Api creation

.dockerignore

@@ -0,0 +1,51 @@
+# .dockerignore — exclut les fichiers inutiles de l'image Docker
+
+# Git
+.git/
+.gitignore
+
+# Python
+__pycache__/
+*.py[cod]
+*.pyo
+.pytest_cache/
+.ruff_cache/
+*.egg-info/
+dist/
+build/
+
+# Environnements virtuels
+.venv/
+venv/
+env/
+my_env/
+
+# Données & modèles locaux (téléchargés au runtime via HuggingFace / S3)
+data/
+*.pth
+*.pt
+datasets/WiderPeople/images/
+datasets/voc/images/
+train/
+val/
+test/
+
+# Notebooks & scripts d'entraînement (pas nécessaires dans l'API)
+main.ipynb
+SsdTrainingPipelineVOC2007.py
+*.ipynb
+
+# Secrets
+.env
+*.env
+
+# CI/CD & outils
+.github/
+mlruns/
+wandb/
+
+# Divers
+*.log
+*.zip
+*.tar.gz
+.DS_Store

.github/workflows/docker-deploy.yml

@@ -0,0 +1,54 @@
+# ─────────────────────────────────────────────────────────────────
+# .github/workflows/docker-deploy.yml
+# Pipeline CI/CD — conforme au cours ENSAE "Mise en production"
+#
+# Reproduit exactement le schéma du cours :
+#   push sur une branche → build image Docker → push sur Docker Hub
+#
+# Secrets GitHub à configurer :
+#   DOCKERHUB_USERNAME : votre identifiant Docker Hub
+#   DOCKERHUB_TOKEN    : Access Token Docker Hub
+#                        (hub.docker.com → Account Settings → Security)
+# ─────────────────────────────────────────────────────────────────
+name: Build & Deploy FastCrowdVision
+
+on:
+  push:
+
+jobs:
+  docker:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ secrets.DOCKERHUB_USERNAME }}/fastcrowdvision
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{version}}
+            type=sha,prefix=sha-,format=short
+            type=raw,value=latest,enable=${{ github.ref == 'refs/heads/main' }}
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Build and push
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+

Dockerfile

@@ -0,0 +1,68 @@
+# ─────────────────────────────────────────────────────────────────
+# FastCrowdVision — Dockerfile
+# Conforme aux bonnes pratiques du cours ENSAE "Mise en production"
+#
+# Points clés :
+#   - Image de base slim  → image légère (~600 Mo vs ~3 Go)
+#   - uv  pour installer les dépendances (recommandé dans le cours)
+#   - Séparation build / runtime (multi-stage)
+#   - Dépendances minimales : uniquement ce dont l'API a besoin
+#   - Utilisateur non-root
+# ─────────────────────────────────────────────────────────────────
+
+# ── Stage 1 : installation des dépendances ────────────────────────
+FROM python:3.11-slim AS builder
+
+# Copier uv depuis son image officielle (plus rapide et fiable que pip install uv)
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv
+
+WORKDIR /app
+
+# Dépendances système minimales pour OpenCV headless et torch
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        libgl1 \
+        libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copier UNIQUEMENT le fichier de dépendances en premier
+# → Docker met en cache cette couche et ne réinstalle pas si le code change
+COPY requirements-api.txt .
+
+# Créer un venv et installer les dépendances avec pip
+RUN python -m venv /opt/venv && \
+    /opt/venv/bin/pip install --upgrade pip && \
+    /opt/venv/bin/pip install --no-cache-dir -r requirements-api.txt
+
+
+# ── Stage 2 : image finale allégée ───────────────────────────────
+FROM python:3.11-slim AS runtime
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PATH="/opt/venv/bin:$PATH" \
+    HF_HOME=/app/.cache/huggingface
+
+WORKDIR /app
+
+# Libs runtime uniquement (pas d'outils de build)
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        libgl1 \
+        libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+# Récupérer le venv du stage builder
+COPY --from=builder /opt/venv /opt/venv
+
+# Copier le code applicatif
+# .dockerignore exclut : données, notebooks, scripts d'entraînement, .git
+COPY . .
+
+# Utilisateur non-root (bonne pratique sécurité Kubernetes)
+RUN useradd -m -u 1000 appuser && \
+    mkdir -p /app/.cache/huggingface && \
+    chown -R appuser:appuser /app
+USER appuser
+
+EXPOSE 8000
+
+CMD ["uvicorn", "server:app", "--host", "0.0.0.0", "--port", "8000"]

README.md

@@ -1,6 +1,10 @@
 # FastCrowdVision
 Implementing vision algorithms for human detection on mobile devices.
 
+
+
+for inference uvicorn server:app --reload
+
 ## Setup
 ```
 cp .env.example .env

inference.py

@@ -0,0 +1,105 @@
+# inference.py — Import-safe model loading and per-frame detection for the server.
+# Separates model setup from CLI/argparse concerns in SsdFastCrowdVision.py.
+
+import torch
+import numpy as np
+from huggingface_hub import hf_hub_download
+from mobilenetv3 import MobileNetV3Large
+from ssd import SSDLite
+from transforms import test_val_transform
+from eval import load_model
+from torchvision.transforms import v2
+import os
+import yaml
+import time
+
+# project root = folder containing this file
+project_root = os.path.dirname(os.path.abspath(__file__))
+
+
+def load_ssd_model(device):
+    """Build SSDLite + MobileNetV3 backbone, download weights from HuggingFace,
+    return (model, config_dict, transform) ready for inference."""
+
+    # build MobileNetV3Large backbone without the final avg-pool + classifier
+    mn = MobileNetV3Large(0.1, 1000, 1280)
+    backbone = mn.features[:-1]
+
+    # build SSDLite on top — use absolute path for the YAML config
+    # so the server works regardless of working directory
+    model = SSDLite(
+        backbone_config_path=os.path.join(project_root, "config", "ssdlite_mobilenetv3large.yaml"),
+        backbone=backbone,
+        c4_name="7.features.1.features.2",
+        nb_classes=4,
+        phase="test",
+        alpha=1.0,
+        prob_thr=0.01,
+        nms_thr=0.45,
+        top_k=200,
+        variances=[0.1, 0.2],
+        N_epochs=50,
+        device=device,
+    ).to(device)
+
+    # dummy optimizer — load_model() requires one to restore checkpoint state
+    optimizer = torch.optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0005)
+
+    # download weights from HuggingFace (cached locally after first download)
+    weights_path = hf_hub_download(
+        repo_id="aayrapet/SsdFastCrowdVision",
+        filename="SSD_FastCrowdVision_v1.pth",
+    )
+
+    # load trained weights into the model
+    model, _, _, max_map, _ = load_model(weights_path, device, model, optimizer)
+    model.eval()
+    model.phase = "test"
+    print(f"Model loaded — mAP on WiderPeople: {max_map}")
+
+    # load class names (1: pedestrians, 2: riders, 3: partially-visible persons)
+    wider_yaml = os.path.join(project_root, "datasets", "WiderPeople", "widerpeople.yaml")
+    with open(wider_yaml) as f:
+        config = yaml.safe_load(f)["names"]
+
+    # test transform: resize to 300x300, convert to float, normalize with ImageNet stats
+    transform = v2.Compose(test_val_transform)
+
+    return model, config, transform
+
+
+def detect_frame(model, pil_image, transform, device, score_thr=0.25):
+    """Run SSD on a single PIL image.
+
+    Returns np.ndarray of shape (N, 6) — columns: [x1, y1, x2, y2, score, class].
+    Coordinates are in pixels (original image size).
+    Returns empty (0, 6) array if no detections pass the threshold.
+    """
+    t0 = time.perf_counter()
+    W, H = pil_image.size
+
+    # transform PIL image to 300x300 normalized tensor, add batch dim
+    x = transform(pil_image).unsqueeze(0).to(device)
+
+    # SSD forward pass + NMS (no gradients needed at inference)
+    with torch.no_grad():
+        _, _, topk = model(x)
+
+    # remove batch dimension: (1, top_k, 6) → (top_k, 6)
+    topk = topk.squeeze(0)
+
+    # keep only detections above the score threshold
+    mask = topk[:, 4] > score_thr
+    if not mask.any():
+        return np.empty((0, 6))
+
+    topk = topk[mask]
+
+    # scale normalized [0,1] box coordinates to pixel coordinates
+    topk[:, [0, 2]] *= W
+    topk[:, [1, 3]] *= H
+
+    elapsed = time.perf_counter() - t0
+    print(f"[inference] {elapsed*1000:.1f} ms")
+
+    return topk.cpu().numpy()

kubernetes/deployment.yaml

@@ -0,0 +1,85 @@
+# ─────────────────────────────────────────────────────────────────
+# kubernetes/deployment.yaml
+# ─────────────────────────────────────────────────────────────────
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: fastcrowdvision
+  labels:
+    app: fastcrowdvision
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: fastcrowdvision
+  template:
+    metadata:
+      labels:
+        app: fastcrowdvision
+    spec:
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        fsGroup: 1000
+        fsGroupChangePolicy: "OnRootMismatch"
+
+      containers:
+        - name: fastcrowdvision
+          image: ghcr.io/josiepierr/fastcrowdvision:latest
+          imagePullPolicy: Always
+          ports:
+            - containerPort: 8000
+              name: http
+
+          env:
+            - name: HF_HOME
+              value: /app/.cache/huggingface
+
+          resources:
+            requests:
+              cpu: "500m"
+              memory: "2Gi"
+            limits:
+              cpu: "4"
+              memory: "4Gi"
+
+          # Attend que le modèle soit chargé (jusqu'à 10 min)
+          startupProbe:
+            httpGet:
+              path: /health
+              port: 8000
+            failureThreshold: 60
+            periodSeconds: 10
+            timeoutSeconds: 5
+
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: 8000
+            initialDelaySeconds: 120
+            periodSeconds: 30
+            timeoutSeconds: 5
+
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: 8000
+            initialDelaySeconds: 20
+            periodSeconds: 10
+            timeoutSeconds: 5
+
+          volumeMounts:
+            - name: hf-cache
+              mountPath: /app/.cache/huggingface
+            - name: tmp-dir
+              mountPath: /tmp
+
+      volumes:
+        - name: hf-cache
+          persistentVolumeClaim:
+            claimName: fastcrowdvision-hf-cache
+        - name: tmp-dir
+          emptyDir:
+            sizeLimit: 2Gi