docs: update AGENTS.md — ONNX directory, test app, corrected config specs
Browse files
AGENTS.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# AGENTS.md — CommunityForensics-DeepfakeDet-ViT
|
| 2 |
+
|
| 3 |
+
## What this repo is
|
| 4 |
+
Hugging Face model repo for `buildborderless/CommunityForensics-DeepfakeDet-ViT` — a ViT-Small classifier for deepfake image detection. Trained on 2.7M samples across 4,803 generators. This is a model distribution repo (no app, no build, no tests).
|
| 5 |
+
|
| 6 |
+
## Key files
|
| 7 |
+
- **`model.safetensors`** — HF-format weights (Git LFS — ensure `git lfs pull` after clone)
|
| 8 |
+
- **`config.json`** — `ViTForImageClassification` config (384×384, 6 heads, 2 output classes: real/fake)
|
| 9 |
+
- **`preprocessor_config.json`** — CLIP-style normalization, resize to 440, center-crop to 384
|
| 10 |
+
- **`modeling_vit_classifier.py`** — **DEPRECATED** (moved to `scripts/`). Use standard HF path below.
|
| 11 |
+
- **`pretrained_weights/`** — original `.pt` checkpoints from training (also LFS)
|
| 12 |
+
- **`onnx/`** — 8 pre-exported ONNX variants (21MB–138MB) for CPU/GPU deployment. See README for variant guide.
|
| 13 |
+
- **`test_app/`** — local Gradio app for testing PyTorch vs ONNX, benchmarking, and comparison
|
| 14 |
+
|
| 15 |
+
## Usage
|
| 16 |
+
The model is hosted on Hugging Face. The standard way to load it is via `transformers`:
|
| 17 |
+
```python
|
| 18 |
+
from transformers import ViTForImageClassification, ViTImageProcessor
|
| 19 |
+
model = ViTForImageClassification.from_pretrained("buildborderless/CommunityForensics-DeepfakeDet-ViT")
|
| 20 |
+
processor = ViTImageProcessor.from_pretrained("buildborderless/CommunityForensics-DeepfakeDet-ViT")
|
| 21 |
+
```
|
| 22 |
+
|
| 23 |
+
The custom wrapper (`modeling_vit_classifier.py`) uses `timm.create_model` with a sigmoid output and `pretrained_weights/model_v11_ViT_384_base_ckpt.pt`. This is for standalone (non-HF-pipeline) inference requiring both `timm` and `transformers`.
|
| 24 |
+
|
| 25 |
+
## Dependencies
|
| 26 |
+
- `transformers` >= 4.50.0
|
| 27 |
+
- `timm` (for the custom ViTClassifier wrapper only)
|
| 28 |
+
- `torch`, `torchvision`, `Pillow`
|
| 29 |
+
|
| 30 |
+
## Scripts (in `scripts/`)
|
| 31 |
+
Data processing utilities for the eval dataset — not needed for inference:
|
| 32 |
+
- `convert_to_pytorch.py` — convert timm checkpoints to HuggingFace format
|
| 33 |
+
- `resample_evalset.py` — face-detection-based dataset filtering
|
| 34 |
+
- `restructure.py` — reorganize real/generated image directories
|
| 35 |
+
- `quick_analysis.py` — dataset statistics report
|
| 36 |
+
|
| 37 |
+
## Git LFS
|
| 38 |
+
All weight files (`.safetensors`, `.pt`, `.ckpt`, `.onnx`) are stored via Git LFS. Always run `git lfs pull` after cloning or the model files will be pointer stubs. The full ONNX model alone is 138MB — pull selectively with `git lfs pull --include="onnx/model_int8.onnx"` if you only need one variant.
|
| 39 |
+
|
| 40 |
+
## Remote
|
| 41 |
+
This repo is pushed to `https://huggingface.co/buildborderless/CommunityForensics-DeepfakeDet-ViT`, not GitHub. Standard `gh` CLI commands will not work.
|