Initial Space: Docker layout cloning FMODetect-v2 main

Dockerfile

@@ -0,0 +1,53 @@
+# Hugging Face Space Dockerfile for FMODetect v2.
+# Builds the Next.js UI as a static export, then runs FastAPI on port 7860.
+# The source repo is cloned at build time so the Space stays tiny.
+
+ARG REPO_URL=https://github.com/jai-krishna-0921/FMODetect-v2.git
+ARG REPO_REF=main
+
+# --- Stage 1: build the Next.js static export ---
+FROM node:20-bookworm-slim AS ui-build
+ARG REPO_URL
+ARG REPO_REF
+WORKDIR /src
+RUN apt-get update && apt-get install -y --no-install-recommends git ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+RUN git clone --depth 1 --branch ${REPO_REF} ${REPO_URL} app
+WORKDIR /src/app/ui
+RUN npm install --no-audit --no-fund
+RUN NEXT_OUTPUT=export npm run build
+
+# --- Stage 2: Python runtime serving API + static UI ---
+FROM python:3.12-slim AS runtime
+ARG REPO_URL
+ARG REPO_REF
+
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PIP_NO_CACHE_DIR=1 \
+    HF_HOME=/data/hf-cache \
+    FMODETECT_STATIC=/data/static \
+    PYTHONPATH=/app
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        git ca-certificates libgl1 libglib2.0-0 ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+
+# Clone source at build time (overlaid in the next step with the UI build).
+RUN git clone --depth 1 --branch ${REPO_REF} ${REPO_URL} /app && \
+    rm -rf /app/ui && mkdir -p /app/ui
+
+# Bring in the built static export from stage 1.
+COPY --from=ui-build /src/app/ui/out /app/ui/out
+
+# Install deploy-only requirements (CPU torch).
+COPY requirements.txt /tmp/requirements.txt
+RUN pip install -r /tmp/requirements.txt
+
+# Writable runtime dirs for HF cache and generated overlays.
+RUN mkdir -p /data/hf-cache /data/static && chmod -R 777 /data
+
+EXPOSE 7860
+CMD ["uvicorn", "api.main:app", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -1,10 +1,43 @@
 ---
-title: Fmodetect V2
-emoji: 👁
-colorFrom: blue
+title: FMODetect v2
+emoji: 🌀
+colorFrom: gray
 colorTo: indigo
 sdk: docker
+app_port: 7860
 pinned: false
+license: mit
+short_description: Fast-moving-object detection from a single blurred frame
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# FMODetect v2 — research demo
+
+PyTorch re-implementation of [FMODetect (Rozumnyi et al., ICCV 2021)](https://arxiv.org/abs/2012.08216)
+with three additions: CBAM attention, a joint TDF + matting head, and an
+uncertainty-weighted boundary loss.
+
+Source: <https://github.com/jai-krishna-0921/FMODetect-v2>
+
+## How this Space is built
+
+This Space contains only a `Dockerfile`, a `requirements.txt` and this README.
+At build time the Dockerfile clones the source repo, builds the Next.js UI as
+a static export, and serves it from FastAPI on port 7860.
+
+## Environment
+
+Set these in the Space settings → Variables and secrets:
+
+| key                      | value                                  |
+|--------------------------|----------------------------------------|
+| `FMODETECT_HF_REPO`      | `<your-username>/fmodetect-v2`         |
+| `FMODETECT_HF_FILENAME`  | `best.pt` (default; only set to override) |
+
+The checkpoint is downloaded once at first request and cached on the Space
+disk. To swap models, upload a new file to the HF Hub model repo and restart
+the Space.
+
+## Hardware
+
+CPU Basic (free) runs inference in ~2–3 s per image pair. T4 small (~$0.40/hr)
+brings it under 200 ms.

requirements.txt

@@ -0,0 +1,15 @@
+# CPU-only deploy dependencies for the Hugging Face Space.
+# Pinned to a slim runtime set — no training / experiment-tracking deps.
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch==2.12.0+cpu
+torchvision==0.27.0+cpu
+numpy>=2.1
+scipy>=1.14
+scikit-image>=0.25
+opencv-python-headless>=4.10
+imageio[ffmpeg]>=2.36
+pillow>=11
+fastapi>=0.115
+uvicorn[standard]>=0.32
+python-multipart>=0.0.20
+huggingface_hub>=0.26