Upload README.md with huggingface_hub

README.md

@@ -10,70 +10,71 @@ tags:
 pipeline_tag: video-classification
 ---
 
-# DAViD — ViFi-CLIP Encoder (ViT-B/16, K400-finetuned)
+# DAViD — Checkpoints (ViFi-CLIP encoder + classification head)
 
-Visual feature extractor for **[DAViD](https://huggingface.co/spaces/aitf-its-tim3-dfk/David)**,
-a deepfake & AI-generated video detector.
+Model weights for **[DAViD](https://huggingface.co/spaces/aitf-its-tim3-dfk/David)**,
+a deepfake & AI-generated video/image detector.
 
-This repo hosts a single checkpoint:
+This repo hosts two checkpoints:
 
 | File | Size | Description |
 |---|---|---|
 | `k400_clip_complete_finetuned_30_epochs.pth` | ~1.6 GB | ViFi-CLIP (ViT-B/16) image encoder, fine-tuned on Kinetics-400 for 30 epochs |
+| `best_detector_model.pt` | ~3 MB | MLP classification head (`dense → dense1 → dense2`), trained on the DAViD dataset + CDDB |
 
-## What it is
+## How they fit together
 
-A **ViFi-CLIP (ViT-B/16)** image encoder whose CLIP visual backbone has been
-fine-tuned on **Kinetics-400**. In the DAViD pipeline it acts as a frozen
-feature extractor: each frame is encoded into a **512-dim** embedding, frame
-embeddings are averaged per scene / per clip, and a lightweight MLP
-classification head predicts one of three classes — `real`, `deepfake`,
-`ai_gen`.
+1. **Encoder** — a ViFi-CLIP (ViT-B/16) visual backbone fine-tuned on
+   Kinetics-400. Each frame (or image) is encoded into a **512-dim** embedding.
+2. **Classification head** — a lightweight MLP that maps the (averaged) 512-dim
+   embedding to 3 classes: `real`, `deepfake`, `ai_gen`. It was trained on a mix
+   of the DAViD video dataset and **CDDB** (an image-based deepfake benchmark),
+   so it supports both video and single-image input.
 
-The MLP classification head is a separate, much smaller checkpoint and is **not**
-in this repo.
+## Why these live on the Hub
 
-## Why it lives on the Hub
-
-The DAViD Space downloads this checkpoint at Docker build time. It was
-previously hosted on Google Drive, but Drive throttles datacenter IPs and broke
-the Space build. Serving it from the HF Hub is reliable from HF's build
-infrastructure.
+The DAViD Space downloads these at Docker build time. They were previously on
+Google Drive, but Drive throttles datacenter IPs and broke the Space build.
+Serving them from the HF Hub is reliable from HF's build infrastructure.
 
 ## Usage
 
-Download the checkpoint (no auth needed — public repo):
+Download (no auth needed — public repo):
 
 ```python
 from huggingface_hub import hf_hub_download
 
-ckpt = hf_hub_download(
-    repo_id="aitf-its-tim3-dfk/david-encoder",
-    filename="k400_clip_complete_finetuned_30_epochs.pth",
-)
+REPO = "aitf-its-tim3-dfk/david-encoder"
+encoder_ckpt    = hf_hub_download(REPO, "k400_clip_complete_finetuned_30_epochs.pth")
+classifier_ckpt = hf_hub_download(REPO, "best_detector_model.pt")
 ```
 
-Load it with the DAViD encoder loader (from the training repo
+Load with the DAViD modules (from the training repo
 [`aitf-its-tim3-dfk/david`](https://github.com/aitf-its-tim3-dfk/david)):
 
 ```python
+import torch
 from encoder import load_feature_extractor
+from model import ClassificationHead
 
 feature_extractor = load_feature_extractor(
     arch="ViT-B/16",
     class_names=("real", "deepfake", "ai_gen"),
-    checkpoint_path=ckpt,
+    checkpoint_path=encoder_ckpt,
 ).eval()
 
-# feature_extractor.image_encoder(frames) -> (N, 512) embeddings
+classifier = ClassificationHead(input_dim=512, num_classes=3)
+classifier.load_state_dict(torch.load(classifier_ckpt, map_location="cpu", weights_only=False))
+classifier.eval()
+
+# feats = feature_extractor.image_encoder(frames)   # (N, 512)
+# logits = classifier(feats.mean(dim=0, keepdim=True))  # (1, 3)
 ```
 
 ## Training
 
-- **Backbone:** CLIP ViT-B/16 (ViFi-CLIP setup)
-- **Fine-tuning data:** Kinetics-400
-- **Schedule:** 30 epochs
-- **Output dim:** 512
+- **Encoder:** CLIP ViT-B/16 (ViFi-CLIP), fine-tuned on Kinetics-400, 30 epochs, output dim 512.
+- **Classification head:** MLP trained on DAViD video dataset + CDDB images (branch `feat-cddb`).
 
 ## Related
 
@@ -82,5 +83,5 @@ feature_extractor = load_feature_extractor(
 
 ## License
 
-Set the appropriate license for these weights (currently `other`). Note that the
-CLIP backbone and Kinetics-400 carry their own upstream licenses/terms.
+Set the appropriate license for these weights (currently `other`). The CLIP
+backbone, Kinetics-400, and CDDB carry their own upstream licenses/terms.