"""Attend DeepFace demo — free HF Space (CPU). Slim surface for portfolio demos: detect + Facenet512 embed + in-memory gallery. Not the full FastAPI attendance product (auth/schedules/WS liveness). """ from __future__ import annotations import json import tempfile from pathlib import Path import cv2 import gradio as gr import numpy as np from deepface import DeepFace MODEL_NAME = "Facenet512" DETECTOR = "opencv" THRESHOLD = 0.45 # Ephemeral gallery for this Space process (clears on restart/sleep). _gallery: list[dict] = [] def _cosine(a: np.ndarray, b: np.ndarray) -> float: return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-10)) def _represent(path: str) -> list[dict]: return DeepFace.represent( img_path=path, model_name=MODEL_NAME, enforce_detection=True, detector_backend=DETECTOR, anti_spoofing=False, ) def _save_upload(image: np.ndarray) -> str: if image is None: raise gr.Error("Upload an image first.") # Gradio gives RGB; OpenCV/DeepFace file path is fine as PNG. rgb = image if image.ndim == 3 else cv2.cvtColor(image, cv2.COLOR_GRAY2RGB) fd, path = tempfile.mkstemp(suffix=".png") Path(path).write_bytes(b"") # ensure file exists for cv2 import os os.close(fd) # Write via cv2 expecting BGR bgr = cv2.cvtColor(rgb.astype(np.uint8), cv2.COLOR_RGB2BGR) if not cv2.imwrite(path, bgr): raise gr.Error("Could not write temp image.") return path def detect_faces(image: np.ndarray) -> tuple[np.ndarray | None, str]: path = _save_upload(image) try: faces = _represent(path) except Exception as exc: # DeepFace raises broadly return image, json.dumps({"ok": False, "error": str(exc)}, indent=2) finally: Path(path).unlink(missing_ok=True) annotated = image.copy() rows = [] for i, face in enumerate(faces): area = face.get("facial_area") or {} x, y, w, h = int(area.get("x", 0)), int(area.get("y", 0)), int(area.get("w", 0)), int(area.get("h", 0)) if w and h: cv2.rectangle(annotated, (x, y), (x + w, y + h), (0, 180, 120), 2) cv2.putText(annotated, f"face-{i}", (x, max(0, y - 8)), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 180, 120), 2) emb = np.array(face["embedding"], dtype=float) rows.append( { "index": i, "facial_area": area, "embedding_dim": int(emb.shape[0]), "embedding_preview": emb[:8].round(4).tolist(), } ) return annotated, json.dumps({"ok": True, "faces": rows, "model": MODEL_NAME}, indent=2) def enroll(image: np.ndarray, name: str) -> str: name = (name or "").strip() or f"person-{len(_gallery) + 1}" path = _save_upload(image) try: faces = _represent(path) except Exception as exc: return json.dumps({"ok": False, "error": str(exc)}, indent=2) finally: Path(path).unlink(missing_ok=True) if not faces: return json.dumps({"ok": False, "error": "No face detected"}, indent=2) emb = np.array(faces[0]["embedding"], dtype=float) _gallery.append({"name": name, "embedding": emb}) return json.dumps({"ok": True, "enrolled": name, "gallery_size": len(_gallery)}, indent=2) def recognize(image: np.ndarray) -> tuple[np.ndarray | None, str]: if not _gallery: return image, json.dumps({"ok": False, "error": "Gallery empty — enroll someone first."}, indent=2) path = _save_upload(image) try: faces = _represent(path) except Exception as exc: return image, json.dumps({"ok": False, "error": str(exc)}, indent=2) finally: Path(path).unlink(missing_ok=True) annotated = image.copy() results = [] for i, face in enumerate(faces): emb = np.array(face["embedding"], dtype=float) best_name, best = "Unknown", -1.0 for item in _gallery: score = _cosine(emb, item["embedding"]) if score > best: best, best_name = score, item["name"] matched = best >= THRESHOLD label = f"{best_name} ({best:.2f})" if matched else f"Unknown ({best:.2f})" area = face.get("facial_area") or {} x, y, w, h = int(area.get("x", 0)), int(area.get("y", 0)), int(area.get("w", 0)), int(area.get("h", 0)) color = (0, 180, 120) if matched else (200, 80, 60) if w and h: cv2.rectangle(annotated, (x, y), (x + w, y + h), color, 2) cv2.putText(annotated, label, (x, max(0, y - 8)), cv2.FONT_HERSHEY_SIMPLEX, 0.55, color, 2) results.append({"index": i, "name": best_name if matched else "Unknown", "score": round(best, 4), "matched": matched}) return annotated, json.dumps({"ok": True, "results": results, "threshold": THRESHOLD}, indent=2) def clear_gallery() -> str: _gallery.clear() return json.dumps({"ok": True, "gallery_size": 0}, indent=2) with gr.Blocks(title="Attend DeepFace (CPU)") as demo: gr.Markdown( """# Attend — DeepFace demo (free CPU) Upload a face photo to detect embeddings (Facenet512) or enroll/recognize against an ephemeral gallery. Full attendance API (schedules, liveness WS, SQLite) is **not** hosted here — see repo + Azure FE shell.""" ) with gr.Tab("Detect"): inp = gr.Image(type="numpy", label="Face image") out_img = gr.Image(type="numpy", label="Annotated") out_json = gr.Code(language="json", label="Result") gr.Button("Detect").click(detect_faces, inputs=inp, outputs=[out_img, out_json]) with gr.Tab("Enroll / Recognize"): name = gr.Textbox(label="Name for enroll", placeholder="Alice") inp2 = gr.Image(type="numpy", label="Face image") out_img2 = gr.Image(type="numpy", label="Annotated") out_json2 = gr.Code(language="json", label="Result") with gr.Row(): gr.Button("Enroll").click(enroll, inputs=[inp2, name], outputs=out_json2) gr.Button("Recognize").click(recognize, inputs=inp2, outputs=[out_img2, out_json2]) gr.Button("Clear gallery").click(clear_gallery, outputs=out_json2) if __name__ == "__main__": demo.queue(max_size=8).launch()