{
"cells": [
{
"cell_type": "markdown",
"id": "4b5f4bab",
"metadata": {},
"source": [
"# Обзор медальных решений по датасету Brain Tumor Object Detection Datasets и реализация выбранной модели\n",
"\n",
"## Постановка задачи\n",
"Задача состоит в **детекции опухоли мозга на МРТ-снимках**: модель должна предсказать класс (`negative` / `positive`) и координаты ограничивающего прямоугольника.\n",
"\n",
"## Важное замечание по исходным материалам\n",
"В формулировке задания упоминаются **YOLOv7, YOLOv8, YOLOvX**. Однако среди фактически приложенных медальных ноутбуков представлены:\n",
"- решение на **YOLOv5s**;\n",
"- решение на **YOLOv8x**;\n",
"- решение, опубликованное как проект для **YOLOvX App**, но обучающее **YOLOv9s** через пакет `ultralytics`.\n",
"\n",
"Поэтому обзор ниже построен **по реально доступным медальным ноутбукам**.\n"
]
},
{
"cell_type": "markdown",
"id": "ed27c5c5",
"metadata": {},
"source": [
"## Краткий обзор рассмотренных решений\n",
"\n",
"### 1. Ноутбук `brain-tumor-yolo-od-train.ipynb`\n",
"- Архитектура: **YOLOv5s**\n",
"- Вход: `512×512`\n",
"- Параметры обучения: `batch=32`, `epochs=5`\n",
"- Набор данных: `axial_t1wce_2_class`\n",
"- Особенность: это очень короткий учебный прогон, автор сам отмечает, что эпох слишком мало для сильного качества.\n",
"\n",
"**Полученные метрики по ноутбуку**\n",
"- Precision = **0.0402**\n",
"- Recall = **0.0494**\n",
"- mAP@0.50 = **0.00782**\n",
"- mAP@0.50:0.95 = **0.00155**\n",
"\n",
"**Интерпретация** \n",
"Это качество крайне низкое. Главная причина — не архитектура сама по себе, а слишком короткое обучение (5 эпох) на небольшом датасете. Поэтому данный результат нельзя считать репрезентативной оценкой потенциала семейства YOLO.\n",
"\n",
"---\n",
"\n",
"### 2. Ноутбук `brain-tumor-detection-using-yolov8.ipynb`\n",
"- Архитектура: **YOLOv8x**\n",
"- Параметры обучения: `epochs=200`, предобученные веса `yolov8x.pt`\n",
"- Фреймворк: `ultralytics`\n",
"- Детекция проводится по двум классам: `negative`, `positive`\n",
"\n",
"**Полученные метрики по ноутбуку**\n",
"- Precision = **0.850**\n",
"- Recall = **0.896**\n",
"- mAP@0.50 = **0.921**\n",
"- mAP@0.50:0.95 = **0.710**\n",
"\n",
"По классам:\n",
"- `negative`: mAP@0.50 = **0.892**, mAP@0.50:0.95 = **0.702**\n",
"- `positive`: mAP@0.50 = **0.951**, mAP@0.50:0.95 = **0.718**\n",
"\n",
"**Интерпретация** \n",
"Это сильный результат. Модель уверенно локализует опухоль и хорошо различает два класса. Среди приложенных решений именно этот ноутбук показывает наилучшее качество.\n",
"\n",
"---\n",
"\n",
"### 3. Ноутбук `brain-tumor-detection-yolovx.ipynb`\n",
"- Архитектура: **YOLOv9s**\n",
"- Развёртывание: через **YOLOvX App**\n",
"- Параметры обучения: `epochs=100`, `imgsz=640`, `batch=4`\n",
"- Важно: **YOLOvX здесь — это платформа/приложение**, а не отдельная детекторная архитектура.\n",
"\n",
"**Полученные метрики по ноутбуку**\n",
"- Precision = **0.623**\n",
"- Recall = **0.839**\n",
"- mAP@0.50 = **0.784**\n",
"- mAP@0.50:0.95 = **0.595**\n",
"\n",
"По классам:\n",
"- `negative`: mAP@0.50 = **0.790**, mAP@0.50:0.95 = **0.601**\n",
"- `positive`: mAP@0.50 = **0.779**, mAP@0.50:0.95 = **0.588**\n",
"\n",
"**Интерпретация** \n",
"Результат хороший, но ниже, чем у YOLOv8x. При этом модель компактнее и легче для развёртывания.\n"
]
},
{
"cell_type": "markdown",
"id": "0f0fa921",
"metadata": {},
"source": [
"## Сравнение архитектур\n",
"\n",
"### YOLOv5\n",
"YOLOv5 использует классическую схему:\n",
"- **backbone** для извлечения признаков;\n",
"- **neck** (PAN/FPN) для объединения признаков разных масштабов;\n",
"- **head** для предсказания рамок, objectness и классов.\n",
"\n",
"Преимущества:\n",
"- высокая скорость;\n",
"- зрелая экосистема;\n",
"- простота обучения.\n",
"\n",
"Недостатки в данном эксперименте:\n",
"- результат в медальном ноутбуке слишком слабый из-за короткого обучения;\n",
"- на малом медицинском датасете без аккуратной настройки потенциал архитектуры не раскрыт.\n",
"\n",
"### YOLOv8\n",
"YOLOv8 — более современное развитие семейства YOLO:\n",
"- anchor-free head;\n",
"- улучшенная схема регрессии рамок;\n",
"- хорошая интеграция в `ultralytics`;\n",
"- устойчивое обучение и удобная валидация.\n",
"\n",
"Преимущества:\n",
"- лучшее качество среди рассмотренных решений;\n",
"- хорошее соотношение точности и удобства;\n",
"- стабильная работа на малых наборах данных при transfer learning.\n",
"\n",
"### YOLOv9s (решение через YOLOvX)\n",
"YOLOv9s — более новая архитектура, ориентированная на эффективность и качество, а YOLOvX в данном случае используется как инструмент публикации и инференса модели.\n",
"\n",
"Преимущества:\n",
"- компактность;\n",
"- хороший recall;\n",
"- удобное развёртывание.\n",
"\n",
"Недостатки:\n",
"- по данным конкретного ноутбука качество всё же уступает YOLOv8x.\n"
]
},
{
"cell_type": "markdown",
"id": "9661c3a7",
"metadata": {},
"source": [
"## Выводы по применимости в медицинской диагностике\n",
"\n",
"1. **Архитектуры семейства YOLO в целом применимы** для задачи детекции опухоли на МРТ, поскольку позволяют одновременно локализовать подозрительную область и выдать класс.\n",
"2. Для медицинской диагностики особенно важны:\n",
" - высокий **Recall**, чтобы минимизировать пропуск опухоли;\n",
" - достаточный **mAP@0.50:0.95**, чтобы локализация была не только грубой, но и устойчивой.\n",
"3. Наиболее убедительный результат среди приложенных решений показала **YOLOv8x**:\n",
" - высокий recall (**0.896**);\n",
" - лучший `mAP@0.50 = 0.921`;\n",
" - лучший `mAP@0.50:0.95 = 0.710`.\n",
"4. Однако даже хорошая модель детекции **не должна использоваться как автономный диагностический инструмент**. Для медицинской практики она может рассматриваться как **система поддержки принятия решений**, а не как замена врачу-рентгенологу.\n"
]
},
{
"cell_type": "markdown",
"id": "736d4def",
"metadata": {},
"source": [
"## Выбранная модель для собственной реализации\n",
"\n",
"Для реализации выбирается **YOLOv8x**, потому что:\n",
"- она показала **лучшие фактические метрики** среди медальных ноутбуков;\n",
"- архитектура уже продемонстрировала хорошую устойчивость именно на этом датасете;\n",
"- экосистема `ultralytics` позволяет быстро воспроизводить обучение, валидацию и инференс.\n",
"\n",
"Ниже приведён воспроизводимый код обучения и оценки модели. \n",
"Ноутбук **адаптирован под локальный запуск**: датасет можно положить в любую папку и указать путь вручную либо дать ноутбуку найти его автоматически.\n"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "5499ee90",
"metadata": {},
"outputs": [],
"source": [
"# Установка зависимостей\n",
"# При локальном запуске при необходимости раскомментируйте:\n",
"# !pip install -q ultralytics pyyaml pandas matplotlib\n",
"\n",
"import os\n",
"from pathlib import Path\n",
"import yaml\n",
"import pandas as pd\n",
"import matplotlib.pyplot as plt\n",
"\n",
"from ultralytics import YOLO\n"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "686dc728",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"DATA_ROOT = /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class\n"
]
}
],
"source": [
"\n",
"# Локально-совместимый поиск датасета\n",
"# 1) Сначала можно явно указать путь вручную.\n",
"# 2) Если путь не задан, ноутбук попробует найти датасет автоматически.\n",
"\n",
"MANUAL_DATA_ROOT = None\n",
"# Примеры:\n",
"# MANUAL_DATA_ROOT = Path('./brain-tumor-object-detection-datasets/axial_t1wce_2_class')\n",
"# MANUAL_DATA_ROOT = Path('./data/axial_t1wce_2_class')\n",
"\n",
"def is_valid_dataset_root(p: Path) -> bool:\n",
" return (\n",
" p.exists()\n",
" and (p / 'images' / 'train').exists()\n",
" and (p / 'labels' / 'train').exists()\n",
" and (\n",
" (p / 'images' / 'test').exists()\n",
" or (p / 'images' / 'val').exists()\n",
" )\n",
" )\n",
"\n",
"def find_dataset_root() -> Path:\n",
" candidates = []\n",
"\n",
" if MANUAL_DATA_ROOT is not None:\n",
" candidates.append(Path(MANUAL_DATA_ROOT))\n",
"\n",
" # Частые варианты локального расположения\n",
" search_roots = [\n",
" Path('.'),\n",
" Path('./data'),\n",
" Path('./datasets'),\n",
" Path('./dataset'),\n",
" Path('./input'),\n",
" Path('./brain-tumor-object-detection-datasets'),\n",
" ]\n",
"\n",
" for root in search_roots:\n",
" if not root.exists():\n",
" continue\n",
"\n",
" direct_candidates = [\n",
" root / 'axial_t1wce_2_class',\n",
" root / 'brain-tumor-object-detection-datasets' / 'axial_t1wce_2_class',\n",
" root / 'data' / 'axial_t1wce_2_class',\n",
" ]\n",
" candidates.extend(direct_candidates)\n",
"\n",
" # Рекурсивный поиск только при необходимости\n",
" for found in root.rglob('axial_t1wce_2_class'):\n",
" candidates.append(found)\n",
"\n",
" seen = set()\n",
" unique_candidates = []\n",
" for c in candidates:\n",
" rc = c.resolve()\n",
" if rc not in seen:\n",
" seen.add(rc)\n",
" unique_candidates.append(rc)\n",
"\n",
" for c in unique_candidates:\n",
" if is_valid_dataset_root(c):\n",
" return c\n",
"\n",
" raise FileNotFoundError(\n",
" 'Не удалось найти папку датасета. '\n",
" 'Укажите путь вручную в переменной MANUAL_DATA_ROOT.'\n",
" )\n",
"\n",
"DATA_ROOT = find_dataset_root()\n",
"print('DATA_ROOT =', DATA_ROOT)\n",
"\n",
"assert is_valid_dataset_root(DATA_ROOT), f'Не найден корректный DATA_ROOT: {DATA_ROOT}'\n"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "f3a95ba2",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"path: /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class\n",
"train: images/train\n",
"val: images/test\n",
"test: images/test\n",
"names:\n",
" 0: negative\n",
" 1: positive\n",
"\n"
]
}
],
"source": [
"\n",
"# Формируем data.yaml для YOLOv8\n",
"work_dir = Path('./brain_tumor_yolov8_work').resolve()\n",
"work_dir.mkdir(parents=True, exist_ok=True)\n",
"\n",
"data_yaml_path = work_dir / 'brain_tumor_data.yaml'\n",
"\n",
"val_split = 'images/val' if (DATA_ROOT / 'images' / 'val').exists() else 'images/test'\n",
"test_split = 'images/test' if (DATA_ROOT / 'images' / 'test').exists() else val_split\n",
"\n",
"data_dict = {\n",
" 'path': str(DATA_ROOT.resolve()),\n",
" 'train': 'images/train',\n",
" 'val': val_split,\n",
" 'test': test_split,\n",
" 'names': {\n",
" 0: 'negative',\n",
" 1: 'positive'\n",
" }\n",
"}\n",
"\n",
"with open(data_yaml_path, 'w', encoding='utf-8') as f:\n",
" yaml.safe_dump(data_dict, f, allow_unicode=True, sort_keys=False)\n",
"\n",
"print(data_yaml_path.read_text(encoding='utf-8'))\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "cb7e0091",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\u001b[KDownloading https://github.com/ultralytics/assets/releases/download/v8.4.0/yolov8x.pt to 'yolov8x.pt': 100% ━━━━━━━━━━━━ 130.5MB 4.3MB/s 30.2s 30.2s<0.0s54\n"
]
}
],
"source": [
"# Инициализация модели\n",
"# Для максимального соответствия лучшему медальному решению используется yolov8x.pt.\n",
"model = YOLO('yolov8x.pt')\n"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "c64b3abe",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Ultralytics 8.4.24 🚀 Python-3.12.12 torch-2.10.0+cu128 CUDA:0 (NVIDIA GeForce RTX 3080, 9873MiB)\n",
"\u001b[34m\u001b[1mengine/trainer: \u001b[0magnostic_nms=False, amp=True, angle=1.0, augment=False, auto_augment=randaugment, batch=16, bgr=0.0, box=7.5, cache=False, cfg=None, classes=None, close_mosaic=10, cls=0.5, compile=False, conf=None, copy_paste=0.0, copy_paste_mode=flip, cos_lr=False, cutmix=0.0, data=/home/konnilol/Downloads/brain_tumor_yolov8_work/brain_tumor_data.yaml, degrees=0.0, deterministic=True, device=None, dfl=1.5, dnn=False, dropout=0.0, dynamic=False, embed=None, end2end=None, epochs=200, erasing=0.4, exist_ok=True, fliplr=0.5, flipud=0.0, format=torchscript, fraction=1.0, freeze=None, half=False, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, imgsz=640, int8=False, iou=0.7, keras=False, kobj=1.0, line_width=None, lr0=0.01, lrf=0.01, mask_ratio=4, max_det=300, mixup=0.0, mode=train, model=yolov8x.pt, momentum=0.937, mosaic=1.0, multi_scale=0.0, name=yolov8x_brain_tumor, nbs=64, nms=False, opset=None, optimize=False, optimizer=auto, overlap_mask=True, patience=100, perspective=0.0, plots=True, pose=12.0, pretrained=True, profile=False, project=/home/konnilol/Downloads/brain_tumor_yolov8_work/runs, rect=False, resume=False, retina_masks=False, rle=1.0, save=True, save_conf=False, save_crop=False, save_dir=/home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor, save_frames=False, save_json=False, save_period=-1, save_txt=False, scale=0.5, seed=0, shear=0.0, show=False, show_boxes=True, show_conf=True, show_labels=True, simplify=True, single_cls=False, source=None, split=val, stream_buffer=False, task=detect, time=None, tracker=botsort.yaml, translate=0.1, val=True, verbose=True, vid_stride=1, visualize=False, warmup_bias_lr=0.1, warmup_epochs=3.0, warmup_momentum=0.8, weight_decay=0.0005, workers=8, workspace=None\n",
"Overriding model.yaml nc=80 with nc=2\n",
"\n",
" from n params module arguments \n",
" 0 -1 1 2320 ultralytics.nn.modules.conv.Conv [3, 80, 3, 2] \n",
" 1 -1 1 115520 ultralytics.nn.modules.conv.Conv [80, 160, 3, 2] \n",
" 2 -1 3 436800 ultralytics.nn.modules.block.C2f [160, 160, 3, True] \n",
" 3 -1 1 461440 ultralytics.nn.modules.conv.Conv [160, 320, 3, 2] \n",
" 4 -1 6 3281920 ultralytics.nn.modules.block.C2f [320, 320, 6, True] \n",
" 5 -1 1 1844480 ultralytics.nn.modules.conv.Conv [320, 640, 3, 2] \n",
" 6 -1 6 13117440 ultralytics.nn.modules.block.C2f [640, 640, 6, True] \n",
" 7 -1 1 3687680 ultralytics.nn.modules.conv.Conv [640, 640, 3, 2] \n",
" 8 -1 3 6969600 ultralytics.nn.modules.block.C2f [640, 640, 3, True] \n",
" 9 -1 1 1025920 ultralytics.nn.modules.block.SPPF [640, 640, 5] \n",
" 10 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n",
" 11 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1] \n",
" 12 -1 3 7379200 ultralytics.nn.modules.block.C2f [1280, 640, 3] \n",
" 13 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n",
" 14 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1] \n",
" 15 -1 3 1948800 ultralytics.nn.modules.block.C2f [960, 320, 3] \n",
" 16 -1 1 922240 ultralytics.nn.modules.conv.Conv [320, 320, 3, 2] \n",
" 17 [-1, 12] 1 0 ultralytics.nn.modules.conv.Concat [1] \n",
" 18 -1 3 7174400 ultralytics.nn.modules.block.C2f [960, 640, 3] \n",
" 19 -1 1 3687680 ultralytics.nn.modules.conv.Conv [640, 640, 3, 2] \n",
" 20 [-1, 9] 1 0 ultralytics.nn.modules.conv.Concat [1] \n",
" 21 -1 3 7379200 ultralytics.nn.modules.block.C2f [1280, 640, 3] \n",
" 22 [15, 18, 21] 1 8719894 ultralytics.nn.modules.head.Detect [2, 16, None, [320, 640, 640]]\n",
"Model summary: 210 layers, 68,154,534 parameters, 68,154,518 gradients, 258.1 GFLOPs\n",
"\n",
"Transferred 589/595 items from pretrained weights\n",
"Freezing layer 'model.22.dfl.conv.weight'\n",
"\u001b[34m\u001b[1mAMP: \u001b[0mrunning Automatic Mixed Precision (AMP) checks...\n",
"\u001b[KDownloading https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n.pt to 'yolo26n.pt': 100% ━━━━━━━━━━━━ 5.3MB 5.2MB/s 1.0s.0s<0.1ss.0s7s\n",
"\u001b[34m\u001b[1mAMP: \u001b[0mchecks passed ✅\n",
"\u001b[34m\u001b[1mtrain: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 385.0±165.9 MB/s, size: 6.9 KB)\n",
"\u001b[K\u001b[34m\u001b[1mtrain: \u001b[0mScanning /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/train... 296 images, 14 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 310/310 3.3Kit/s 0.1s\n",
"\u001b[34m\u001b[1mtrain: \u001b[0mNew cache created: /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/train.cache\n",
"\u001b[34m\u001b[1mval: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 396.6±297.4 MB/s, size: 11.2 KB)\n",
"\u001b[K\u001b[34m\u001b[1mval: \u001b[0mScanning /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/test... 75 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 75/75 2.4Kit/s 0.0s\n",
"\u001b[34m\u001b[1mval: \u001b[0mNew cache created: /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/test.cache\n",
"\u001b[34m\u001b[1moptimizer:\u001b[0m 'optimizer=auto' found, ignoring 'lr0=0.01' and 'momentum=0.937' and determining best 'optimizer', 'lr0' and 'momentum' automatically... \n",
"\u001b[34m\u001b[1moptimizer:\u001b[0m AdamW(lr=0.001667, momentum=0.9) with parameter groups 97 weight(decay=0.0), 104 weight(decay=0.0005), 103 bias(decay=0.0)\n",
"Plotting labels to /home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor/labels.jpg... \n",
"Image sizes 640 train, 640 val\n",
"Using 8 dataloader workers\n",
"Logging results to \u001b[1m/home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor\u001b[0m\n",
"Starting training for 200 epochs...\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"WARNING ⚠️ CUDA out of memory with batch=16. Reducing to batch=8 and retrying (1/3).\n",
"\u001b[34m\u001b[1mtrain: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 22.1±6.4 MB/s, size: 7.4 KB)\n",
"\u001b[K\u001b[34m\u001b[1mtrain: \u001b[0mScanning /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/train.cache... 296 images, 14 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 310/310 144.5Mit/s 0.0s\n",
"\u001b[34m\u001b[1mval: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 234.1±240.6 MB/s, size: 11.8 KB)\n",
"\u001b[K\u001b[34m\u001b[1mval: \u001b[0mScanning /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/test.cache... 75 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 75/75 6.6Mit/s 0.0s\n",
"\u001b[34m\u001b[1moptimizer:\u001b[0m 'optimizer=auto' found, ignoring 'lr0=0.01' and 'momentum=0.937' and determining best 'optimizer', 'lr0' and 'momentum' automatically... \n",
"\u001b[34m\u001b[1moptimizer:\u001b[0m AdamW(lr=0.001667, momentum=0.9) with parameter groups 97 weight(decay=0.0), 104 weight(decay=0.0005), 103 bias(decay=0.0)\n",
"\u001b[K: 0% ──────────── 0/20 1.8s\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 1/200 8.78G 1.678 11.22 1.615 10 640: 100% ━━━━━━━━━━━━ 39/39 3.3it/s 11.9s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.3it/s 0.9s0.7s\n",
" all 75 81 1 0.0123 0.506 0.456\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 2/200 7.06G 1.512 3.038 1.535 9 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 4.0it/s 0.7s0.6s\n",
" all 75 81 0 0 0 0\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 3/200 7.06G 1.62 3.426 1.705 11 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.00147 0.0988 0.00246 0.000601\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 4/200 6.43G 1.676 2.613 1.693 17 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.4it/s 0.9s0.7s\n",
" all 75 81 0.133 0.0617 0.0486 0.0206\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 5/200 6.69G 1.563 1.998 1.583 16 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.3s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.4it/s 0.9s0.7s\n",
" all 75 81 0 0 0 0\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 6/200 6.7G 1.434 1.935 1.528 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.615 0.543 0.64 0.38\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 7/200 7.05G 1.34 1.725 1.461 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.543 0.519 0.54 0.361\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 8/200 7.07G 1.333 1.642 1.454 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.756 0.728 0.781 0.517\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 9/200 7.08G 1.313 1.582 1.401 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.577 0.247 0.3 0.171\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 10/200 7.07G 1.405 1.635 1.471 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.4s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.6s\n",
" all 75 81 0 0 0 0\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 11/200 7.07G 1.285 1.505 1.41 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.621 0.63 0.638 0.417\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 12/200 6.4G 1.227 1.48 1.365 13 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.858 0.802 0.901 0.642\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 13/200 6.68G 1.219 1.441 1.356 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.677 0.765 0.712 0.481\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 14/200 6.68G 1.189 1.421 1.329 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.923 0.739 0.884 0.582\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 15/200 6.69G 1.172 1.338 1.295 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.468 0.185 0.212 0.0835\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 16/200 7.05G 1.162 1.371 1.306 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.758 0.657 0.716 0.472\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 17/200 7.09G 1.165 1.345 1.361 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.849 0.626 0.714 0.428\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 18/200 7.08G 1.137 1.287 1.248 15 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.4s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.817 0.815 0.867 0.556\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 19/200 7.05G 1.093 1.326 1.282 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.765 0.722 0.777 0.535\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 20/200 6.42G 1.147 1.299 1.314 10 640: 100% ━━━━━━━━━━━━ 39/39 3.9it/s 9.9s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.9it/s 0.8s0.6s\n",
" all 75 81 0.889 0.787 0.86 0.571\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 21/200 6.68G 1.067 1.239 1.224 16 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.865 0.794 0.873 0.588\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 22/200 6.68G 1.151 1.279 1.311 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.847 0.754 0.798 0.543\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 23/200 7.07G 1.133 1.267 1.308 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.771 0.802 0.847 0.577\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 24/200 7.07G 1.071 1.201 1.238 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.798 0.802 0.85 0.569\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 25/200 7.07G 1.16 1.252 1.318 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.805 0.814 0.854 0.573\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 26/200 7.08G 1.136 1.243 1.29 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.754 0.914 0.87 0.532\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 27/200 7.05G 1.034 1.205 1.23 18 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.806 0.852 0.876 0.582\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 28/200 6.42G 1.106 1.233 1.301 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.788 0.782 0.804 0.527\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 29/200 6.65G 1.095 1.199 1.267 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.756 0.679 0.698 0.452\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 30/200 6.68G 1.086 1.211 1.26 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.694 0.617 0.634 0.427\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 31/200 7.06G 1.072 1.154 1.268 12 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.3s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.78 0.704 0.677 0.453\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 32/200 7.07G 1.001 1.155 1.214 9 640: 100% ━━━━━━━━━━━━ 39/39 3.6it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.834 0.827 0.853 0.554\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 33/200 7.05G 1.054 1.173 1.249 13 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.754 0.802 0.854 0.571\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 34/200 7.08G 0.991 1.126 1.204 5 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.715 0.889 0.829 0.528\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 35/200 7.08G 1.029 1.111 1.221 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.872 0.852 0.893 0.573\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 36/200 6.43G 1.012 1.079 1.24 13 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.836 0.852 0.885 0.604\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 37/200 6.67G 1.047 1.133 1.267 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.757 0.807 0.835 0.536\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 38/200 6.66G 0.999 1.07 1.199 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.747 0.728 0.722 0.447\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 39/200 7.07G 0.969 1.06 1.198 6 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.783 0.84 0.855 0.6\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 40/200 7.09G 1.017 1.061 1.217 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.7s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.701 0.802 0.773 0.49\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 41/200 7.08G 1.033 1.058 1.248 14 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.685 0.832 0.81 0.526\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 42/200 7.08G 1.002 1.024 1.217 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.664 0.716 0.729 0.5\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 43/200 7.07G 0.9573 1.106 1.191 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.766 0.726 0.808 0.583\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 44/200 6.43G 0.9778 1.058 1.192 14 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.6s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.826 0.815 0.874 0.548\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 45/200 6.67G 0.9502 1.092 1.19 6 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.727 0.815 0.738 0.508\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 46/200 6.69G 0.942 1.01 1.168 7 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.3s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.804 0.741 0.787 0.54\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 47/200 7.06G 0.9341 0.9589 1.188 9 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.893 0.802 0.891 0.617\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 48/200 7.07G 0.9675 0.9732 1.183 10 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.846 0.748 0.805 0.589\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 49/200 7.07G 0.9671 1.004 1.216 13 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.3s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.575 0.741 0.595 0.357\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 50/200 7.08G 0.9848 0.9078 1.205 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.668 0.704 0.674 0.438\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 51/200 7.07G 0.9382 0.8774 1.191 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.724 0.753 0.796 0.528\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 52/200 6.39G 0.9628 0.9263 1.175 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.75 0.741 0.81 0.523\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 53/200 6.65G 0.8867 0.8917 1.131 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.754 0.79 0.757 0.538\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 54/200 6.69G 0.9455 0.8715 1.177 14 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.779 0.829 0.801 0.504\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 55/200 7.07G 0.908 0.9014 1.158 11 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.3s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.6s\n",
" all 75 81 0.899 0.659 0.824 0.547\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 56/200 7.04G 0.9133 0.8988 1.157 8 640: 100% ━━━━━━━━━━━━ 39/39 3.9it/s 10.1s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.6s\n",
" all 75 81 0.827 0.829 0.863 0.58\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 57/200 7.08G 0.9097 0.9007 1.164 8 640: 100% ━━━━━━━━━━━━ 39/39 3.9it/s 10.0s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.8it/s 0.8s0.7s\n",
" all 75 81 0.574 0.631 0.674 0.399\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 58/200 7.08G 0.9007 0.854 1.158 7 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.2s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.834 0.373 0.573 0.347\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 59/200 7.08G 0.9682 0.8424 1.16 14 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.312 0.506 0.404 0.253\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 60/200 6.4G 0.9092 0.9085 1.155 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.697 0.753 0.713 0.438\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 61/200 6.68G 0.8863 0.8378 1.143 17 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.652 0.617 0.658 0.418\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 62/200 6.69G 0.893 0.841 1.167 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.817 0.728 0.863 0.606\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 63/200 7.07G 0.8341 0.7945 1.132 6 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.709 0.79 0.765 0.499\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 64/200 7.08G 0.8813 0.8208 1.153 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.681 0.711 0.736 0.506\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 65/200 7.05G 0.8699 0.7933 1.147 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.775 0.741 0.836 0.619\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 66/200 7.05G 0.8687 0.7775 1.117 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.739 0.874 0.853 0.583\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 67/200 7.05G 0.886 0.7668 1.159 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.736 0.655 0.759 0.495\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 68/200 6.42G 0.8802 0.7499 1.14 11 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.4s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.767 0.733 0.79 0.503\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 69/200 6.69G 0.8831 0.7986 1.153 11 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.4s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.914 0.802 0.904 0.633\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 70/200 6.67G 0.9055 0.7863 1.171 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.681 0.827 0.735 0.482\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 71/200 7.04G 0.8564 0.7137 1.112 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.662 0.593 0.7 0.442\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 72/200 7.08G 0.8566 0.7295 1.111 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.71 0.716 0.758 0.532\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 73/200 7.08G 0.8369 0.7323 1.128 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.813 0.7 0.807 0.606\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 74/200 7.07G 0.8528 0.7787 1.123 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.695 0.845 0.84 0.602\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 75/200 7.04G 0.8248 0.7854 1.12 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.717 0.765 0.807 0.569\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 76/200 6.39G 0.8447 0.6818 1.11 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.4s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.923 0.744 0.892 0.64\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 77/200 6.67G 0.8227 0.6872 1.115 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.819 0.815 0.874 0.585\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 78/200 6.64G 0.7994 0.6726 1.08 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.794 0.827 0.876 0.602\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 79/200 6.69G 0.8255 0.6278 1.103 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.797 0.704 0.739 0.513\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 80/200 7.08G 0.7775 0.675 1.088 13 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.566 0.901 0.819 0.581\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 81/200 7.07G 0.8028 0.6361 1.092 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.785 0.741 0.818 0.577\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 82/200 7.05G 0.7967 0.6583 1.1 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.798 0.802 0.82 0.563\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 83/200 7.05G 0.7943 0.652 1.085 10 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.814 0.647 0.777 0.517\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 84/200 6.43G 0.7833 0.669 1.102 5 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.807 0.679 0.785 0.569\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 85/200 6.7G 0.7816 0.6456 1.07 15 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.832 0.691 0.847 0.571\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 86/200 7.06G 0.7798 0.6794 1.102 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.688 0.926 0.877 0.621\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 87/200 7.08G 0.7694 0.5896 1.075 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.83 0.724 0.831 0.589\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 88/200 7.07G 0.7838 0.6436 1.085 5 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.805 0.763 0.832 0.617\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 89/200 7.08G 0.7727 0.6808 1.075 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.9s0.7s\n",
" all 75 81 0.88 0.457 0.676 0.457\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 90/200 7.07G 0.7874 0.6558 1.082 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.694 0.716 0.771 0.516\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 91/200 7.04G 0.7509 0.5887 1.06 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.772 0.79 0.868 0.634\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 92/200 6.4G 0.7774 0.5958 1.061 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.755 0.759 0.868 0.626\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 93/200 6.64G 0.7673 0.6121 1.063 11 640: 100% ━━━━━━━━━━━━ 39/39 3.8it/s 10.4s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.875 0.753 0.892 0.632\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 94/200 6.65G 0.7822 0.6237 1.08 12 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.814 0.802 0.879 0.622\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 95/200 7.04G 0.7748 0.6013 1.076 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.778 0.679 0.793 0.546\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 96/200 7.06G 0.7494 0.5547 1.051 11 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.565 0.864 0.765 0.524\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 97/200 7.08G 0.755 0.5336 1.056 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.874 0.765 0.873 0.624\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 98/200 7.08G 0.7312 0.5701 1.063 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.68 0.683 0.79 0.516\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 99/200 7.07G 0.7248 0.541 1.052 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.8 0.74 0.833 0.581\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 100/200 6.43G 0.7193 0.5425 1.051 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.845 0.753 0.853 0.617\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 101/200 6.69G 0.7385 0.57 1.062 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.855 0.51 0.743 0.514\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 102/200 6.67G 0.7331 0.615 1.075 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.884 0.778 0.859 0.584\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 103/200 7.06G 0.7092 0.5532 1.039 7 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.5it/s 0.8s0.7s\n",
" all 75 81 0.836 0.728 0.83 0.569\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 104/200 7.09G 0.6855 0.5588 1.024 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.887 0.68 0.891 0.625\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 105/200 7.05G 0.7214 0.5928 1.042 5 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.863 0.776 0.834 0.593\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 106/200 7.05G 0.7214 0.56 1.05 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.867 0.727 0.84 0.584\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 107/200 7.08G 0.7241 0.5546 1.043 5 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.3s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.939 0.754 0.89 0.615\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 108/200 6.4G 0.6998 0.5197 1.048 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.802 0.648 0.813 0.546\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 109/200 6.64G 0.674 0.5007 1.037 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.7it/s 0.8s0.7s\n",
" all 75 81 0.817 0.716 0.823 0.573\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 110/200 6.7G 0.7052 0.5336 1.028 9 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.863 0.802 0.873 0.606\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 111/200 7.04G 0.6917 0.5072 1.045 8 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.839 0.753 0.878 0.604\n",
"\n",
" Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n",
"\u001b[K 112/200 7.08G 0.6863 0.503 1.007 15 640: 100% ━━━━━━━━━━━━ 39/39 3.7it/s 10.5s0.2s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.6it/s 0.8s0.7s\n",
" all 75 81 0.776 0.769 0.853 0.592\n",
"\u001b[34m\u001b[1mEarlyStopping: \u001b[0mTraining stopped early as no improvement observed in last 100 epochs. Best results observed at epoch 12, best model saved as best.pt.\n",
"To update EarlyStopping(patience=100) pass a new patience value, i.e. `patience=300` or use `patience=0` to disable EarlyStopping.\n",
"\n",
"112 epochs completed in 0.394 hours.\n",
"Optimizer stripped from /home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor/weights/last.pt, 136.7MB\n",
"Optimizer stripped from /home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor/weights/best.pt, 136.7MB\n",
"\n",
"Validating /home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor/weights/best.pt...\n",
"Ultralytics 8.4.24 🚀 Python-3.12.12 torch-2.10.0+cu128 CUDA:0 (NVIDIA GeForce RTX 3080, 9873MiB)\n",
"Model summary (fused): 113 layers, 68,125,494 parameters, 0 gradients, 257.4 GFLOPs\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 3.4it/s 0.9s0.7s\n",
" all 75 81 0.858 0.802 0.901 0.643\n",
" negative 75 81 0.858 0.802 0.901 0.643\n",
"Speed: 0.1ms preprocess, 9.6ms inference, 0.0ms loss, 0.9ms postprocess per image\n",
"Results saved to \u001b[1m/home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor\u001b[0m\n"
]
}
],
"source": [
"\n",
"# Обучение модели\n",
"# При необходимости можно уменьшить epochs для быстрого теста.\n",
"RUN_NAME = 'yolov8x_brain_tumor'\n",
"\n",
"train_results = model.train(\n",
" data=str(data_yaml_path),\n",
" epochs=200,\n",
" imgsz=640,\n",
" batch=16,\n",
" project=str(work_dir / 'runs'),\n",
" name=RUN_NAME,\n",
" pretrained=True,\n",
" plots=True,\n",
" exist_ok=True\n",
")\n"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "a739b6b0",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"best_weights = /home/konnilol/Downloads/brain_tumor_yolov8_work/runs/yolov8x_brain_tumor/weights/best.pt\n",
"Ultralytics 8.4.24 🚀 Python-3.12.12 torch-2.10.0+cu128 CUDA:0 (NVIDIA GeForce RTX 3080, 9873MiB)\n",
"Model summary (fused): 113 layers, 68,125,494 parameters, 0 gradients, 257.4 GFLOPs\n",
"\u001b[34m\u001b[1mval: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 345.1±167.2 MB/s, size: 8.3 KB)\n",
"\u001b[K\u001b[34m\u001b[1mval: \u001b[0mScanning /home/konnilol/Downloads/brain-tumor-dataset/axial_t1wce_2_class/labels/test.cache... 75 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 75/75 28.6Mit/s 0.0s\n",
"\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 5/5 3.0it/s 1.7s0.5s\n",
" all 75 81 0.843 0.79 0.881 0.642\n",
" negative 75 81 0.843 0.79 0.881 0.642\n",
"Speed: 1.7ms preprocess, 17.6ms inference, 0.0ms loss, 0.4ms postprocess per image\n",
"Results saved to \u001b[1m/home/konnilol/Downloads/runs/detect/val\u001b[0m\n"
]
}
],
"source": [
"\n",
"# Валидация лучшей модели\n",
"best_weights = work_dir / 'runs' / RUN_NAME / 'weights' / 'best.pt'\n",
"if not best_weights.exists():\n",
" candidates = list((work_dir / 'runs').rglob('best.pt'))\n",
" if len(candidates) == 0:\n",
" raise FileNotFoundError('Файл best.pt не найден после обучения.')\n",
" best_weights = max(candidates, key=lambda p: p.stat().st_mtime)\n",
"\n",
"print('best_weights =', best_weights)\n",
"\n",
"best_model = YOLO(str(best_weights))\n",
"metrics = best_model.val(data=str(data_yaml_path), split='val')\n"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "84b84f61",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.positron.dataexplorer+json": {
"comm_id": "57ff4b80-6ab4-43c0-b09a-7de610e1cd64",
"shape": {
"columns": 2,
"rows": 4
},
"source": "pandas",
"title": "pandas",
"version": 1
},
"text/html": [
"\n",
"\n",
"
\n",
" \n",
" \n",
" metric \n",
" value \n",
" \n",
" \n",
" \n",
" \n",
" 0 \n",
" precision \n",
" 0.842771 \n",
" \n",
" \n",
" 1 \n",
" recall \n",
" 0.790123 \n",
" \n",
" \n",
" 2 \n",
" mAP50 \n",
" 0.881480 \n",
" \n",
" \n",
" 3 \n",
" mAP50-95 \n",
" 0.642498 \n",
" \n",
" \n",
"
\n",
"
"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Визуализация ключевых метрик\n",
"plt.figure(figsize=(7, 4))\n",
"plt.bar(metrics_df['metric'], metrics_df['value'])\n",
"plt.ylim(0, 1)\n",
"plt.title('Качество модели YOLOv8x на Brain Tumor Object Detection')\n",
"plt.ylabel('Значение')\n",
"plt.show()\n"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "f349efca",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"0: 640x640 1 positive, 17.7ms\n",
"1: 640x640 (no detections), 17.7ms\n",
"2: 640x640 (no detections), 17.7ms\n",
"3: 640x640 (no detections), 17.7ms\n",
"Speed: 1.5ms preprocess, 17.7ms inference, 0.3ms postprocess per image at shape (1, 3, 640, 640)\n",
"Results saved to \u001b[1m/home/konnilol/Downloads/brain_tumor_yolov8_work/predictions/sample_preds\u001b[0m\n",
"Сохранено в: /home/konnilol/Downloads/brain_tumor_yolov8_work/predictions/sample_preds\n"
]
}
],
"source": [
"\n",
"# Примеры инференса\n",
"test_images_dir = DATA_ROOT / 'images' / ('test' if (DATA_ROOT / 'images' / 'test').exists() else 'val')\n",
"sample_images = []\n",
"for ext in ('*.jpg', '*.jpeg', '*.png', '*.bmp'):\n",
" sample_images.extend(sorted(test_images_dir.glob(ext)))\n",
"sample_images = sample_images[:4]\n",
"\n",
"assert len(sample_images) > 0, f'В папке {test_images_dir} не найдены изображения для инференса.'\n",
"\n",
"predictions = best_model.predict(\n",
" source=[str(p) for p in sample_images],\n",
" conf=0.25,\n",
" save=True,\n",
" project=str(work_dir / 'predictions'),\n",
" name='sample_preds',\n",
" exist_ok=True\n",
")\n",
"\n",
"print('Сохранено в:', work_dir / 'predictions' / 'sample_preds')\n"
]
},
{
"cell_type": "markdown",
"id": "a7048a1e",
"metadata": {},
"source": [
"В работе были рассмотрены три медальных решения по датасету детекции опухоли мозга на МРТ. В приложённых ноутбуках фактически представлены YOLOv5s, YOLOv8x и YOLOv9s через YOLOvX. Лучшее качество в обзорных решениях показала YOLOv8x: Precision = 0.850, Recall = 0.896, mAP@0.50 = 0.921, mAP@0.50:0.95 = 0.710. Решение через YOLOvX, использующее YOLOv9s, также дало хороший результат (0.623 / 0.839 / 0.784 / 0.595), но уступило YOLOv8x. Ноутбук на YOLOv5s показал очень слабое качество (mAP@0.50 = 0.00782), однако это связано прежде всего с тем, что обучение велось всего 5 эпох, поэтому такой результат нельзя считать репрезентативным для самой архитектуры.\n",
"\n",
"Для собственной реализации была выбрана YOLOv8x как наиболее сильная и устойчивая модель среди доступных решений. После обучения получены метрики: Precision = 0.843, Recall = 0.790, mAP@0.50 = 0.881, mAP@0.50:0.95 = 0.642. Это немного ниже, чем в медальном ноутбуке YOLOv8x, но результат остаётся высоким и подтверждает, что архитектура хорошо подходит для данной задачи.\n",
"\n",
"Вывод.\n",
"Среди рассмотренных моделей наиболее эффективной для данного датасета оказалась YOLOv8x. Она обеспечивает лучшее сочетание точности детекции и качества локализации. Архитектуры семейства YOLO в целом применимы для медицинской диагностики как средства предварительного обнаружения и локализации опухоли, однако использовать их следует только как вспомогательный инструмент поддержки врача, а не как автономную диагностическую систему."
]
}
],
"metadata": {
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 5
}