aiorscam/wildfc-spai

WildFC SPAI Inference

FastAPI wrapper for the WildFC SPAI checkpoint served as a Hugging Face Dedicated Inference Endpoint.

The endpoint expects the SPAI checkpoint and DINOv2 ViT-L/14 weights to be mounted from the Hugging Face model repository:

• /repository/wildfc_spai.pth
• /repository/dinov2_vitl14_pretrain.pth

API

POST /predict

{
  "image": "<base64 image or https URL>"
}

The response includes score, the WildFC synthetic probability, plus confidence, the predicted-class probability. The default threshold is the upstream WildFC cutoff of 0.5 and can be configured with WILDFC_FAKE_THRESHOLD and WILDFC_REAL_THRESHOLD.

Local Development

make sync
make validate

Running the production predictor requires CUDA, the pinned WildFC source tree, the pinned DINOv2 source tree, and the mounted model assets.

Image

make image-build
make image-push

The Dockerfile pins WildFC and DINOv2 commits, uses uv dependency cache mounts, and keeps model weights out of the image so Hugging Face can mount them from the model repository at runtime.

Hugging Face Endpoint

Use the custom Docker image with:

• repository: aiorscam/wildfc-spai
• image: docker.io/aiorscam/wildfc-inference:0.1.0
• health route: /health
• port: 8000
• hardware: AWS eu-west-1, nvidia-t4, x1
• replicas: min 0, max 1
• scale-to-zero timeout: 15 minutes

Eval

cd ../../eval
WILDFC_ENDPOINT_URL=https://... HF_TOKEN=... \
  uv run python run_eval.py \
    --manifest ~/datasets/social/manifest.csv \
    --models wildfc_endpoint \
    --output runs/wildfc-social \
    --overwrite