{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":91249,"databundleVersionId":11294684,"sourceType":"competition"},{"sourceId":11932060,"sourceType":"datasetVersion","datasetId":7501747},{"sourceId":327336,"sourceType":"modelInstanceVersion","modelInstanceId":274744,"modelId":295634}],"dockerImageVersionId":30919,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# !tar xfz /kaggle/input/ultralytics-for-offline-install/archive.tar.gz\n# !pip install --no-index --find-links=./packages -q ultralytics\n# !rm -rf ./packages\n# print(\"package installed ...........................\")\n\n!cp -r /kaggle/input/mhafyolo/pytorch/default/1/MHAF-YOLO-main /kaggle/working/\nprint('PIP INSTALL OK!!!')\n\nimport os\nimport numpy as np\nimport pandas as pd\nfrom PIL import Image\nimport torch\nimport cv2\nfrom tqdm.notebook import tqdm\nfrom pathlib import Path\nimport logging\nimport sys\ncurrent_dir = Path.cwd()\nprint(\"this_dir:\", current_dir)\n\ntarget_dir = Path(\"/kaggle/working/MHAF-YOLO-main\") \nos.chdir(target_dir)\n\nfrom ultralytics import YOLOv10\nimport threading\nimport time\nfrom contextlib import nullcontext\nfrom concurrent.futures import ThreadPoolExecutor\nfrom ultralytics.utils.ops import non_max_suppression\n\n\n# Set random seed for reproducibility\nnp.random.seed(42)\ntorch.manual_seed(42)\n\n# Define paths\ndata_path = \"/kaggle/input/byu-locating-bacterial-flagellar-motors-2025/\"\ntest_dir = os.path.join(data_path, \"test\")\nsubmission_path = \"/kaggle/working/submission.csv\"\n\n# Model path - adjust if your best model is saved in a different location\nmodel_path = \"/kaggle/input/mahf-yolo-train/mayolov2f.pt\"\n\n# Detection parameters\nCONFIDENCE_THRESHOLD = 0.9  # Lower threshold to catch more potential motors\nMAX_DETECTIONS_PER_TOMO = 1  # Keep track of top N detections per tomogram\nNMS_IOU_THRESHOLD = 0.3  # Non-maximum suppression threshold for 3D clustering\nCONCENTRATION = 1 # ONLY PROCESS 1/20 slices for fast submission\nSIZE = 1024\n\n# GPU profiling context manager\nclass GPUProfiler:\n    def __init__(self, name):\n        self.name = name\n        self.start_time = None\n        \n    def __enter__(self):\n        if torch.cuda.is_available():\n            torch.cuda.synchronize()\n        self.start_time = time.time()\n        return self\n        \n    def __exit__(self, *args):\n        if torch.cuda.is_available():\n            torch.cuda.synchronize()\n        elapsed = time.time() - self.start_time\n        print(f\"[PROFILE] {self.name}: {elapsed:.3f}s\")\n\n# Check GPU availability and set up optimizations\ndevice = 'cuda:0' if torch.cuda.is_available() else 'cpu'\n# device = ['cuda:0', 'cuda:1']\nBATCH_SIZE = 8  # Default batch size, will be adjusted dynamically if GPU available\n\nif device.startswith('cuda'):\n    # Set CUDA optimization flags\n    torch.backends.cudnn.benchmark = True\n    torch.backends.cudnn.deterministic = False\n    torch.backends.cuda.matmul.allow_tf32 = True  # Allow TF32 on Ampere GPUs\n    torch.backends.cudnn.allow_tf32 = True\n    \n    # Print GPU info\n    gpu_name = torch.cuda.get_device_name(0)\n    gpu_mem = torch.cuda.get_device_properties(0).total_memory / 1e9  # Convert to GB\n    print(f\"Using GPU: {gpu_name} with {gpu_mem:.2f} GB memory\")\n    \n    # Get available GPU memory and set batch size accordingly\n    free_mem = gpu_mem - torch.cuda.memory_allocated(0) / 1e9\n    BATCH_SIZE = max(8, min(32, int(free_mem * 4)))  # 4 images per GB as rough estimate\n    print(f\"Dynamic batch size set to {BATCH_SIZE} based on {free_mem:.2f}GB free memory\")\nelse:\n    print(\"GPU not available, using CPU\")\n    BATCH_SIZE = 8  # Reduce batch size for CPU\n\n# Add this IOU function definition\ndef iou(boxA, boxB):\n    # Determine the coordinates of the intersection rectangle\n    # boxA and boxB are expected to be (x1, y1, x2, y2, ...)\n    xA = max(boxA[0], boxB[0])\n    yA = max(boxA[1], boxB[1])\n    xB = min(boxA[2], boxB[2])\n    yB = min(boxA[3], boxB[3])\n\n    # Compute the area of intersection rectangle\n    interArea = max(0, xB - xA + 1) * max(0, yB - yA + 1)\n\n    # Compute the area of both the prediction and ground-truth rectangles\n    boxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)\n    boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)\n\n    # Compute the intersection over union\n    unionArea = float(boxAArea + boxBArea - interArea)\n    if unionArea == 0:\n        return 0.0 # Avoid division by zero\n    iou = interArea / unionArea\n\n    # Return the intersection over union value\n    return iou\n\ndef weighted_box_fusion(boxes, iou_threshold=0.4):\n    \"\"\"Applies Weighted Box Fusion to combine overlapping bounding boxes.\"\"\"\n    fused_boxes = []\n    # print(type(boxes))\n    used = [False] * len(boxes)\n\n    for i in range(len(boxes)):\n        if used[i]:\n            continue\n        \n        similar_boxes = [boxes[i]]\n        used[i] = True\n\n        for j in range(i + 1, len(boxes)):\n            if used[j]:\n                continue\n\n            if iou(boxes[i], boxes[j]) > iou_threshold:\n                similar_boxes.append(boxes[j])\n                used[j] = True\n\n        # Compute weighted average for the final box\n        # similar_boxes = np.array(similar_boxes)\n        # similar_boxes = similar_boxes.cpu().numpy()\n        \n        # similar_boxes = np.array([t.cpu() for t in similar_boxes])\n        similar_boxes_np = np.array(similar_boxes)\n        # print(\"similar_boxes: \",similar_boxes)\n        confidences = similar_boxes_np[:, 4]\n        weights = confidences / confidences.sum()\n\n        fused_x1 = np.sum(similar_boxes_np[:, 0] * weights)\n        fused_y1 = np.sum(similar_boxes_np[:, 1] * weights)\n        fused_x2 = np.sum(similar_boxes_np[:, 2] * weights)\n        fused_y2 = np.sum(similar_boxes_np[:, 3] * weights)\n        fused_confidence = np.mean(confidences) # You can keep np.mean or change to np.max\n\n        fused_boxes.append((fused_x1, fused_y1, fused_x2, fused_y2, fused_confidence))\n\n    return fused_boxes\n\ndef predict_ensemble_tta(single_model, image_np, device, img_size):\n    \"\"\"\n    For a 640x640 numpy image:\n    - Multiple models (list of models)\n    - Multiple TTA (original, hflip, vflip, rot90)\n    Do the NMS for the last time\n    Return: [K,6] => x1,y1,x2,y2,conf,cls\n    \"\"\"\n    all_boxes = []\n    all_confs = []\n    all_clss = []\n\n    def do_infer(img_tta, invert_func):\n        #for m in models:\n            res = single_model(img_tta, \n                            imgsz=img_size, \n                            # conf=conf_thres,\n                            device=device, \n                            verbose=False)\n            # res = single_model(img_tta,verbose=False)\n            for r in res:\n                # print('res box:', r.boxes)\n                boxes = r.boxes\n                if boxes is None or len(boxes)==0:\n                    # print('predict box is none!')\n                    continue\n                xyxy = boxes.xyxy.cpu().numpy()\n                confs = boxes.conf.cpu().numpy()\n                clss = boxes.cls.cpu().numpy().astype(int)\n                # Inverse transformation\n                xyxy_orig = invert_func(xyxy)\n                all_boxes.append(xyxy_orig)\n                all_confs.append(confs)\n                all_clss.append(clss)\n\n    # Master drawing\n    do_infer(image_np, invert_func=lambda x: x)\n\n    # Horizontal flip\n    img_hflip = cv2.flip(image_np, 1)\n    def invert_hflip(xyxy):\n        new_ = xyxy.copy()\n        x1 = img_size - xyxy[:,2]\n        x2 = img_size - xyxy[:,0]\n        new_[:,0] = x1\n        new_[:,2] = x2\n        return new_\n    do_infer(img_hflip, invert_func=invert_hflip)\n\n    # Vertical flip\n    img_vflip = cv2.flip(image_np, 0)\n    def invert_vflip(xyxy):\n        new_ = xyxy.copy()\n        y1 = img_size - xyxy[:,3]\n        y2 = img_size - xyxy[:,1]\n        new_[:,1] = y1\n        new_[:,3] = y2\n        return new_\n    do_infer(img_vflip, invert_func=invert_vflip)\n\n    # Rotate 90 degrees (clockwise)\n    img_rot90 = cv2.rotate(image_np, cv2.ROTATE_90_CLOCKWISE)\n    def invert_rot90(xyxy):\n        new_ = xyxy.copy()\n        # (x,y)->(y,640-x)\n        # Inverse transform (x',y')->(640-y',x')\n        x1_old,y1_old = xyxy[:,0], xyxy[:,1]\n        x2_old,y2_old = xyxy[:,2], xyxy[:,3]\n        X1 = img_size - y2_old\n        Y1 = x1_old\n        X2 = img_size - y1_old\n        Y2 = x2_old\n        new_[:,0] = X1\n        new_[:,1] = Y1\n        new_[:,2] = X2\n        new_[:,3] = Y2\n        return new_\n    do_infer(img_rot90, invert_func=invert_rot90)\n\n    if len(all_boxes)==0:\n        # print('all boxes is None!')\n        return None\n\n    boxes_cat = np.concatenate(all_boxes, axis=0)\n    confs_cat = np.concatenate(all_confs, axis=0)\n    clss_cat  = np.concatenate(all_clss, axis=0)\n    # Prepare data for weighted_box_fusion. WBF expects a list of individual detections.\n    # Each detection for WBF should be a tuple/list-like (x1, y1, x2, y2, confidence, [optional class]).\n    detections_for_wbf = []\n    for i in range(boxes_cat.shape[0]):\n        # Append as a list of 6 elements (xyxy, conf, cls) to detections_for_wbf\n        detections_for_wbf.append([boxes_cat[i,0], boxes_cat[i,1], boxes_cat[i,2], boxes_cat[i,3], confs_cat[i], clss_cat[i]])\n\n    # Perform Weighted Box Fusion\n    # weighted_box_fusion returns a list of (fused_x1, fused_y1, fused_x2, fused_y2, fused_confidence) tuples\n    fused_boxes = weighted_box_fusion(detections_for_wbf, iou_threshold=0.5)\n\n    if not fused_boxes: # If WBF resulted in no boxes\n        return None\n\n    # Select the single best detection after WBF (highest confidence)\n    best_fused_detection_tuple = max(fused_boxes, key=lambda x: x[4]) # x[4] is confidence\n\n    # CRITICAL: Return as a LIST of 5 elements, to be consistent with make_predict's expectation\n    return list(best_fused_detection_tuple) # Convert the tuple into a list of its elements\n\n\ndef make_predict(sub_path, model, device, img_size):\n    res_list = []\n    for img_file_path in sub_path:\n        img_np = cv2.imread(img_file_path)\n\n        # Handle cases where image fails to load\n        if img_np is None:\n            logger.warning(f\"Could not read image {img_file_path}. Appending zero-confidence placeholder.\")\n            # Append a LIST of 5 zeros/floats, matching the format of an actual detection list\n            res_list.append([0.0, 0.0, 0.0, 0.0, 0.0])\n            continue # Move to the next image\n\n        # Perform TTA and WBF inference\n        res_detection_list = predict_ensemble_tta(model, img_np, device, img_size=img_size)\n\n        # Handle case where predict_ensemble_tta returns None (no detections found after TTA/WBF)\n        if res_detection_list is None:\n            # Append a LIST of 5 zeros/floats as placeholder\n            res_list.append([0.0, 0.0, 0.0, 0.0, 0.0])\n        else:\n            # predict_ensemble_tta now returns a LIST of 5 floats: [x1,y1,x2,y2,confidence]\n            res_list.append(res_detection_list)\n            \n    return res_list\n\n\ndef normalize_slice(slice_data):\n    \"\"\"\n    Normalize slice data using 2nd and 98th percentiles for better contrast\n    \"\"\"\n    p2 = np.percentile(slice_data, 2)\n    p98 = np.percentile(slice_data, 98)\n    clipped_data = np.clip(slice_data, p2, p98)\n    normalized = 255 * (clipped_data - p2) / (p98 - p2)\n    return np.uint8(normalized)\n\ndef preload_image_batch(file_paths):\n    \"\"\"Preload a batch of images to CPU memory\"\"\"\n    images = []\n    for path in file_paths:\n        img = cv2.imread(path)\n        if img is None:\n            # Try with PIL as fallback\n            img = np.array(Image.open(path))\n        images.append(img)\n    return images\n\ndef process_tomogram(tomo_id, model, index=0, total=1,SIZE=SIZE):\n    \"\"\"\n    Process a single tomogram and return the most confident motor detection\n    \"\"\"\n    print(f\"Processing tomogram {tomo_id} ({index}/{total})\")\n    \n    # Get all slice files for this tomogram\n    tomo_dir = os.path.join(test_dir, tomo_id)\n    slice_files = sorted([f for f in os.listdir(tomo_dir) if f.endswith('.jpg')])\n    \n    # Apply CONCENTRATION to reduce the number of slices processed\n    # This will process approximately CONCENTRATION fraction of all slices\n    selected_indices = np.linspace(0, len(slice_files)-1, int(len(slice_files) * CONCENTRATION))\n    selected_indices = np.round(selected_indices).astype(int)\n    slice_files = [slice_files[i] for i in selected_indices]\n    \n    print(f\"Processing {len(slice_files)} out of {len(os.listdir(tomo_dir))} slices based on CONCENTRATION={CONCENTRATION}\")\n    \n    # Create a list to store all detections\n    all_detections = []\n    \n    # Create CUDA streams for parallel processing if using GPU\n    if device.startswith('cuda'):\n        streams = [torch.cuda.Stream() for _ in range(min(8, BATCH_SIZE))]\n    else:\n        streams = [None]\n    \n    # Variables for preloading\n    next_batch_thread = None\n    next_batch_images = None\n    \n    # Process slices in batches\n    for batch_start in range(0, len(slice_files), BATCH_SIZE):\n        # Wait for previous preload thread if it exists\n        if next_batch_thread is not None:\n            next_batch_thread.join()\n            next_batch_images = None\n            \n        batch_end = min(batch_start + BATCH_SIZE, len(slice_files))\n        batch_files = slice_files[batch_start:batch_end]\n        \n        # Start preloading next batch\n        next_batch_start = batch_end\n        next_batch_end = min(next_batch_start + BATCH_SIZE, len(slice_files))\n        next_batch_files = slice_files[next_batch_start:next_batch_end] if next_batch_start < len(slice_files) else []\n        \n        if next_batch_files:\n            next_batch_paths = [os.path.join(tomo_dir, f) for f in next_batch_files]\n            next_batch_thread = threading.Thread(target=preload_image_batch, args=(next_batch_paths,))\n            next_batch_thread.start()\n        else:\n            next_batch_thread = None\n        \n        # Split batch across streams for parallel processing\n        sub_batches = np.array_split(batch_files, len(streams))\n        sub_batch_results = []\n        \n        for i, sub_batch in enumerate(sub_batches):\n            if len(sub_batch) == 0:\n                continue\n                \n            stream = streams[i % len(streams)]\n            with torch.cuda.stream(stream) if stream and device.startswith('cuda') else nullcontext():\n                # Process sub-batch\n                sub_batch_paths = [os.path.join(tomo_dir, slice_file) for slice_file in sub_batch]\n                sub_batch_slice_nums = [int(slice_file.split('_')[1].split('.')[0]) for slice_file in sub_batch]\n                \n                # Run inference with profiling\n                with GPUProfiler(f\"Inference batch {i+1}/{len(sub_batches)}\"):\n                    # sub_results = model(sub_batch_paths, verbose=False)\n                    sub_results = make_predict(sub_batch_paths, model, device, SIZE)\n                    # print(sub_results)\n                \n                # Process each result in this sub-batch\n                # for result in sub_results:\n                    # print('nms_res1:', result)\n                for j,res in enumerate(sub_results):\n                    # print('nms_res2', res)\n                    # x1,y1,x2,y2, confidence, cls_ = res\n                    x1,y1,x2,y2, confidence = res\n                    if confidence >= CONFIDENCE_THRESHOLD:\n                        # Calculate center coordinates\n                        x_center = (x1 + x2) / 2\n                        y_center = (y1 + y2) / 2\n                                \n                        # Store detection with 3D coordinates\n                        all_detections.append({\n                                'z': round(sub_batch_slice_nums[j]),\n                                'y': round(y_center),\n                                'x': round(x_center),\n                                'confidence': float(confidence)\n                            })\n        \n        # Synchronize streams\n        if device.startswith('cuda'):\n            torch.cuda.synchronize()\n    \n    # Clean up thread if still running\n    if next_batch_thread is not None:\n        next_batch_thread.join()\n    \n    # 3D Non-Maximum Suppression to merge nearby detections across slices\n    final_detections = perform_3d_nms(all_detections, NMS_IOU_THRESHOLD)\n    \n    # Sort detections by confidence (highest first)\n    final_detections.sort(key=lambda x: x['confidence'], reverse=True)\n    \n    # If there are no detections, return NA values\n    if not final_detections:\n        return {\n            'tomo_id': tomo_id,\n            'Motor axis 0': -1,\n            'Motor axis 1': -1,\n            'Motor axis 2': -1\n        }\n    \n    # Take the detection with highest confidence\n    best_detection = final_detections[0]\n    \n    # Return result with integer coordinates\n    return {\n        'tomo_id': tomo_id,\n        'Motor axis 0': round(best_detection['z']),\n        'Motor axis 1': round(best_detection['y']),\n        'Motor axis 2': round(best_detection['x'])\n    }\n\ndef perform_3d_nms(detections, iou_threshold):\n    \"\"\"\n    Perform 3D Non-Maximum Suppression on detections to merge nearby motors\n    \"\"\"\n    if not detections:\n        return []\n    \n    # Sort by confidence (highest first)\n    detections = sorted(detections, key=lambda x: x['confidence'], reverse=True)\n    \n    # List to store final detections after NMS\n    final_detections = []\n    \n    # Define 3D distance function\n    def distance_3d(d1, d2):\n        return np.sqrt((d1['z'] - d2['z'])**2 + \n                       (d1['y'] - d2['y'])**2 + \n                       (d1['x'] - d2['x'])**2)\n    \n    # Maximum distance threshold (based on box size and slice gap)\n    box_size = 24  # Same as annotation box size\n    distance_threshold = box_size * iou_threshold\n    \n    # Process each detection\n    while detections:\n        # Take the detection with highest confidence\n        best_detection = detections.pop(0)\n        final_detections.append(best_detection)\n        \n        # Filter out detections that are too close to the best detection\n        detections = [d for d in detections if distance_3d(d, best_detection) > distance_threshold]\n    \n    return final_detections\n\ndef generate_submission():\n    \"\"\"\n    Main function to generate the submission file\n    \"\"\"\n    # Get list of test tomograms\n    test_tomos = sorted([d for d in os.listdir(test_dir) if os.path.isdir(os.path.join(test_dir, d))])\n    total_tomos = len(test_tomos)\n    \n    print(f\"Found {total_tomos} tomograms in test directory\")\n    \n    # Debug image loading for the first tomogram\n    # if test_tomos:\n    #     debug_image_loading(test_tomos[0])\n    \n    # Clear GPU cache before starting\n    if torch.cuda.is_available():\n        torch.cuda.empty_cache()\n    \n    # Initialize model once outside the processing loop\n    print(f\"Loading YOLO model from {model_path}\")\n    model = YOLOv10(model_path)\n    model.to(device)\n    \n    # Additional optimizations for inference\n    if device.startswith('cuda'):\n        # Fuse conv and bn layers for faster inference\n        model.fuse()\n        \n        # Enable model half precision (FP16) if on compatible GPU\n        if torch.cuda.get_device_capability(0)[0] >= 7:  # Volta or newer\n            model.model.half()\n            print(\"Using half precision (FP16) for inference\")\n    \n    # Process tomograms with parallelization\n    results = []\n    motors_found = 0\n    \n    # Using ThreadPoolExecutor with max_workers=1 since each worker uses the GPU already\n    # and we're parallelizing within each tomogram processing\n    with ThreadPoolExecutor(max_workers=4) as executor:\n        future_to_tomo = {}\n        \n        # Submit all tomograms for processing\n        for i, tomo_id in enumerate(test_tomos, 1):\n            future = executor.submit(process_tomogram, tomo_id, model, i, total_tomos)\n            future_to_tomo[future] = tomo_id\n        \n        # Process completed futures as they complete\n        for future in future_to_tomo:\n            tomo_id = future_to_tomo[future]\n            if torch.cuda.is_available():\n                    torch.cuda.empty_cache()\n                    \n            result = future.result()\n            results.append(result)\n                \n                # Update motors found count\n            has_motor = not pd.isna(result['Motor axis 0'])\n            if has_motor:\n                motors_found += 1\n                print(f\"Motor found in {tomo_id} at position: \"\n                      f\"z={result['Motor axis 0']}, y={result['Motor axis 1']}, x={result['Motor axis 2']}\")\n            else:\n                print(f\"No motor detected in {tomo_id}\")\n                    \n            print(f\"Current detection rate: {motors_found}/{len(results)} ({motors_found/len(results)*100:.1f}%)\")\n            \n    # Create submission dataframe\n    submission_df = pd.DataFrame(results)\n    \n    # Ensure proper column order\n    submission_df = submission_df[['tomo_id', 'Motor axis 0', 'Motor axis 1', 'Motor axis 2']]\n    \n    # Save the submission file\n    submission_df.to_csv(submission_path, index=False)\n\n    print(f\"\\nSubmission complete!\")\n    print(f\"Motors detected: {motors_found}/{total_tomos} ({motors_found/total_tomos*100:.1f}%)\")\n    print(f\"Submission saved to: {submission_path}\")\n    \n    # Display first few rows of submission\n    print(\"\\nSubmission preview:\")\n    print(submission_df.head())\n    \n    return submission_df","metadata":{"trusted":true,"_kg_hide-input":false,"execution":{"iopub.status.busy":"2025-06-04T13:15:26.874829Z","iopub.execute_input":"2025-06-04T13:15:26.875174Z","iopub.status.idle":"2025-06-04T13:15:41.233494Z","shell.execute_reply.started":"2025-06-04T13:15:26.875135Z","shell.execute_reply":"2025-06-04T13:15:41.232676Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Run the submission pipeline\nif __name__ == \"__main__\":\n    # Time entire process\n    start_time = time.time()\n\n    # Generate submission\n    submission = generate_submission()\n    \n    # Print total execution time\n    elapsed = time.time() - start_time\n    print(f\"\\nTotal execution time: {elapsed:.2f} seconds ({elapsed/60:.2f} minutes)\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-04T13:15:59.488871Z","iopub.execute_input":"2025-06-04T13:15:59.489416Z","iopub.status.idle":"2025-06-04T13:19:10.898168Z","shell.execute_reply.started":"2025-06-04T13:15:59.489378Z","shell.execute_reply":"2025-06-04T13:19:10.897272Z"}},"outputs":[],"execution_count":null}]}