLoader

ckpt_path = Path('checkpoints/best.pt')
data_dir = Path('/mnt/c/Users/warre/workspace/train')  # replace with your eval folder containing fake/ and real/
img_size = 256

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
_, eval_tfms = build_transforms(img_size)
val_ds = RealFakeDataset(data_dir, transform=eval_tfms)
val_loader = DataLoader(val_ds, batch_size=64, shuffle=False, num_workers=4, pin_memory=True)
ckpt = torch.load(ckpt_path, map_location=device)
model = DinoV2Classifier(backbone_name=ckpt.get('cfg', {}).get('backbone', 'dinov2_vits14'))
model.load_state_dict(ckpt['model'], strict=False)
model.to(device)
val_loss, val_acc = evaluate(model, val_loader, device)
print({'val_loss': val_loss, 'val_acc': val_acc})
PY

Predict / submission (test set)

python team_3_DINOv2.py --mode predict \
  --checkpoint checkpoints/best.pt \
  --test-dir /mnt/c/Users/warre/workspace/test \
  --submission submission.csv \
  --img-size 256

Output:

• Submission CSV matching sample_submission.csv format with columns filename,label; labels are real/fake using a 0.5 probability threshold for the real class.

File map

• Training/inference script: team_3_DINOv2.py
• Sample submission template: sample_submission.csv
• Generated assets after training: checkpoints/best.pt, checkpoints/history.json, and your chosen submission.csv.