{"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":11846439,"sourceType":"datasetVersion","datasetId":7402389},{"sourceId":224916709,"sourceType":"kernelVersion"}],"dockerImageVersionId":30919,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# BYU Locating Flagellar Motors\n\nPrevious step is training.\n\nThis notebook execute analyze train results and submittion.\n\n## my previous Notebook\n\nhttps://www.kaggle.com/code/hiranorm/let-s-train-yolov8-and-compress-train-results\n\n## citation\n\nhttps://www.kaggle.com/code/andrewjdarley/submission-notebook","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19"}},{"cell_type":"code","source":"!tar xfvz /kaggle/input/ultralytics-for-offline-install/archive.tar.gz\n!pip install --no-index --find-links=./packages ultralytics\n!rm -rf ./packages","metadata":{"trusted":true,"_kg_hide-input":false,"execution":{"iopub.status.busy":"2025-05-13T02:01:38.000798Z","iopub.execute_input":"2025-05-13T02:01:38.001126Z","iopub.status.idle":"2025-05-13T02:02:39.530029Z","shell.execute_reply.started":"2025-05-13T02:01:38.001099Z","shell.execute_reply":"2025-05-13T02:02:39.529032Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport numpy as np\nimport pandas as pd\nfrom PIL import Image\nimport torch\nimport cv2\nfrom tqdm.notebook import tqdm\nfrom ultralytics import YOLO\nimport threading\nimport time\nfrom contextlib import nullcontext\nfrom concurrent.futures import ThreadPoolExecutor\n\n# Set random seed for reproducibility\nnp.random.seed(42)\ntorch.manual_seed(42)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-13T02:03:27.145602Z","iopub.execute_input":"2025-05-13T02:03:27.145906Z","iopub.status.idle":"2025-05-13T02:03:32.966096Z","shell.execute_reply.started":"2025-05-13T02:03:27.145880Z","shell.execute_reply":"2025-05-13T02:03:32.965367Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"##### chengable\ntrain_result_dir = \"/kaggle/input/byu-yolo-train-result\"\n#####\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\nmotor_detector_path = train_result_dir+\"/yolo_weights/motor_detector\"\nmodel_path = motor_detector_path+\"/weights/best.pt\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-13T02:21:40.032081Z","iopub.execute_input":"2025-05-13T02:21:40.032425Z","iopub.status.idle":"2025-05-13T02:21:40.037021Z","shell.execute_reply.started":"2025-05-13T02:21:40.032403Z","shell.execute_reply":"2025-05-13T02:21:40.036005Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import matplotlib.pyplot as plt\nimport matplotlib.image as mpimg\nimport os\nimport math\n\ndef display_images_in_directory(directory_path, cols=5):\n    \"\"\"\n    指定されたディレクトリ内のJPGおよびPNG画像をタイル状に表示する関数。\n\n    Args:\n        directory_path (str): 画像ファイルが含まれるディレクトリへのパス。\n        cols (int): 1行に表示する画像の数。デフォルトは5。\n    \"\"\"\n\n    # 対応する画像ファイルの拡張子リスト（大文字小文字を考慮）\n    valid_extensions = ['.jpg', '.jpeg', '.png', '.JPG', '.JPEG', '.PNG']\n\n    # ディレクトリ内のすべてのファイルを取得\n    try:\n        all_files = os.listdir(directory_path)\n    except FileNotFoundError:\n        print(f\"エラー: ディレクトリ '{directory_path}' が見つかりません。\")\n        return\n    except Exception as e:\n        print(f\"エラーが発生しました: {e}\")\n        return\n\n    # 画像ファイルのみをフィルタリング\n    image_files = [f for f in all_files if os.path.splitext(f)[1] in valid_extensions]\n    image_files.sort() # アルファベット順にソートして表示順を安定させる\n\n    if not image_files:\n        print(f\"'{directory_path}' ディレクトリに表示可能な画像ファイルが見つかりませんでした。\")\n        return\n\n    n_images = len(image_files)\n    rows = math.ceil(n_images / cols) # 必要な行数を計算\n\n    # Matplotlibのフィギュアとサブプロットを作成\n    # figsizeは画像の数に応じて調整すると良いでしょう\n    fig, axes = plt.subplots(rows, cols, figsize=(cols * 6, rows * 6))\n\n    # axesが2次元配列でない場合（例：画像が1枚のみの場合）に対応\n    if rows == 1 and cols == 1:\n        axes = [[axes]]\n    elif rows == 1:\n        axes = [axes]\n    elif cols == 1:\n        axes = [[ax] for ax in axes]\n\n\n    # 各サブプロットに画像を表示\n    for i, image_file in enumerate(image_files):\n        row_index = i // cols\n        col_index = i % cols\n        ax = axes[row_index][col_index]\n\n        image_path = os.path.join(directory_path, image_file)\n\n        try:\n            img = mpimg.imread(image_path)\n            ax.imshow(img)\n            ax.set_title(image_file, fontsize=8) # ファイル名を表示\n            ax.axis('off') # 軸を非表示\n        except Exception as e:\n            print(f\"'{image_file}' の読み込み中にエラーが発生しました: {e}\")\n            ax.set_title(f\"Error: {image_file}\", fontsize=8, color='red')\n            ax.axis('off')\n\n\n    # 使用しないサブプロットを非表示にする\n    for i in range(n_images, rows * cols):\n        row_index = i // cols\n        col_index = i % cols\n        axes[row_index][col_index].axis('off')\n\n    # レイアウトを調整\n    plt.tight_layout()\n\n    # 画像を表示\n    plt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-13T02:24:10.972838Z","iopub.execute_input":"2025-05-13T02:24:10.973349Z","iopub.status.idle":"2025-05-13T02:24:10.982716Z","shell.execute_reply.started":"2025-05-13T02:24:10.973308Z","shell.execute_reply":"2025-05-13T02:24:10.981818Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# 1行に表示する画像の数を指定\ncolumns_per_row = 2 # 例えば、1行に4枚表示する場合\n\n# 関数を実行して画像を表示\ndisplay_images_in_directory(motor_detector_path, cols=columns_per_row)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-13T02:24:11.344077Z","iopub.execute_input":"2025-05-13T02:24:11.344391Z","iopub.status.idle":"2025-05-13T02:24:23.991490Z","shell.execute_reply.started":"2025-05-13T02:24:11.344368Z","shell.execute_reply":"2025-05-13T02:24:23.990187Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\n# Detection parameters\nCONFIDENCE_THRESHOLD = 0.40  # Lower threshold to catch more potential motors\nMAX_DETECTIONS_PER_TOMO = 4  # Keep track of top N detections per tomogram\nNMS_IOU_THRESHOLD = 0.2  # Non-maximum suppression threshold for 3D clustering\nCONCENTRATION = 1 # ONLY PROCESS 1/20 slices for fast submission\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'\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 = 4  # Reduce batch size for CPU\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):\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(4, 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                \n                # Process each result in this sub-batch\n                for j, result in enumerate(sub_results):\n                    if len(result.boxes) > 0:\n                        boxes = result.boxes\n                        for box_idx, confidence in enumerate(boxes.conf):\n                            if confidence >= CONFIDENCE_THRESHOLD:\n                                # Get bounding box coordinates\n                                x1, y1, x2, y2 = boxes.xyxy[box_idx].cpu().numpy()\n                                \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 debug_image_loading(tomo_id):\n    \"\"\"\n    Debug function to check image loading\n    \"\"\"\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    if not slice_files:\n        print(f\"No image files found in {tomo_dir}\")\n        return\n        \n    print(f\"Found {len(slice_files)} image files in {tomo_dir}\")\n    sample_file = slice_files[len(slice_files)//2]  # Middle slice\n    img_path = os.path.join(tomo_dir, sample_file)\n    \n    # Try different loading methods\n    try:\n        # Method 1: PIL\n        img_pil = Image.open(img_path)\n        img_array_pil = np.array(img_pil)\n        print(f\"PIL Image shape: {img_array_pil.shape}, dtype: {img_array_pil.dtype}\")\n        \n        # Method 2: OpenCV\n        img_cv2 = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)\n        print(f\"OpenCV Image shape: {img_cv2.shape}, dtype: {img_cv2.dtype}\")\n        \n        # Method 3: Convert to RGB\n        img_rgb = cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB)\n        print(f\"OpenCV RGB Image shape: {img_rgb.shape}, dtype: {img_rgb.dtype}\")\n        \n        print(\"Image loading successful!\")\n    except Exception as e:\n        print(f\"Error loading image {img_path}: {e}\")\n        \n    # Also test with YOLO's built-in loader\n    try:\n        test_model = YOLO(model_path)\n        test_results = test_model([img_path], verbose=False)\n        print(\"YOLO model successfully processed the test image\")\n    except Exception as e:\n        print(f\"Error with YOLO processing: {e}\")\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 = YOLO(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=1) 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            try:\n                # Clear CUDA cache between tomograms\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            except Exception as e:\n                print(f\"Error processing {tomo_id}: {e}\")\n                # Create a default entry for failed tomograms\n                results.append({\n                    'tomo_id': tomo_id,\n                    'Motor axis 0': -1,\n                    'Motor axis 1': -1,\n                    'Motor axis 2': -1\n                })\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\n\n# 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)\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-05-13T02:24:57.737966Z","iopub.execute_input":"2025-05-13T02:24:57.738308Z","iopub.status.idle":"2025-05-13T02:25:58.334778Z","shell.execute_reply.started":"2025-05-13T02:24:57.738279Z","shell.execute_reply":"2025-05-13T02:25:58.334068Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}