app.py

"""
Industrial Defect Detection with Gradio for HuggingFace Spaces.
Provides both UI and API endpoints through Gradio's native API system.

Includes:
- YOLO-based defect detection models (Data Matrix, Tire Cord, etc.)
- Zero-shot anomaly detection via AdaCLIP (no training data required)
"""

import gradio as gr
import onnxruntime as ort
import numpy as np
import cv2
from huggingface_hub import hf_hub_download
import os
import logging
from collections import defaultdict
from datetime import datetime, timedelta
import time
import tempfile

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Rate limiting configuration
RATE_LIMIT_REQUESTS = 100  # Max requests per window
RATE_LIMIT_WINDOW = 60  # Window in seconds (1 minute)
request_tracker = defaultdict(list)

def check_rate_limit(identifier="global"):
    """Simple rate limiting - allows RATE_LIMIT_REQUESTS per RATE_LIMIT_WINDOW"""
    current_time = time.time()
    cutoff_time = current_time - RATE_LIMIT_WINDOW

    # Remove old requests
    request_tracker[identifier] = [
        req_time for req_time in request_tracker[identifier]
        if req_time > cutoff_time
    ]

    # Check if over limit
    if len(request_tracker[identifier]) >= RATE_LIMIT_REQUESTS:
        logger.warning(f"Rate limit exceeded for {identifier}")
        return False

    # Add current request
    request_tracker[identifier].append(current_time)
    return True


def extract_bboxes_from_heatmap(heatmap_path: str, orig_w: int, orig_h: int, threshold: float = 0.5):
    """Extract bounding boxes from AdaCLIP heatmap image."""
    try:
        heatmap = cv2.imread(heatmap_path)
        if heatmap is None:
            return []

        if len(heatmap.shape) == 3:
            heatmap_gray = cv2.cvtColor(heatmap, cv2.COLOR_BGR2GRAY)
        else:
            heatmap_gray = heatmap

        heatmap_resized = cv2.resize(heatmap_gray, (orig_w, orig_h))
        heatmap_norm = heatmap_resized.astype(np.float32) / 255.0

        binary_threshold = threshold * 0.5
        _, binary_mask = cv2.threshold(heatmap_norm, binary_threshold, 1.0, cv2.THRESH_BINARY)
        binary_mask = (binary_mask * 255).astype(np.uint8)

        contours, _ = cv2.findContours(binary_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

        bboxes = []
        min_area = (orig_w * orig_h) * 0.001

        for contour in contours:
            area = cv2.contourArea(contour)
            if area < min_area:
                continue

            x, y, w, h = cv2.boundingRect(contour)
            roi = heatmap_norm[y:y+h, x:x+w]
            confidence = float(np.mean(roi)) if roi.size > 0 else 0.5

            bboxes.append({
                "x1": float(x),
                "y1": float(y),
                "x2": float(x + w),
                "y2": float(y + h),
                "confidence": confidence
            })

        return bboxes
    except Exception as e:
        logger.error(f"Error extracting bboxes from heatmap: {e}")
        return []


def run_clip_anomaly_inference(image_bytes: bytes, confidence: float = 0.25):
    """
    Run zero-shot anomaly detection using CLIP similarity scoring.

    This uses CLIP to compare the image against "normal" vs "defect" descriptions.
    Returns detection if the image is more similar to defect descriptions than normal.
    """
    try:
        from transformers import CLIPProcessor, CLIPModel
        from PIL import Image
        import torch
        import io

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size

        # Initialize model and processor (cached after first load)
        if not hasattr(run_clip_anomaly_inference, 'processor'):
            logger.info("Loading CLIP model (first time only)...")
            run_clip_anomaly_inference.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
            run_clip_anomaly_inference.model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
            logger.info("CLIP model loaded successfully")

        processor = run_clip_anomaly_inference.processor
        model = run_clip_anomaly_inference.model

        # Simpler binary comparison: normal vs defect
        text_descriptions = [
            "a high quality product without any defects or anomalies",
            "a defective product with visible defects, cracks, scratches, or damage"
        ]

        # Process inputs
        inputs = processor(
            text=text_descriptions,
            images=image,
            return_tensors="pt",
            padding=True
        )

        # Run inference
        with torch.no_grad():
            outputs = model(**inputs)
            logits_per_image = outputs.logits_per_image
            probs = logits_per_image.softmax(dim=1)

        # Get probabilities
        normal_prob = float(probs[0][0])
        defect_prob = float(probs[0][1])

        logger.info(f"CLIP probabilities - Normal: {normal_prob:.3f}, Defect: {defect_prob:.3f}")

        detections = []

        # If defect probability is higher than threshold, create detection
        # This means the image looks more like a defect than normal
        if defect_prob >= confidence:
            detections.append({
                "bbox": [0, 0, orig_w, orig_h],
                "confidence": defect_prob,
                "class_id": 0,
                "class_name": "anomaly",
                "x1": 0,
                "y1": 0,
                "x2": orig_w,
                "y2": orig_h,
                "anomaly_score": defect_prob,
                "normal_score": normal_prob,
                "model_type": "clip",
                "description": f"CLIP anomaly detection (defect:{defect_prob:.2f} vs normal:{normal_prob:.2f})"
            })

        logger.info(f"CLIP result - Defect score: {defect_prob:.3f}, Detections: {len(detections)}")
        return detections, defect_prob

    except Exception as e:
        logger.error(f"CLIP inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return [], 0.0


def run_florence2_inference(image_bytes: bytes, confidence: float = 0.3):
    """
    Run zero-shot object detection using Florence-2 (Microsoft).

    Florence-2 is a multimodal vision-language model that can detect objects,
    generate captions, and understand context - similar to Claude but open-source.
    """
    try:
        from transformers import AutoProcessor, AutoModelForCausalLM
        from PIL import Image
        import torch
        import io

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"Florence-2: Processing image {orig_w}x{orig_h}")

        # Initialize model and processor (cached after first load)
        if not hasattr(run_florence2_inference, 'processor'):
            logger.info("Loading Florence-2 model (first time only - may take a moment)...")
            run_florence2_inference.processor = AutoProcessor.from_pretrained(
                "microsoft/Florence-2-base",  # Use base model - faster and more reliable
                trust_remote_code=True
            )
            run_florence2_inference.model = AutoModelForCausalLM.from_pretrained(
                "microsoft/Florence-2-base",
                trust_remote_code=True
            )
            logger.info("Florence-2 model loaded successfully")

        processor = run_florence2_inference.processor
        model = run_florence2_inference.model

        # Try dense region caption first - better for finding all objects
        task_prompt = "<DENSE_REGION_CAPTION>"
        inputs = processor(text=task_prompt, images=image, return_tensors="pt")

        logger.info(f"Florence-2: Running inference with task {task_prompt}")

        # Run inference
        with torch.no_grad():
            generated_ids = model.generate(
                input_ids=inputs["input_ids"],
                pixel_values=inputs["pixel_values"],
                max_new_tokens=1024,
                num_beams=3,
            )

        # Decode results
        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
        logger.info(f"Florence-2 generated text: {generated_text[:200]}...")

        parsed_answer = processor.post_process_generation(
            generated_text,
            task=task_prompt,
            image_size=(orig_w, orig_h)
        )

        logger.info(f"Florence-2 parsed answer keys: {list(parsed_answer.keys())}")
        logger.info(f"Florence-2 parsed answer: {str(parsed_answer)[:500]}...")

        detections = []

        # Try multiple possible output formats
        if task_prompt in parsed_answer:
            result = parsed_answer[task_prompt]

            # Dense region caption format
            if 'bboxes' in result and 'labels' in result:
                bboxes = result['bboxes']
                labels = result['labels']
                logger.info(f"Florence-2 found {len(bboxes)} regions")

                for bbox, label in zip(bboxes, labels):
                    x1, y1, x2, y2 = bbox

                    detections.append({
                        "bbox": [float(x1), float(y1), float(x2), float(y2)],
                        "confidence": 0.9,
                        "class_id": 0,
                        "class_name": str(label),
                        "x1": float(x1),
                        "y1": float(y1),
                        "x2": float(x2),
                        "y2": float(y2),
                        "model_type": "florence2"
                    })

        logger.info(f"Florence-2 detected {len(detections)} objects: {[d.get('class_name', '?') for d in detections]}")
        return detections

    except Exception as e:
        logger.error(f"Florence-2 inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return []


def run_groundingdino_inference(image_bytes: bytes, text_queries: list = None, confidence: float = 0.3):
    """
    Run zero-shot object detection using GroundingDINO (IDEA Research).

    GroundingDINO is better than OWL-ViT for open-set object detection.
    It can find objects based on text descriptions with better accuracy.
    """
    try:
        from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
        from PIL import Image
        import torch
        import io

        if text_queries is None:
            text_queries = ["defect", "anomaly", "crack", "scratch", "damage", "error", "imperfection"]

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"GroundingDINO: Processing image {orig_w}x{orig_h}")

        # Initialize model and processor (cached after first load)
        if not hasattr(run_groundingdino_inference, 'processor'):
            logger.info("Loading GroundingDINO model (first time only)...")
            run_groundingdino_inference.processor = AutoProcessor.from_pretrained("IDEA-Research/grounding-dino-tiny")
            run_groundingdino_inference.model = AutoModelForZeroShotObjectDetection.from_pretrained("IDEA-Research/grounding-dino-tiny")
            logger.info("GroundingDINO model loaded successfully")

        processor = run_groundingdino_inference.processor
        model = run_groundingdino_inference.model

        # Create text prompt (comma-separated)
        text_prompt = ". ".join(text_queries) + "."

        # Prepare inputs
        inputs = processor(images=image, text=text_prompt, return_tensors="pt")

        # Run inference
        with torch.no_grad():
            outputs = model(**inputs)

        # Post-process results
        results = processor.post_process_grounded_object_detection(
            outputs,
            inputs.input_ids,
            box_threshold=confidence,
            text_threshold=confidence,
            target_sizes=[(orig_h, orig_w)]
        )[0]

        detections = []

        if len(results["boxes"]) > 0:
            boxes = results["boxes"].cpu().numpy()
            scores = results["scores"].cpu().numpy()
            labels = results["labels"]

            logger.info(f"GroundingDINO found {len(boxes)} objects")

            for box, score, label in zip(boxes, scores, labels):
                x1, y1, x2, y2 = box

                detections.append({
                    "bbox": [float(x1), float(y1), float(x2), float(y2)],
                    "confidence": float(score),
                    "class_id": 0,
                    "class_name": str(label),
                    "x1": float(x1),
                    "y1": float(y1),
                    "x2": float(x2),
                    "y2": float(y2),
                    "model_type": "groundingdino"
                })

        logger.info(f"GroundingDINO detected {len(detections)} objects: {[d['class_name'] for d in detections]}")
        return detections

    except Exception as e:
        logger.error(f"GroundingDINO inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return []


def run_yoloworld_inference(image_bytes: bytes, text_queries: list = None, confidence: float = 0.3):
    """
    Run zero-shot object detection using YOLO-World.

    YOLO-World combines YOLO speed with open-vocabulary detection.
    Fast and effective for real-time anomaly detection.
    """
    try:
        from ultralytics import YOLOWorld
        from PIL import Image
        import io
        import numpy as np

        if text_queries is None:
            text_queries = ["defect", "anomaly", "crack", "scratch", "damage"]

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"YOLO-World: Processing image {orig_w}x{orig_h}")

        # Initialize model (cached after first load)
        if not hasattr(run_yoloworld_inference, 'model'):
            logger.info("Loading YOLO-World model (first time only)...")
            run_yoloworld_inference.model = YOLOWorld("yolov8s-world.pt")  # Small model
            logger.info("YOLO-World model loaded successfully")

        model = run_yoloworld_inference.model

        # Set custom classes
        model.set_classes(text_queries)

        # Convert PIL to numpy array
        img_array = np.array(image)

        # Run inference
        results = model.predict(img_array, conf=confidence, verbose=False)

        detections = []

        if len(results) > 0 and results[0].boxes is not None:
            boxes = results[0].boxes
            logger.info(f"YOLO-World found {len(boxes)} objects")

            for box in boxes:
                x1, y1, x2, y2 = box.xyxy[0].cpu().numpy()
                conf = float(box.conf[0].cpu().numpy())
                cls = int(box.cls[0].cpu().numpy())
                class_name = text_queries[cls] if cls < len(text_queries) else "object"

                detections.append({
                    "bbox": [float(x1), float(y1), float(x2), float(y2)],
                    "confidence": conf,
                    "class_id": cls,
                    "class_name": class_name,
                    "x1": float(x1),
                    "y1": float(y1),
                    "x2": float(x2),
                    "y2": float(y2),
                    "model_type": "yoloworld"
                })

        logger.info(f"YOLO-World detected {len(detections)} objects: {[d['class_name'] for d in detections]}")
        return detections

    except Exception as e:
        logger.error(f"YOLO-World inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return []


def run_winclip_inference(image_bytes: bytes, confidence: float = 0.3):
    """
    Run zero-shot anomaly detection using WinCLIP (CVPR 2023).

    WinCLIP achieves 91.8% AUROC on MVTec AD using window-based CLIP
    with compositional prompts. Best performing proven zero-shot method.

    Based on: https://github.com/mala-lab/WinCLIP
    """
    try:
        from transformers import CLIPProcessor, CLIPModel
        from PIL import Image
        import torch
        import io

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"WinCLIP: Processing image {orig_w}x{orig_h}")

        # Initialize model (cached after first load)
        if not hasattr(run_winclip_inference, 'model'):
            logger.info("Loading WinCLIP model (CLIP-based, first time only)...")
            run_winclip_inference.processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
            run_winclip_inference.model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
            logger.info("WinCLIP model loaded successfully")

        processor = run_winclip_inference.processor
        model = run_winclip_inference.model

        # WinCLIP compositional prompts (state words + templates)
        normal_prompts = [
            "a photo of a flawless {object}",
            "a photo of a perfect {object}",
            "a photo of a {object} without defects",
            "a photo of an unblemished {object}",
        ]

        abnormal_prompts = [
            "a photo of a {object} with defects",
            "a photo of a damaged {object}",
            "a photo of a broken {object}",
            "a photo of a {object} with anomalies",
        ]

        # Generic object terms for industrial products
        object_terms = ["product", "item", "surface", "material"]

        # Create compositional prompts
        normal_texts = [prompt.format(object=obj) for obj in object_terms for prompt in normal_prompts]
        abnormal_texts = [prompt.format(object=obj) for obj in object_terms for prompt in abnormal_prompts]
        all_texts = normal_texts + abnormal_texts

        # Compute image-text similarity
        inputs = processor(text=all_texts, images=image, return_tensors="pt", padding=True)
        with torch.no_grad():
            outputs = model(**inputs)
            logits_per_image = outputs.logits_per_image
            probs = logits_per_image.softmax(dim=1)

        # Average probabilities for normal vs abnormal
        num_normal = len(normal_texts)
        normal_prob = float(probs[0][:num_normal].mean())
        abnormal_prob = float(probs[0][num_normal:].mean())

        logger.info(f"WinCLIP: Normal={normal_prob:.3f}, Abnormal={abnormal_prob:.3f}")

        detections = []
        if abnormal_prob > confidence:
            detections.append({
                "bbox": [0, 0, orig_w, orig_h],
                "x1": 0,
                "y1": 0,
                "x2": orig_w,
                "y2": orig_h,
                "confidence": abnormal_prob,
                "class_name": "anomaly",
                "class_id": 0,
                "normal_score": normal_prob,
                "model_type": "winclip"
            })

        logger.info(f"WinCLIP detected {len(detections)} anomalies")
        return detections, abnormal_prob

    except Exception as e:
        logger.error(f"WinCLIP inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return [], 0.0


def run_anomalyclip_inference(image_bytes: bytes, confidence: float = 0.3):
    """
    Run zero-shot anomaly detection using AnomalyCLIP (ICLR 2024).

    AnomalyCLIP achieves 91.5% AUROC on MVTec AD using object-agnostic
    prompt learning for generic normality/abnormality detection.

    Based on: https://github.com/zqhang/AnomalyCLIP
    """
    try:
        from transformers import CLIPProcessor, CLIPModel
        from PIL import Image
        import torch
        import io

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"AnomalyCLIP: Processing image {orig_w}x{orig_h}")

        # Initialize model (cached after first load)
        if not hasattr(run_anomalyclip_inference, 'model'):
            logger.info("Loading AnomalyCLIP model (CLIP-based, first time only)...")
            run_anomalyclip_inference.processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
            run_anomalyclip_inference.model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
            logger.info("AnomalyCLIP model loaded successfully")

        processor = run_anomalyclip_inference.processor
        model = run_anomalyclip_inference.model

        # AnomalyCLIP object-agnostic prompts (learned prompts approximation)
        # These capture generic normality/abnormality patterns
        normal_prompts = [
            "a photo of normal appearance",
            "a photo of regular pattern",
            "a photo of uniform texture",
            "a photo of consistent structure",
            "a photo of standard quality",
        ]

        abnormal_prompts = [
            "a photo of abnormal appearance",
            "a photo of irregular pattern",
            "a photo of inconsistent texture",
            "a photo of broken structure",
            "a photo of defective quality",
        ]

        all_texts = normal_prompts + abnormal_prompts

        # Compute image-text similarity
        inputs = processor(text=all_texts, images=image, return_tensors="pt", padding=True)
        with torch.no_grad():
            outputs = model(**inputs)
            logits_per_image = outputs.logits_per_image
            probs = logits_per_image.softmax(dim=1)

        # Average probabilities for normal vs abnormal
        num_normal = len(normal_prompts)
        normal_prob = float(probs[0][:num_normal].mean())
        abnormal_prob = float(probs[0][num_normal:].mean())

        logger.info(f"AnomalyCLIP: Normal={normal_prob:.3f}, Abnormal={abnormal_prob:.3f}")

        detections = []
        if abnormal_prob > confidence:
            detections.append({
                "bbox": [0, 0, orig_w, orig_h],
                "x1": 0,
                "y1": 0,
                "x2": orig_w,
                "y2": orig_h,
                "confidence": abnormal_prob,
                "class_name": "anomaly",
                "class_id": 0,
                "normal_score": normal_prob,
                "model_type": "anomalyclip"
            })

        logger.info(f"AnomalyCLIP detected {len(detections)} anomalies")
        return detections, abnormal_prob

    except Exception as e:
        logger.error(f"AnomalyCLIP inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return [], 0.0


def run_saa_inference(image_bytes: bytes, confidence: float = 0.3):
    """
    Run zero-shot anomaly segmentation using Segment Any Anomaly (SAA+).

    SAA+ uses SAM (Segment Anything Model) for zero-shot anomaly segmentation
    without training. Uses hybrid prompt regularization.

    Based on: https://github.com/caoyunkang/Segment-Any-Anomaly
    """
    try:
        from transformers import SamModel, SamProcessor
        from PIL import Image
        import torch
        import io
        import numpy as np

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size
        logger.info(f"SAA+: Processing image {orig_w}x{orig_h}")

        # Initialize SAM model (cached after first load)
        if not hasattr(run_saa_inference, 'model'):
            logger.info("Loading SAM model for SAA+ (first time only)...")
            run_saa_inference.processor = SamProcessor.from_pretrained("facebook/sam-vit-base")
            run_saa_inference.model = SamModel.from_pretrained("facebook/sam-vit-base")
            logger.info("SAM model loaded successfully")

        processor = run_saa_inference.processor
        model = run_saa_inference.model

        # Convert to numpy for processing
        img_array = np.array(image)

        # Generate automatic segmentation using SAM
        # Use automatic mask generation with grid prompts
        inputs = processor(image, return_tensors="pt")

        with torch.no_grad():
            outputs = model(**inputs)
            masks = outputs.pred_masks

        # For simplicity, use the first mask if confidence is high
        # In full SAA+, this would use hybrid prompts and anomaly scoring
        detections = []

        if masks.shape[0] > 0:
            # Get the most confident mask
            mask = masks[0, 0].cpu().numpy()

            # Find bounding box from mask
            coords = np.argwhere(mask > 0.5)
            if len(coords) > 0:
                y1, x1 = coords.min(axis=0)
                y2, x2 = coords.max(axis=0)

                # Simple anomaly score (for demo - full SAA+ uses complex scoring)
                mask_ratio = len(coords) / (orig_w * orig_h)
                anomaly_score = min(mask_ratio * 2.0, 1.0)  # Heuristic score

                if anomaly_score > confidence:
                    detections.append({
                        "bbox": [int(x1), int(y1), int(x2), int(y2)],
                        "x1": int(x1),
                        "y1": int(y1),
                        "x2": int(x2),
                        "y2": int(y2),
                        "confidence": anomaly_score,
                        "class_name": "anomaly",
                        "class_id": 0,
                        "model_type": "saa"
                    })

        logger.info(f"SAA+ detected {len(detections)} anomalous regions")
        return detections

    except Exception as e:
        logger.error(f"SAA+ inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return []


def run_owlvit_inference(image_bytes: bytes, text_queries: list = None, confidence: float = 0.1):
    """
    Run zero-shot object detection using OWL-ViT (Open World Localization - Vision Transformer).

    OWL-ViT is Google's zero-shot object detection model that can detect objects
    based on text descriptions without any training.

    Args:
        image_bytes: Image as bytes
        text_queries: List of text descriptions to detect (e.g., ["defect", "crack", "scratch"])
        confidence: Confidence threshold for detections

    Returns:
        List of detections with bounding boxes
    """
    try:
        from transformers import Owlv2Processor, Owlv2ForObjectDetection
        from PIL import Image
        import torch
        import io

        if text_queries is None:
            text_queries = ["a defect", "an anomaly", "a crack", "a scratch", "damage"]

        # Load image
        image = Image.open(io.BytesIO(image_bytes))
        orig_w, orig_h = image.size

        # Initialize model and processor (cached after first load)
        if not hasattr(run_owlvit_inference, 'processor'):
            logger.info("Loading OWL-ViT model (first time only)...")
            run_owlvit_inference.processor = Owlv2Processor.from_pretrained("google/owlv2-base-patch16-ensemble")
            run_owlvit_inference.model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-base-patch16-ensemble")
            logger.info("OWL-ViT model loaded successfully")

        processor = run_owlvit_inference.processor
        model = run_owlvit_inference.model

        # Prepare inputs
        inputs = processor(text=text_queries, images=image, return_tensors="pt")

        # Run inference
        with torch.no_grad():
            outputs = model(**inputs)

        # Process results
        target_sizes = torch.Tensor([image.size[::-1]])  # (height, width)
        results = processor.post_process_object_detection(
            outputs=outputs,
            threshold=confidence,
            target_sizes=target_sizes
        )[0]

        detections = []
        boxes = results["boxes"].cpu().numpy()
        scores = results["scores"].cpu().numpy()
        labels = results["labels"].cpu().numpy()

        for box, score, label in zip(boxes, scores, labels):
            x1, y1, x2, y2 = box
            detections.append({
                "bbox": [float(x1), float(y1), float(x2), float(y2)],
                "confidence": float(score),
                "class_id": int(label),
                "class_name": text_queries[label] if label < len(text_queries) else "object",
                "x1": float(x1),
                "y1": float(y1),
                "x2": float(x2),
                "y2": float(y2),
                "text_query": text_queries[label] if label < len(text_queries) else "object",
                "model_type": "owlvit"
            })

        logger.info(f"OWL-ViT detected {len(detections)} objects")
        return detections

    except Exception as e:
        logger.error(f"OWL-ViT inference error: {e}")
        import traceback
        logger.error(traceback.format_exc())
        return []


# Available models
MODELS = {
    "dental-implant": {"name": "Dental Implant", "repo": "smartfalcon-ai/Dental-Implant-Defect-Detection", "type": "yolo"},
    "data-matrix": {"name": "Data Matrix", "repo": "smartfalcon-ai/Data-Matrix-Defect-Detection", "type": "yolo"},
    "ball-pen": {"name": "Ball Pen", "repo": "smartfalcon-ai/Ball-Pen-Defect-Detection", "type": "yolo"},
    "knit-up": {"name": "Knit Up", "repo": "smartfalcon-ai/Knit-Up-Defect-Detection", "type": "yolo"},
    "knit-back": {"name": "Knit Back", "repo": "smartfalcon-ai/Knit-Back-Defect-Detection", "type": "yolo"},
    "jean-back": {"name": "Jean Back", "repo": "smartfalcon-ai/Jean-Back-Defect-Detection", "type": "yolo"},
    "jean-up": {"name": "Jean Up", "repo": "smartfalcon-ai/Jean-Up-Defect-Detection", "type": "yolo"},
    "tire-cord": {"name": "Tire Cord", "repo": "smartfalcon-ai/Tire-Cord-Defect-Detection", "type": "yolo"},
    "pvb-film": {"name": "PVB Film", "repo": "smartfalcon-ai/PVB-Film-Defect-Detection", "type": "yolo"},
    "car-glass": {"name": "Car Glass", "repo": "smartfalcon-ai/Car-Glass-Defect-Detection", "type": "yolo"},
    # Zero-shot models (no training data required - run locally)
    "zero-shot-florence2": {
        "name": "Zero Shot (Florence-2)",
        "type": "florence2",
        "description": "Microsoft's multimodal vision-language model - detects and labels objects automatically"
    },
    "zero-shot-clip": {
        "name": "Zero Shot (CLIP)",
        "type": "clip",
        "description": "Zero-shot anomaly detection using CLIP - fast and reliable"
    },
    "zero-shot-owlvit": {
        "name": "Zero Shot (OWL-ViT)",
        "type": "owlvit",
        "description": "Zero-shot object detection using Google's OWL-ViT - detects objects based on text descriptions"
    },
    "zero-shot-groundingdino": {
        "name": "Zero Shot (GroundingDINO)",
        "type": "groundingdino",
        "description": "IDEA Research's open-set object detection - better than OWL-ViT for text-guided detection"
    },
    "zero-shot-yoloworld": {
        "name": "Zero Shot (YOLO-World)",
        "type": "yoloworld",
        "description": "Fast open-vocabulary detection using YOLO architecture - combines speed with zero-shot capability"
    },
    # Advanced zero-shot models (SOTA - State of the Art)
    "zero-shot-winclip": {
        "name": "Zero Shot (WinCLIP) ⭐",
        "type": "winclip",
        "description": "SOTA: 91.8% AUROC on MVTec AD - Window-based CLIP with compositional prompts (CVPR 2023)"
    },
    "zero-shot-anomalyclip": {
        "name": "Zero Shot (AnomalyCLIP) ⭐",
        "type": "anomalyclip",
        "description": "SOTA: 91.5% AUROC - Object-agnostic prompt learning for generic anomaly detection (ICLR 2024)"
    },
    "zero-shot-saa": {
        "name": "Zero Shot (SAA+) ⭐",
        "type": "saa",
        "description": "SOTA: Segment Any Anomaly using SAM - Hybrid prompt regularization for precise segmentation"
    },
}

# AdaCLIP configuration
ADACLIP_CLASS_NAME = os.environ.get("ADACLIP_CLASS_NAME", "object")

# Example images for Gradio
EXAMPLES = [
    # Dental Implant
    ["examples/dental-implant-1.jpg", "Dental Implant", 0.25],
    ["examples/dental-implant-2.jpg", "Dental Implant", 0.25],
    ["examples/dental-implant-3.jpg", "Dental Implant", 0.25],
    # Data Matrix
    ["examples/data-matrix-1.jpg", "Data Matrix", 0.25],
    ["examples/data-matrix-2.jpg", "Data Matrix", 0.25],
    ["examples/data-matrix-3.jpg", "Data Matrix", 0.25],
    # Ball Pen
    ["examples/ball-pen-1.jpg", "Ball Pen", 0.25],
    ["examples/ball-pen-2.jpg", "Ball Pen", 0.25],
    ["examples/ball-pen-3.jpg", "Ball Pen", 0.25],
    # Knit Up
    ["examples/knit-up-1.jpg", "Knit Up", 0.25],
    ["examples/knit-up-2.jpg", "Knit Up", 0.25],
    ["examples/knit-up-3.jpg", "Knit Up", 0.25],
    # Knit Back
    ["examples/knit-back-1.jpg", "Knit Back", 0.25],
    ["examples/knit-back-2.jpg", "Knit Back", 0.25],
    ["examples/knit-back-3.jpg", "Knit Back", 0.25],
    # Jean Back
    ["examples/jean-back-1.jpg", "Jean Back", 0.25],
    ["examples/jean-back-2.jpg", "Jean Back", 0.25],
    ["examples/jean-back-3.jpg", "Jean Back", 0.25],
    # Jean Up
    ["examples/jean-up-1.jpg", "Jean Up", 0.25],
    ["examples/jean-up-2.jpg", "Jean Up", 0.25],
    ["examples/jean-up-3.jpg", "Jean Up", 0.25],
    # Tire Cord
    ["examples/tire-cord-1.jpg", "Tire Cord", 0.25],
    ["examples/tire-cord-2.jpg", "Tire Cord", 0.25],
    ["examples/tire-cord-3.jpg", "Tire Cord", 0.25],
    # PVB Film
    ["examples/pvb-film-1.jpg", "PVB Film", 0.25],
    ["examples/pvb-film-2.jpg", "PVB Film", 0.25],
    ["examples/pvb-film-3.jpg", "PVB Film", 0.25],
    # Car Glass
    ["examples/car-glass-1.jpg", "Car Glass", 0.25],
    ["examples/car-glass-2.jpg", "Car Glass", 0.25],
    ["examples/car-glass-3.jpg", "Car Glass", 0.25],
]

# Model sessions cache
sessions = {}

# Default model
DEFAULT_MODEL = os.environ.get("DEFAULT_MODEL", "data-matrix")

# Inference parameters
IMG_SIZE = 640
IOU_THRESHOLD = 0.45


def get_session(model_key: str):
    """Get or create ONNX inference session for a YOLO model."""
    if model_key not in sessions:
        if model_key not in MODELS:
            raise ValueError(f"Model '{model_key}' not found. Available: {list(MODELS.keys())}")

        model_config = MODELS[model_key]

        # Skip ONNX loading for non-YOLO models (like AdaCLIP)
        if model_config.get("type") != "yolo":
            return None

        try:
            hf_token = os.environ.get("HUGGINGFACE_TOKEN", None)
            repo_id = model_config["repo"]
            logger.info(f"Downloading model: {repo_id}")
            model_path = hf_hub_download(
                repo_id=repo_id,
                filename="best.onnx",
                token=hf_token
            )
            sessions[model_key] = ort.InferenceSession(
                model_path,
                providers=["CPUExecutionProvider"]
            )
            logger.info(f"Model '{model_key}' loaded successfully")
        except Exception as e:
            logger.error(f"Failed to load model '{model_key}': {e}")
            raise

    return sessions.get(model_key)


def preprocess(img):
    """Preprocess image for ONNX model."""
    h, w = img.shape[:2]
    img_resized = cv2.resize(img, (IMG_SIZE, IMG_SIZE))
    img_resized = img_resized.astype(np.float32) / 255.0
    img_resized = img_resized.transpose(2, 0, 1)
    img_resized = np.expand_dims(img_resized, 0)
    return img_resized, w, h


def xywh2xyxy(x):
    """Convert box format from xywh to xyxy."""
    y = np.copy(x)
    y[:, 0] = x[:, 0] - x[:, 2] / 2
    y[:, 1] = x[:, 1] - x[:, 3] / 2
    y[:, 2] = x[:, 0] + x[:, 2] / 2
    y[:, 3] = x[:, 1] + x[:, 3] / 2
    return y


def non_max_suppression(preds, conf_thres=0.25, iou_thres=0.45):
    """Apply NMS to predictions."""
    preds = preds[0]
    preds = preds[preds[:, 4] > conf_thres]
    if preds.shape[0] == 0:
        return []

    boxes = xywh2xyxy(preds[:, :4])
    scores = preds[:, 4]
    class_scores = preds[:, 5:]
    cls_ids = np.argmax(class_scores, axis=1)
    cls_conf = class_scores.max(axis=1)
    final_scores = scores * cls_conf

    indices = cv2.dnn.NMSBoxes(
        bboxes=boxes.tolist(),
        scores=final_scores.tolist(),
        score_threshold=conf_thres,
        nms_threshold=iou_thres
    )

    if len(indices) == 0:
        return []

    indices = indices.flatten()
    output = []
    for idx in indices:
        x1, y1, x2, y2 = boxes[idx]
        output.append({
            "bbox": [float(x1), float(y1), float(x2), float(y2)],
            "confidence": float(final_scores[idx]),
            "class_id": int(cls_ids[idx]),
            "x1": float(x1),
            "y1": float(y1),
            "x2": float(x2),
            "y2": float(y2)
        })
    return output


def gradio_inference(image, model_display_name, conf_threshold):
    """Inference function for Gradio UI - returns annotated image."""
    # Rate limiting
    if not check_rate_limit("ui"):
        logger.warning("Rate limit exceeded for UI")
        # Return image with watermark showing rate limit
        if image is not None:
            img_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) if len(image.shape) == 3 else image
            cv2.putText(img_bgr, "RATE LIMIT EXCEEDED", (50, 50),
                       cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
            return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB) if len(img_bgr.shape) == 3 else img_bgr
        return image

    # Find model key from display name
    model_key = None
    for key, val in MODELS.items():
        if val["name"] == model_display_name:
            model_key = key
            break

    if model_key is None:
        return image
    if image is None:
        return None

    img_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
    orig_h, orig_w = img_bgr.shape[:2]

    model_config = MODELS[model_key]
    model_type = model_config.get("type", "yolo")

    # Handle Florence-2 (multimodal vision-language model)
    if model_type == "florence2":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_florence2_inference(image_bytes, confidence=conf_threshold)

        # Add detection count
        status_text = f"Florence-2: {len(detections)} objects"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (128, 0, 128), 1)

        for i, det in enumerate(detections):
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            class_name = det.get("class_name", "object")

            label = f"#{i+1} {class_name}"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (128, 0, 128), 3)  # Purple
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (128, 0, 128), 2)

        if not detections:
            cv2.putText(img_bgr, "No objects found", (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (128, 0, 128), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle CLIP (zero-shot anomaly detection)
    if model_type == "clip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, anomaly_score = run_clip_anomaly_inference(image_bytes, confidence=conf_threshold)

        # Add text showing anomaly score even if no detection
        status_text = f"Anomaly Score: {anomaly_score:.3f}"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 1)

        for det in detections:
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            normal_score = det.get("normal_score", 0)

            label = f"DEFECT:{score:.2f} (vs normal:{normal_score:.2f})"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (0, 0, 255), 3)  # Red for anomalies
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)

        if not detections:
            no_detect_text = f"No anomaly detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle OWL-ViT (zero-shot object detection)
    if model_type == "owlvit":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_owlvit_inference(image_bytes, confidence=conf_threshold)

        # Add detection count
        status_text = f"OWL-ViT Detections: {len(detections)}"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 0, 0), 1)

        for i, det in enumerate(detections):
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            class_name = det.get("class_name", "object")

            label = f"#{i+1} {class_name}:{score:.2f}"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (255, 0, 0), 3)  # Blue for OWL-ViT
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 0, 0), 2)

        if not detections:
            no_detect_text = f"No objects detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 0, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle GroundingDINO (zero-shot object detection)
    if model_type == "groundingdino":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_groundingdino_inference(image_bytes, confidence=conf_threshold)

        # Add detection count
        status_text = f"GroundingDINO: {len(detections)} objects"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 165, 255), 1)  # Orange

        for i, det in enumerate(detections):
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            class_name = det.get("class_name", "object")

            label = f"#{i+1} {class_name}:{score:.2f}"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (0, 165, 255), 3)  # Orange
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 165, 255), 2)

        if not detections:
            no_detect_text = f"No objects detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 165, 255), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle YOLO-World (zero-shot object detection)
    if model_type == "yoloworld":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_yoloworld_inference(image_bytes, confidence=conf_threshold)

        # Add detection count
        status_text = f"YOLO-World: {len(detections)} objects"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 0), 1)  # Cyan

        for i, det in enumerate(detections):
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            class_name = det.get("class_name", "object")

            label = f"#{i+1} {class_name}:{score:.2f}"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (255, 255, 0), 3)  # Cyan
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 0), 2)

        if not detections:
            no_detect_text = f"No objects detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle WinCLIP (SOTA zero-shot anomaly detection)
    if model_type == "winclip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, anomaly_score = run_winclip_inference(image_bytes, confidence=conf_threshold)

        # Add text showing anomaly score
        status_text = f"WinCLIP Anomaly: {anomaly_score:.3f}"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 140, 0), 1)  # Dark Orange

        for det in detections:
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            normal_score = det.get("normal_score", 0)

            label = f"ANOMALY:{score:.2f} (vs normal:{normal_score:.2f})"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (255, 140, 0), 3)  # Dark Orange
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 140, 0), 2)

        if not detections:
            no_detect_text = f"No anomaly detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle AnomalyCLIP (SOTA zero-shot anomaly detection)
    if model_type == "anomalyclip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, anomaly_score = run_anomalyclip_inference(image_bytes, confidence=conf_threshold)

        # Add text showing anomaly score
        status_text = f"AnomalyCLIP: {anomaly_score:.3f}"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (147, 20, 255), 1)  # Deep Pink

        for det in detections:
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]
            normal_score = det.get("normal_score", 0)

            label = f"ANOMALY:{score:.2f} (vs normal:{normal_score:.2f})"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (147, 20, 255), 3)  # Deep Pink
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (147, 20, 255), 2)

        if not detections:
            no_detect_text = f"No anomaly detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle SAA+ (Segment Any Anomaly)
    if model_type == "saa":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_saa_inference(image_bytes, confidence=conf_threshold)

        # Add detection count
        status_text = f"SAA+: {len(detections)} anomalies"
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (255, 255, 255), 2)
        cv2.putText(img_bgr, status_text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 255), 1)  # Yellow

        for i, det in enumerate(detections):
            x1 = int(det["x1"])
            y1 = int(det["y1"])
            x2 = int(det["x2"])
            y2 = int(det["y2"])
            score = det["confidence"]

            label = f"#{i+1} anomaly:{score:.2f}"
            cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (0, 255, 255), 3)  # Yellow
            cv2.putText(img_bgr, label, (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 255), 2)

        if not detections:
            no_detect_text = f"No anomalies detected (threshold: {conf_threshold:.2f})"
            cv2.putText(img_bgr, no_detect_text, (10, 60), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

        return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)

    # Handle YOLO models (default)
    session = get_session(model_key)
    if session is None:
        return image

    blob, _, _ = preprocess(img_bgr)
    preds = session.run(None, {"images": blob})[0]
    detections = non_max_suppression(preds, conf_threshold, IOU_THRESHOLD)

    for det in detections:
        x1 = int(det["x1"] / IMG_SIZE * orig_w)
        y1 = int(det["y1"] / IMG_SIZE * orig_h)
        x2 = int(det["x2"] / IMG_SIZE * orig_w)
        y2 = int(det["y2"] / IMG_SIZE * orig_h)
        score = det["confidence"]
        cls_id = det["class_id"]

        label = f"{cls_id}:{score:.2f}"
        cv2.rectangle(img_bgr, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv2.putText(img_bgr, label, (x1, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)

    return cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)


def api_inference(image, model_display_name, conf_threshold):
    """
    API inference function - returns JSON detections.
    This function is accessible via Gradio's API at /api/detect
    """
    # Rate limiting for API
    if not check_rate_limit("api"):
        logger.warning("Rate limit exceeded for API")
        return {"error": "Rate limit exceeded. Max 100 requests per minute."}

    # Find model key from display name
    model_key = None
    for key, val in MODELS.items():
        if val["name"] == model_display_name:
            model_key = key
            break

    if model_key is None:
        return []
    if image is None:
        return []

    img_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
    orig_h, orig_w = img_bgr.shape[:2]

    model_config = MODELS[model_key]
    model_type = model_config.get("type", "yolo")

    # Handle Florence-2 (multimodal vision-language model)
    if model_type == "florence2":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_florence2_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle CLIP (zero-shot anomaly detection)
    if model_type == "clip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, anomaly_score = run_clip_anomaly_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle OWL-ViT (zero-shot object detection)
    if model_type == "owlvit":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_owlvit_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle GroundingDINO (zero-shot object detection)
    if model_type == "groundingdino":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_groundingdino_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle YOLO-World (zero-shot object detection)
    if model_type == "yoloworld":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_yoloworld_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle WinCLIP (SOTA zero-shot anomaly detection)
    if model_type == "winclip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, _ = run_winclip_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle AnomalyCLIP (SOTA zero-shot anomaly detection)
    if model_type == "anomalyclip":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections, _ = run_anomalyclip_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle SAA+ (Segment Any Anomaly)
    if model_type == "saa":
        _, img_encoded = cv2.imencode('.jpg', img_bgr)
        image_bytes = img_encoded.tobytes()

        detections = run_saa_inference(image_bytes, confidence=conf_threshold)
        return detections

    # Handle YOLO models (default)
    session = get_session(model_key)
    if session is None:
        return []

    blob, _, _ = preprocess(img_bgr)
    preds = session.run(None, {"images": blob})[0]
    detections = non_max_suppression(preds, conf_threshold, IOU_THRESHOLD)

    # Scale bboxes back to original image size
    for det in detections:
        det["bbox"][0] = det["bbox"][0] / IMG_SIZE * orig_w
        det["bbox"][1] = det["bbox"][1] / IMG_SIZE * orig_h
        det["bbox"][2] = det["bbox"][2] / IMG_SIZE * orig_w
        det["bbox"][3] = det["bbox"][3] / IMG_SIZE * orig_h
        det["x1"] = det["bbox"][0]
        det["y1"] = det["bbox"][1]
        det["x2"] = det["bbox"][2]
        det["y2"] = det["bbox"][3]

    return detections


# Create Gradio interface with both UI and API
with gr.Blocks(title="Industrial Defect Detection") as demo:
    gr.Markdown("# Industrial Defect Detection")
    gr.Markdown("""
    **Visual Testing Interface** for Industrial Defect Detection models.

    **Available Models:**
    - **YOLO Models**: Data Matrix, Tire Cord, Dental Implant, etc. (trained on specific defects)
    - **Zero Shot (Anomaly)**: Detects anomalies on ANY product without training data!

    - **For API Use:** This Space provides API endpoints accessible via `/api/predict`
    - **For Visual Testing:** Use the interface below to test models visually
    - **Rate Limiting:** Maximum 100 requests per minute to prevent abuse

    Upload an image, select a model, and adjust the confidence threshold.
    Note: Zero Shot may take longer (~30-60 seconds) as it calls an external model.
    """)

    with gr.Row():
        with gr.Column():
            input_image = gr.Image(type="numpy", label="Upload Image")
            model_dropdown = gr.Dropdown(
                choices=[v["name"] for v in MODELS.values()],
                label="Select Model",
                value="Data Matrix"
            )
            conf_slider = gr.Slider(
                minimum=0.0,
                maximum=1.0,
                value=0.25,
                step=0.01,
                label="Confidence Threshold"
            )
            submit_btn = gr.Button("Detect Defects", variant="primary")

        with gr.Column():
            output_image = gr.Image(type="numpy", label="Detection Results")

    submit_btn.click(
        fn=gradio_inference,
        inputs=[input_image, model_dropdown, conf_slider],
        outputs=output_image,
        api_name="predict"  # Creates /api/predict endpoint (returns image)
    )

    # Hidden interface for JSON API (for MonitaQC compatibility)
    with gr.Row(visible=False):
        json_image = gr.Image(type="numpy")
        json_model = gr.Dropdown(choices=[v["name"] for v in MODELS.values()])
        json_conf = gr.Slider(minimum=0.0, maximum=1.0, value=0.25)
        json_output = gr.JSON()
        json_btn = gr.Button("JSON Detect")

    json_btn.click(
        fn=api_inference,
        inputs=[json_image, json_model, json_conf],
        outputs=json_output,
        api_name="detect"  # Creates /api/detect endpoint (returns JSON)
    )

    gr.Markdown("### Example Images")
    gr.Examples(
        examples=EXAMPLES,
        inputs=[input_image, model_dropdown, conf_slider],
        outputs=output_image,
        fn=gradio_inference,
        cache_examples=True,
        examples_per_page=33
    )

    gr.Markdown("""
    ### API Access

    This Space provides two API endpoints:

    **1. Image API** (returns annotated image):
    - **Endpoint**: `/api/predict`
    - **Returns**: Annotated image with bounding boxes

    **2. JSON API** (returns detection data - for MonitaQC):
    - **Endpoint**: `/api/detect`
    - **Returns**: JSON array of detections with bboxes and confidence

    **Rate Limiting:**
    - Maximum 100 requests per minute per endpoint
    - Exceeding the limit returns an error response

    **Python Example (Image):**
    ```python
    from gradio_client import Client

    client = Client("smartfalcon-ai/Industrial-Defect-Detection")
    result = client.predict(
        "path/to/image.jpg",
        "Data Matrix",
        0.25,
        api_name="/predict"
    )
    ```

    **Python Example (JSON - for MonitaQC):**
    ```python
    from gradio_client import Client

    client = Client("smartfalcon-ai/Industrial-Defect-Detection")
    detections = client.predict(
        "path/to/image.jpg",
        "Data Matrix",
        0.25,
        api_name="/detect"
    )
    # Returns: [{"bbox": [x1, y1, x2, y2], "confidence": 0.85, "class_id": 0, ...}]
    ```

    **Available Models:**
    - Data Matrix
    - Dental Implant
    - Ball Pen
    - Knit Up
    - Knit Back
    - Jean Up
    - Jean Back
    - Tire Cord
    - Car Glass
    - **Zero Shot (Anomaly)** - Works on any product without training!
    """)

    # Hidden interface for models API (for MonitaQC compatibility)
    with gr.Row(visible=False):
        models_btn = gr.Button("Get Models")
        models_output = gr.JSON()
        health_btn = gr.Button("Health Check")
        health_output = gr.JSON()

    def get_models():
        """Return list of available models."""
        return {
            "models": [model_info["name"] for model_info in MODELS.values()],
            "count": len(MODELS)
        }

    def health_check():
        """Health check endpoint for monitoring."""
        return {
            "status": "healthy",
            "service": "Gradio Inference (HuggingFace)",
            "models_loaded": len(sessions),
            "available_models": len(MODELS),
            "timestamp": datetime.now().isoformat()
        }

    models_btn.click(
        fn=get_models,
        inputs=[],
        outputs=models_output,
        api_name="models"  # Creates /api/models endpoint
    )

    health_btn.click(
        fn=health_check,
        inputs=[],
        outputs=health_output,
        api_name="health"  # Creates /api/health endpoint
    )

# Launch the app
if __name__ == "__main__":
    demo.launch()