"""Shared document policy and inference; importing this module does not load weights.""" import math import os import threading import time from PIL import Image, ImageOps MODEL_ID = "thaitea/laya-vision-201m" REVISION = "0b6228f7a0762566de1c4539e9aa4eb1c1aef5f4" POLICY_VERSION = "document-v2-short-prompt" EXTENSIONS = {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".tif", ".tiff"} QUESTION = { "document": { "type": "noul", "instructions": "Is the main subject a printed form or document?", } } _agent = None _lock = threading.Lock() def load_agent(device=None): """Load the pinned checkpoint once; prefer available GPU, then Apple MPS, then CPU.""" global _agent with _lock: if _agent is None: import laya import torch torch.set_num_threads(min(8, os.cpu_count() or 1)) selected = device or os.environ.get("LAYA_DEVICE") print(f"Loading {MODEL_ID}@{REVISION} device={selected or 'auto'}", flush=True) _agent = laya.load_vlm(MODEL_ID, revision=REVISION, device=selected) print(f"Model ready on {_agent.device}", flush=True) return _agent def decision(probability, threshold=0.5, review_confidence=0.8): p = float(probability) if not math.isfinite(p) or not 0 <= p <= 1: raise ValueError("Model returned an invalid probability") if not 0 < threshold < 1 or not 0.5 <= review_confidence <= 1: raise ValueError("Invalid threshold or review confidence") is_document = p >= threshold selected_probability = p if is_document else 1 - p return { "label": "document" if is_document else "non_document", "document_probability": p, "selected_probability": selected_probability, "needs_review": selected_probability < review_confidence, "threshold": threshold, "review_confidence": review_confidence, } def classify(image, threshold=0.5, review_confidence=0.8, agent=None): """Predict whether an image is primarily a document under the agreed policy.""" model = agent if agent is not None else load_agent() start = time.monotonic() with _lock: normalized = ImageOps.exif_transpose(image).convert("RGB") answer = model.predict({"image": normalized}, QUESTION, strict=True)["answers"]["document"] return { **decision(answer["noul"], threshold, review_confidence), "elapsed_seconds": round(time.monotonic() - start, 3), "model": MODEL_ID, "revision": REVISION, "policy": POLICY_VERSION, }