{ "cells": [ { "cell_type": "markdown", "id": "66f5a637", "metadata": {}, "source": [ "\n", "# Реализация архитектуры YOLOv10 для детекции автомобилей\n", "\n", "**Статья (ResearchGate):** YOLOv10: Real-Time End-to-End Object Detection \n", "https://www.researchgate.net/publication/380821008_YOLOv10_Real-Time_End-to-End_Object_Detection\n", "\n", "**Выбранный датасет:** Cars (`sshikamaru/car-object-detection`)\n", "\n", "Ноутбук адаптирован **для локального запуска**. Нужны только папка датасета и доступ в интернет для установки пакетов и загрузки весов `yolov10n.pt`.\n" ] }, { "cell_type": "markdown", "id": "00d6d9c6", "metadata": {}, "source": [ "\n", "## 1. Краткая постановка задачи\n", "\n", "Нужно реализовать новую архитектуру детекции объектов из научной статьи и применить её к одному из практических датасетов. \n", "В этом ноутбуке используется **YOLOv10** и датасет **Cars** с одной категорией `car`.\n", "\n", "**Почему YOLOv10:**\n", "- это новая end-to-end архитектура детекции;\n", "- она убирает зависимость от `NMS` за счёт **consistent dual assignments**;\n", "- сочетает высокую точность и низкую задержку.\n" ] }, { "cell_type": "markdown", "id": "4e11fec6", "metadata": {}, "source": [ "\n", "## 2. Архитектура YOLOv10: детально, но компактно\n", "\n", "YOLOv10 сохраняет общую одностадийную схему `backbone -> neck -> detection head`, но меняет и обучение, и внутренние блоки модели.\n", "\n", "**Структура:**\n", "1. **Backbone** извлекает иерархию признаков из изображения.\n", "2. **Neck (PAN/FPN-подобное слияние)** объединяет признаки разных масштабов.\n", "3. **Dual head** во время обучения:\n", " - **one-to-many head** даёт плотный обучающий сигнал;\n", " - **one-to-one head** учится выдавать одно лучшее предсказание на объект.\n", "4. На инференсе используется **one-to-one branch**, поэтому модель работает **без NMS**.\n", "\n", "**Ключевые механизмы новизны:**\n", "- **Consistent dual assignments.** Оба режима назначения позитивных примеров согласованы одной метрикой сопоставления. Это позволяет обучать модель как обычный YOLO, но на инференсе перейти к end-to-end детекции без отдельного этапа подавления дубликатов.\n", "- **Lightweight classification head.** Классификационная ветвь облегчена глубинно-разделимыми свёртками, чтобы снизить FLOPs и задержку.\n", "- **Spatial-channel decoupled downsampling.** Уменьшение разрешения и изменение числа каналов разделены: сначала точечная свёртка по каналам, затем depthwise downsampling. Это уменьшает потери информации и вычислительные затраты.\n", "- **Rank-guided block design.** Глубина и тип блоков распределяются по стадиям неравномерно: более избыточные стадии упрощаются сильнее.\n", "- **Large-kernel conv + PSA (partial self-attention).** В глубоких стадиях модель расширяет receptive field и добавляет ограниченное self-attention без резкого роста стоимости.\n", "\n", "**Практический смысл:**\n", "YOLOv10 — это не просто «ещё один YOLO», а попытка одновременно улучшить **точность, задержку и end-to-end поведение**.\n" ] }, { "cell_type": "code", "execution_count": 7, "id": "f6c77b40", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[33mWARNING: Ignoring invalid distribution ~ulsectl (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution ~hirp (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution ~ulsectl (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution ~hirp (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution ~ulsectl (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution ~hirp (/home/konnilol/.pyenv/versions/3.12.12/lib/python3.12/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m" ] } ], "source": [ "\n", "# Если пакеты уже стоят, ячейку можно пропустить\n", "import sys\n", "!{sys.executable} -m pip -q install -U ultralytics pandas scikit-learn pyyaml matplotlib tqdm\n" ] }, { "cell_type": "code", "execution_count": 8, "id": "ab07ed58", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "ultralytics version: 8.4.24\n" ] } ], "source": [ "\n", "from pathlib import Path\n", "import os\n", "import random\n", "import shutil\n", "import time\n", "import yaml\n", "import math\n", "from collections import defaultdict\n", "\n", "import numpy as np\n", "import pandas as pd\n", "from sklearn.model_selection import train_test_split\n", "from tqdm.auto import tqdm\n", "import matplotlib.pyplot as plt\n", "from PIL import Image\n", "\n", "from ultralytics import YOLO\n", "import ultralytics\n", "\n", "SEED = 42\n", "random.seed(SEED)\n", "np.random.seed(SEED)\n", "\n", "print('ultralytics version:', ultralytics.__version__)\n" ] }, { "cell_type": "markdown", "id": "e324bdfd", "metadata": {}, "source": [ "\n", "## 3. Поиск датасета на локальной машине\n", "\n", "Ожидается исходная структура Kaggle-датасета с папкой `training_images` и CSV-файлом разметки `train_solution_bounding_boxes (1).csv`.\n", "\n", "Если автоопределение не сработает, задайте путь вручную в `MANUAL_DATA_ROOT`.\n" ] }, { "cell_type": "code", "execution_count": 9, "id": "67abf7a2", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "DATA_ROOT = /home/konnilol/Downloads/car-object-detection/data\n", "CSV_PATH = /home/konnilol/Downloads/car-object-detection/data/train_solution_bounding_boxes (1).csv\n", "TRAIN_IMAGES_DIR exists: True\n", "TEST_IMAGES_DIR exists: True\n" ] } ], "source": [ "\n", "MANUAL_DATA_ROOT = None # например: Path('./car-object-detection/data')\n", "\n", "candidate_roots = [\n", " Path('./car-object-detection/data'),\n", " Path('./car-object-detection'),\n", " Path('./data/car-object-detection'),\n", " Path('./data'),\n", " Path('.'),\n", "]\n", "\n", "def find_dataset_root(manual_root=None):\n", " if manual_root is not None:\n", " p = Path(manual_root)\n", " if (p / 'training_images').exists():\n", " return p\n", " if (p / 'data' / 'training_images').exists():\n", " return p / 'data'\n", " for root in candidate_roots:\n", " root = Path(root)\n", " if (root / 'training_images').exists():\n", " return root\n", " if (root / 'data' / 'training_images').exists():\n", " return root / 'data'\n", " for root, dirs, files in os.walk('.'):\n", " root = Path(root)\n", " if (root / 'training_images').exists():\n", " return root\n", " return None\n", "\n", "DATA_ROOT = find_dataset_root(MANUAL_DATA_ROOT)\n", "if DATA_ROOT is None:\n", " raise FileNotFoundError('Не найдена папка датасета. Укажите MANUAL_DATA_ROOT.')\n", "\n", "csv_candidates = list(DATA_ROOT.glob('train_solution_bounding_boxes*.csv'))\n", "if not csv_candidates:\n", " raise FileNotFoundError('Не найден CSV-файл с bbox-разметкой.')\n", "\n", "CSV_PATH = csv_candidates[0]\n", "TRAIN_IMAGES_DIR = DATA_ROOT / 'training_images'\n", "TEST_IMAGES_DIR = DATA_ROOT / 'testing_images'\n", "\n", "print('DATA_ROOT =', DATA_ROOT.resolve())\n", "print('CSV_PATH =', CSV_PATH.resolve())\n", "print('TRAIN_IMAGES_DIR exists:', TRAIN_IMAGES_DIR.exists())\n", "print('TEST_IMAGES_DIR exists:', TEST_IMAGES_DIR.exists())\n" ] }, { "cell_type": "markdown", "id": "ae5dd86a", "metadata": {}, "source": [ "\n", "## 4. Подготовка данных в формате YOLO\n" ] }, { "cell_type": "code", "execution_count": 10, "id": "d6258a24", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ " image xmin ymin xmax ymax\n", "0 vid_4_1000.jpg 281.259045 187.035071 327.727931 223.225547\n", "1 vid_4_10000.jpg 15.163531 187.035071 120.329957 236.430180\n", "2 vid_4_10040.jpg 239.192475 176.764801 361.968162 236.430180\n", "3 vid_4_10020.jpg 496.483358 172.363256 630.020260 231.539575\n", "4 vid_4_10060.jpg 16.630970 186.546010 132.558611 238.386422\n", "(559, 5)\n", "['image', 'xmin', 'ymin', 'xmax', 'ymax']\n" ] } ], "source": [ "\n", "df = pd.read_csv(CSV_PATH)\n", "print(df.head())\n", "print(df.shape)\n", "print(df.columns.tolist())\n" ] }, { "cell_type": "code", "execution_count": 11, "id": "5551305b", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Число изображений с bbox: 355\n", "Число bbox: 559\n" ] } ], "source": [ "\n", "# Нормализуем имена колонок\n", "rename_map = {}\n", "for c in df.columns:\n", " low = c.lower().strip()\n", " if low == 'image': rename_map[c] = 'image'\n", " elif low == 'xmin': rename_map[c] = 'xmin'\n", " elif low == 'ymin': rename_map[c] = 'ymin'\n", " elif low == 'xmax': rename_map[c] = 'xmax'\n", " elif low == 'ymax': rename_map[c] = 'ymax'\n", "\n", "df = df.rename(columns=rename_map)\n", "required = {'image', 'xmin', 'ymin', 'xmax', 'ymax'}\n", "missing = required - set(df.columns)\n", "if missing:\n", " raise ValueError(f'В CSV не хватает колонок: {missing}')\n", "\n", "# Оставляем только валидные bbox\n", "df = df.dropna(subset=['image', 'xmin', 'ymin', 'xmax', 'ymax']).copy()\n", "df['xmin'] = df['xmin'].astype(float)\n", "df['ymin'] = df['ymin'].astype(float)\n", "df['xmax'] = df['xmax'].astype(float)\n", "df['ymax'] = df['ymax'].astype(float)\n", "df = df[(df['xmax'] > df['xmin']) & (df['ymax'] > df['ymin'])].copy()\n", "\n", "image_to_boxes = defaultdict(list)\n", "for row in df.itertuples(index=False):\n", " image_to_boxes[row.image].append((row.xmin, row.ymin, row.xmax, row.ymax))\n", "\n", "all_images = sorted([p.name for p in TRAIN_IMAGES_DIR.glob('*') if p.suffix.lower() in {'.jpg', '.jpeg', '.png'}])\n", "all_images = [img for img in all_images if img in image_to_boxes]\n", "\n", "print('Число изображений с bbox:', len(all_images))\n", "print('Число bbox:', sum(len(v) for v in image_to_boxes.values()))\n" ] }, { "cell_type": "code", "execution_count": 12, "id": "bc59de1c", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "train: 284 val: 71\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "Подготовка train: 100%|██████████| 284/284 [00:00<00:00, 3154.36it/s]\n", "Подготовка val: 100%|██████████| 71/71 [00:00<00:00, 3173.78it/s]" ] }, { "name": "stdout", "output_type": "stream", "text": [ "path: /home/konnilol/Downloads/work_yolov10_cars/converted\n", "train: images/train\n", "val: images/val\n", "names:\n", " 0: car\n", "\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "\n" ] } ], "source": [ "\n", "WORKDIR = Path('./work_yolov10_cars').resolve()\n", "CONVERTED = WORKDIR / 'converted'\n", "for split in ['train', 'val']:\n", " (CONVERTED / 'images' / split).mkdir(parents=True, exist_ok=True)\n", " (CONVERTED / 'labels' / split).mkdir(parents=True, exist_ok=True)\n", "\n", "train_imgs, val_imgs = train_test_split(all_images, test_size=0.2, random_state=SEED, shuffle=True)\n", "print('train:', len(train_imgs), 'val:', len(val_imgs))\n", "\n", "def yolo_line(box, w, h, cls_id=0):\n", " xmin, ymin, xmax, ymax = box\n", " xc = ((xmin + xmax) / 2.0) / w\n", " yc = ((ymin + ymax) / 2.0) / h\n", " bw = (xmax - xmin) / w\n", " bh = (ymax - ymin) / h\n", " return f\"{cls_id} {xc:.6f} {yc:.6f} {bw:.6f} {bh:.6f}\"\n", "\n", "for split_name, split_imgs in [('train', train_imgs), ('val', val_imgs)]:\n", " for img_name in tqdm(split_imgs, desc=f'Подготовка {split_name}'):\n", " src = TRAIN_IMAGES_DIR / img_name\n", " dst = CONVERTED / 'images' / split_name / img_name\n", " shutil.copy2(src, dst)\n", "\n", " with Image.open(src) as im:\n", " w, h = im.size\n", " label_path = CONVERTED / 'labels' / split_name / f'{Path(img_name).stem}.txt'\n", " lines = [yolo_line(box, w, h, cls_id=0) for box in image_to_boxes[img_name]]\n", " label_path.write_text(''.join(lines), encoding='utf-8')\n", "\n", "DATA_YAML = WORKDIR / 'cars_yolov10.yaml'\n", "data_cfg = {\n", " 'path': str(CONVERTED),\n", " 'train': 'images/train',\n", " 'val': 'images/val',\n", " 'names': {0: 'car'}\n", "}\n", "DATA_YAML.write_text(yaml.safe_dump(data_cfg, sort_keys=False, allow_unicode=True), encoding='utf-8')\n", "print(DATA_YAML.read_text(encoding='utf-8'))\n" ] }, { "cell_type": "markdown", "id": "ee065dd4", "metadata": {}, "source": [ "\n", "## 5. Обучение YOLOv10\n", "\n", "Для локальной машины выбран **YOLOv10n** — самый компактный вариант семейства. \n", "Это позволяет показать архитектуру статьи и при этом сохранить разумное время обучения.\n" ] }, { "cell_type": "code", "execution_count": 13, "id": "e8cb02d5", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[KDownloading https://github.com/ultralytics/assets/releases/download/v8.4.0/yolov10n.pt to 'yolov10n.pt': 100% ━━━━━━━━━━━━ 5.6MB 3.6MB/s 1.5s.5s<0.0s1.4s8s\n", "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/work_yolov10_cars/cars_yolov10.yaml, degrees=0.0, deterministic=True, device=None, dfl=1.5, dnn=False, dropout=0.0, dynamic=False, embed=None, end2end=None, epochs=40, 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=yolov10n.pt, momentum=0.937, mosaic=1.0, multi_scale=0.0, name=yolov10n_cars, 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/work_yolov10_cars/runs, rect=False, resume=False, retina_masks=False, rle=1.0, save=True, save_conf=False, save_crop=False, save_dir=/home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars, save_frames=False, save_json=False, save_period=-1, save_txt=False, scale=0.2, seed=42, 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=0.05, 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=1\n", "\n", " from n params module arguments \n", " 0 -1 1 464 ultralytics.nn.modules.conv.Conv [3, 16, 3, 2] \n", " 1 -1 1 4672 ultralytics.nn.modules.conv.Conv [16, 32, 3, 2] \n", " 2 -1 1 7360 ultralytics.nn.modules.block.C2f [32, 32, 1, True] \n", " 3 -1 1 18560 ultralytics.nn.modules.conv.Conv [32, 64, 3, 2] \n", " 4 -1 2 49664 ultralytics.nn.modules.block.C2f [64, 64, 2, True] \n", " 5 -1 1 9856 ultralytics.nn.modules.block.SCDown [64, 128, 3, 2] \n", " 6 -1 2 197632 ultralytics.nn.modules.block.C2f [128, 128, 2, True] \n", " 7 -1 1 36096 ultralytics.nn.modules.block.SCDown [128, 256, 3, 2] \n", " 8 -1 1 460288 ultralytics.nn.modules.block.C2f [256, 256, 1, True] \n", " 9 -1 1 164608 ultralytics.nn.modules.block.SPPF [256, 256, 5] \n", " 10 -1 1 249728 ultralytics.nn.modules.block.PSA [256, 256] \n", " 11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", " 12 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1] \n", " 13 -1 1 148224 ultralytics.nn.modules.block.C2f [384, 128, 1] \n", " 14 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", " 15 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1] \n", " 16 -1 1 37248 ultralytics.nn.modules.block.C2f [192, 64, 1] \n", " 17 -1 1 36992 ultralytics.nn.modules.conv.Conv [64, 64, 3, 2] \n", " 18 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1] \n", " 19 -1 1 123648 ultralytics.nn.modules.block.C2f [192, 128, 1] \n", " 20 -1 1 18048 ultralytics.nn.modules.block.SCDown [128, 128, 3, 2] \n", " 21 [-1, 10] 1 0 ultralytics.nn.modules.conv.Concat [1] \n", " 22 -1 1 282624 ultralytics.nn.modules.block.C2fCIB [384, 256, 1, True, True] \n", " 23 [16, 19, 22] 1 861718 ultralytics.nn.modules.head.v10Detect [1, [64, 128, 256]] \n", "YOLOv10n summary: 224 layers, 2,707,430 parameters, 2,707,414 gradients, 8.4 GFLOPs\n", "\n", "Transferred 493/595 items from pretrained weights\n", "Freezing layer 'model.23.dfl.conv.weight'\n", "\u001b[34m\u001b[1mAMP: \u001b[0mrunning Automatic Mixed Precision (AMP) checks...\n", "\u001b[34m\u001b[1mAMP: \u001b[0mchecks passed ✅\n", "\u001b[34m\u001b[1mtrain: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 2544.0±1441.6 MB/s, size: 99.8 KB)\n", "\u001b[K\u001b[34m\u001b[1mtrain: \u001b[0mScanning /home/konnilol/Downloads/work_yolov10_cars/converted/labels/train... 284 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 284/284 3.1Kit/s 0.1s\n", "\u001b[34m\u001b[1mtrain: \u001b[0mNew cache created: /home/konnilol/Downloads/work_yolov10_cars/converted/labels/train.cache\n", "WARNING ⚠️ Box and segment counts should be equal, but got len(segments) = 93, len(boxes) = 284. To resolve this only boxes will be used and all segments will be removed. To avoid this please supply either a detect or segment dataset, not a detect-segment mixed dataset.\n", "\u001b[34m\u001b[1mval: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 1967.5±1363.5 MB/s, size: 100.8 KB)\n", "\u001b[K\u001b[34m\u001b[1mval: \u001b[0mScanning /home/konnilol/Downloads/work_yolov10_cars/converted/labels/val... 71 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 71/71 2.2Kit/s 0.0s\n", "\u001b[34m\u001b[1mval: \u001b[0mNew cache created: /home/konnilol/Downloads/work_yolov10_cars/converted/labels/val.cache\n", "WARNING ⚠️ Box and segment counts should be equal, but got len(segments) = 26, len(boxes) = 71. To resolve this only boxes will be used and all segments will be removed. To avoid this please supply either a detect or segment dataset, not a detect-segment mixed dataset.\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=0.002, momentum=0.9) with parameter groups 95 weight(decay=0.0), 108 weight(decay=0.0005), 107 bias(decay=0.0)\n", "Plotting labels to /home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars/labels.jpg... \n", "Image sizes 640 train, 640 val\n", "Using 8 dataloader workers\n", "Logging results to \u001b[1m/home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars\u001b[0m\n", "Starting training for 40 epochs...\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 1/40 2.69G 1.683 7.06 1.481 21 640: 100% ━━━━━━━━━━━━ 18/18 1.5it/s 12.3s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 1.2s/it 3.5s11.7s\n", " all 71 71 0.00324 0.972 0.0605 0.0209\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 2/40 2.7G 1.178 6.123 1.192 23 640: 100% ━━━━━━━━━━━━ 18/18 10.7it/s 1.7s0.2s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 18.8it/s 0.2s.2s\n", " all 71 71 0.00296 0.887 0.018 0.00601\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 3/40 2.7G 1.212 5.516 1.207 21 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.2s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 19.0it/s 0.2s.2s\n", " all 71 71 0.00319 0.958 0.0246 0.00701\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 4/40 2.7G 1.307 5.02 1.217 18 640: 100% ━━━━━━━━━━━━ 18/18 9.6it/s 1.9s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.2it/s 0.2s.2s\n", " all 71 71 0.117 0.127 0.0468 0.0208\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 5/40 2.7G 1.361 4.394 1.248 17 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.3it/s 0.2s.2s\n", " all 71 71 0.152 0.169 0.108 0.0357\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 6/40 2.7G 1.379 3.794 1.264 24 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.1it/s 0.2s.2s\n", " all 71 71 0.219 0.324 0.183 0.0938\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 7/40 2.7G 1.398 3.413 1.262 23 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.3it/s 0.2s.3s\n", " all 71 71 0.392 0.381 0.356 0.194\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 8/40 2.7G 1.414 3.18 1.248 18 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.6it/s 0.2s.2s\n", " all 71 71 0.3 0.296 0.223 0.114\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 9/40 2.7G 1.358 2.84 1.207 26 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.7it/s 0.2s.2s\n", " all 71 71 0.481 0.479 0.341 0.202\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 10/40 2.7G 1.307 2.631 1.229 28 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.9it/s 0.2s.2s\n", " all 71 71 0.224 0.282 0.219 0.14\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 11/40 2.7G 1.255 2.341 1.169 27 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.7it/s 0.2s.2s\n", " all 71 71 0.375 0.648 0.413 0.241\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 12/40 2.7G 1.241 2.263 1.177 20 640: 100% ━━━━━━━━━━━━ 18/18 10.5it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.4it/s 0.2s.2s\n", " all 71 71 0.412 0.423 0.351 0.223\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 13/40 2.7G 1.28 2.189 1.201 23 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.2it/s 0.2s.2s\n", " all 71 71 0.56 0.676 0.574 0.374\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 14/40 2.7G 1.284 2.012 1.2 21 640: 100% ━━━━━━━━━━━━ 18/18 10.5it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.0it/s 0.2s.2s\n", " all 71 71 0.625 0.469 0.533 0.338\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 15/40 2.7G 1.242 1.782 1.161 20 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.7it/s 0.2s.2s\n", " all 71 71 0.594 0.563 0.539 0.337\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 16/40 2.7G 1.217 1.633 1.166 27 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.6it/s 0.2s.2s\n", " all 71 71 0.557 0.662 0.585 0.363\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 17/40 2.7G 1.237 1.603 1.162 25 640: 100% ━━━━━━━━━━━━ 18/18 10.7it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.0it/s 0.2s.2s\n", " all 71 71 0.514 0.718 0.637 0.423\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 18/40 2.7G 1.227 1.588 1.156 30 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.6it/s 0.2s.2s\n", " all 71 71 0.59 0.606 0.591 0.371\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 19/40 2.7G 1.169 1.53 1.142 23 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.8it/s 0.2s.2s\n", " all 71 71 0.549 0.704 0.623 0.395\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 20/40 2.7G 1.181 1.472 1.128 21 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.3it/s 0.2s.2s\n", " all 71 71 0.708 0.662 0.699 0.449\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 21/40 2.7G 1.122 1.366 1.1 26 640: 100% ━━━━━━━━━━━━ 18/18 10.5it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.9it/s 0.2s.2s\n", " all 71 71 0.665 0.672 0.66 0.416\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 22/40 2.7G 1.151 1.339 1.11 28 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.3it/s 0.2s.2s\n", " all 71 71 0.48 0.592 0.494 0.338\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 23/40 2.7G 1.076 1.212 1.087 17 640: 100% ━━━━━━━━━━━━ 18/18 10.7it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.3it/s 0.2s.2s\n", " all 71 71 0.545 0.704 0.606 0.418\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 24/40 2.7G 1.126 1.182 1.124 20 640: 100% ━━━━━━━━━━━━ 18/18 10.8it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.6it/s 0.2s.2s\n", " all 71 71 0.655 0.722 0.645 0.425\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 25/40 2.7G 1.116 1.161 1.091 24 640: 100% ━━━━━━━━━━━━ 18/18 10.2it/s 1.8s.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 15.9it/s 0.2s.3s\n", " all 71 71 0.706 0.775 0.729 0.473\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 26/40 2.7G 1.12 1.133 1.108 23 640: 100% ━━━━━━━━━━━━ 18/18 10.3it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.1it/s 0.2s.2s\n", " all 71 71 0.484 0.648 0.561 0.38\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 27/40 2.7G 1.072 1.111 1.081 32 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.3it/s 0.2s.2s\n", " all 71 71 0.546 0.797 0.579 0.39\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 28/40 2.7G 1.061 1.095 1.082 15 640: 100% ━━━━━━━━━━━━ 18/18 10.5it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.0it/s 0.2s.2s\n", " all 71 71 0.683 0.704 0.699 0.48\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 29/40 2.7G 1.05 1.075 1.053 23 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.0it/s 0.2s.2s\n", " all 71 71 0.712 0.73 0.724 0.505\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 30/40 2.7G 1.026 1.021 1.048 21 640: 100% ━━━━━━━━━━━━ 18/18 10.5it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.4it/s 0.2s.2s\n", " all 71 71 0.69 0.721 0.699 0.481\n", "Closing dataloader mosaic\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 31/40 2.7G 0.9717 1.005 1.061 12 640: 100% ━━━━━━━━━━━━ 18/18 7.4it/s 2.4s<0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.4it/s 0.2s.2s\n", " all 71 71 0.513 0.652 0.516 0.358\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 32/40 2.7G 0.9624 0.9446 1.039 12 640: 100% ━━━━━━━━━━━━ 18/18 10.7it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.8it/s 0.2s.2s\n", " all 71 71 0.658 0.789 0.77 0.531\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 33/40 2.7G 0.9605 0.8613 1.086 12 640: 100% ━━━━━━━━━━━━ 18/18 10.9it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 18.0it/s 0.2s.2s\n", " all 71 71 0.726 0.775 0.786 0.538\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 34/40 2.7G 0.9543 0.816 1.041 12 640: 100% ━━━━━━━━━━━━ 18/18 10.8it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.2it/s 0.2s.2s\n", " all 71 71 0.657 0.803 0.799 0.557\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 35/40 2.7G 0.9323 0.8171 1.02 12 640: 100% ━━━━━━━━━━━━ 18/18 10.7it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.7it/s 0.2s.2s\n", " all 71 71 0.8 0.718 0.792 0.557\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 36/40 2.7G 0.902 0.7194 1.028 12 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.0it/s 0.2s.2s\n", " all 71 71 0.71 0.759 0.728 0.492\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 37/40 2.7G 0.9003 0.7076 1.027 12 640: 100% ━━━━━━━━━━━━ 18/18 10.8it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.3it/s 0.2s.2s\n", " all 71 71 0.723 0.773 0.747 0.509\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 38/40 2.7G 0.8693 0.6802 1.002 12 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 17.7it/s 0.2s.2s\n", " all 71 71 0.804 0.761 0.795 0.551\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 39/40 2.7G 0.8885 0.7155 1.039 11 640: 100% ━━━━━━━━━━━━ 18/18 10.6it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.2it/s 0.2s.2s\n", " all 71 71 0.802 0.761 0.811 0.564\n", "\n", " Epoch GPU_mem box_loss cls_loss dfl_loss Instances Size\n", "\u001b[K 40/40 2.7G 0.8847 0.6909 1.004 12 640: 100% ━━━━━━━━━━━━ 18/18 10.4it/s 1.7s0.1s\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 16.6it/s 0.2s.2s\n", " all 71 71 0.859 0.718 0.822 0.572\n", "\n", "40 epochs completed in 0.028 hours.\n", "Optimizer stripped from /home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars/weights/last.pt, 5.7MB\n", "Optimizer stripped from /home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars/weights/best.pt, 5.7MB\n", "\n", "Validating /home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars/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", "YOLOv10n summary (fused): 102 layers, 2,265,363 parameters, 0 gradients, 6.5 GFLOPs\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 3/3 10.4it/s 0.3s.4s\n", " all 71 71 0.859 0.718 0.822 0.572\n", "Speed: 0.1ms preprocess, 1.4ms inference, 0.0ms loss, 0.2ms postprocess per image\n", "Results saved to \u001b[1m/home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars\u001b[0m\n", "Папка эксперимента: /home/konnilol/Downloads/work_yolov10_cars/runs/yolov10n_cars\n" ] } ], "source": [ "\n", "MODEL_NAME = 'yolov10n.pt'\n", "IMG_SIZE = 640\n", "EPOCHS = 40\n", "BATCH = 16\n", "PROJECT_DIR = str(WORKDIR / 'runs')\n", "RUN_NAME = 'yolov10n_cars'\n", "\n", "model = YOLO(MODEL_NAME)\n", "train_results = model.train(\n", " data=str(DATA_YAML),\n", " epochs=EPOCHS,\n", " imgsz=IMG_SIZE,\n", " batch=BATCH,\n", " project=PROJECT_DIR,\n", " name=RUN_NAME,\n", " exist_ok=True,\n", " pretrained=True,\n", " seed=SEED,\n", " degrees=0.0,\n", " translate=0.05,\n", " scale=0.20,\n", " fliplr=0.5,\n", " mosaic=1.0,\n", " mixup=0.0,\n", " cache=False,\n", " verbose=True,\n", ")\n", "\n", "print('Папка эксперимента:', train_results.save_dir)\n" ] }, { "cell_type": "markdown", "id": "1d1e88d7", "metadata": {}, "source": [ "\n", "## 6. Валидация и расчёт метрик\n" ] }, { "cell_type": "code", "execution_count": 14, "id": "3aa7e14a", "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", "YOLOv10n summary (fused): 102 layers, 2,265,363 parameters, 0 gradients, 6.5 GFLOPs\n", "\u001b[34m\u001b[1mval: \u001b[0mFast image access ✅ (ping: 0.0±0.0 ms, read: 5182.2±873.8 MB/s, size: 114.5 KB)\n", "\u001b[K\u001b[34m\u001b[1mval: \u001b[0mScanning /home/konnilol/Downloads/work_yolov10_cars/converted/labels/val.cache... 71 images, 0 backgrounds, 0 corrupt: 100% ━━━━━━━━━━━━ 71/71 42.5Mit/s 0.0s\n", "WARNING ⚠️ Box and segment counts should be equal, but got len(segments) = 26, len(boxes) = 71. To resolve this only boxes will be used and all segments will be removed. To avoid this please supply either a detect or segment dataset, not a detect-segment mixed dataset.\n", "\u001b[K Class Images Instances Box(P R mAP50 mAP50-95): 100% ━━━━━━━━━━━━ 5/5 19.3it/s 0.3s.6s\n", " all 71 71 0.858 0.718 0.822 0.571\n", "Speed: 0.1ms preprocess, 2.5ms inference, 0.0ms loss, 0.1ms postprocess per image\n" ] }, { "data": { "application/vnd.positron.dataexplorer+json": { "comm_id": "e0cba7cf-bc02-4ddc-8430-2117dd2d179c", "shape": { "columns": 2, "rows": 5 }, "source": "pandas", "title": "pandas", "version": 1 }, "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
metricvalue
0Precision0.857842
1Recall0.718310
2F10.781900
3mAP@0.500.822248
4mAP@0.50:0.950.571469
\n", "
" ], "text/plain": [ " metric value\n", "0 Precision 0.857842\n", "1 Recall 0.718310\n", "2 F1 0.781900\n", "3 mAP@0.50 0.822248\n", "4 mAP@0.50:0.95 0.571469" ] }, "execution_count": null, "metadata": {}, "output_type": "execute_result" } ], "source": [ "\n", "best_weights = Path(train_results.save_dir) / 'weights' / 'best.pt'\n", "if not best_weights.exists():\n", " raise FileNotFoundError(f'Не найдены веса: {best_weights}')\n", "\n", "best_model = YOLO(str(best_weights))\n", "val_metrics = best_model.val(data=str(DATA_YAML), split='val', imgsz=IMG_SIZE, plots=False)\n", "\n", "precision = float(val_metrics.box.mp)\n", "recall = float(val_metrics.box.mr)\n", "map50 = float(val_metrics.box.map50)\n", "map5095 = float(val_metrics.box.map)\n", "f1 = 2 * precision * recall / (precision + recall + 1e-12)\n", "\n", "metrics_df = pd.DataFrame({\n", " 'metric': ['Precision', 'Recall', 'F1', 'mAP@0.50', 'mAP@0.50:0.95'],\n", " 'value': [precision, recall, f1, map50, map5095]\n", "})\n", "metrics_df\n" ] }, { "cell_type": "code", "execution_count": 15, "id": "4151f174", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Сохранено: /home/konnilol/Downloads/work_yolov10_cars/yolov10_cars_metrics.csv\n" ] } ], "source": [ "\n", "metrics_csv = WORKDIR / 'yolov10_cars_metrics.csv'\n", "metrics_df.to_csv(metrics_csv, index=False)\n", "print('Сохранено:', metrics_csv)\n" ] }, { "cell_type": "markdown", "id": "9fb2bcbb", "metadata": {}, "source": [ "\n", "## 7. Небольшая оценка задержки инференса\n" ] }, { "cell_type": "code", "execution_count": 16, "id": "e23ca3b1", "metadata": {}, "outputs": [ { "data": { "application/vnd.positron.dataexplorer+json": { "comm_id": "79424d58-c09e-4e89-819f-01b750bea844", "shape": { "columns": 2, "rows": 2 }, "source": "pandas", "title": "pandas", "version": 1 }, "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
metricvalue
0mean_latency_ms8.742471
1fps114.384141
\n", "
" ], "text/plain": [ " metric value\n", "0 mean_latency_ms 8.742471\n", "1 fps 114.384141" ] }, "execution_count": null, "metadata": {}, "output_type": "execute_result" } ], "source": [ "\n", "sample_images = sorted((CONVERTED / 'images' / 'val').glob('*'))[:50]\n", "latencies = []\n", "for p in sample_images:\n", " t0 = time.perf_counter()\n", " _ = best_model.predict(source=str(p), imgsz=IMG_SIZE, verbose=False, save=False)\n", " latencies.append((time.perf_counter() - t0) * 1000.0)\n", "\n", "lat_ms = float(np.mean(latencies)) if latencies else float('nan')\n", "fps = 1000.0 / lat_ms if lat_ms and not math.isnan(lat_ms) else float('nan')\n", "\n", "speed_df = pd.DataFrame({'metric': ['mean_latency_ms', 'fps'], 'value': [lat_ms, fps]})\n", "speed_df\n" ] }, { "cell_type": "markdown", "id": "ba28ecf1", "metadata": {}, "source": [ "\n", "## 8. Визуализация предсказаний\n" ] }, { "cell_type": "code", "execution_count": 17, "id": "76c3b64e", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "
" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "\n", "vis_images = sorted((CONVERTED / 'images' / 'val').glob('*'))[:6]\n", "results = best_model.predict([str(p) for p in vis_images], imgsz=IMG_SIZE, conf=0.25, verbose=False, save=False)\n", "\n", "plt.figure(figsize=(14, 10))\n", "for i, res in enumerate(results, start=1):\n", " plt.subplot(2, 3, i)\n", " img = res.plot()[:, :, ::-1]\n", " plt.imshow(img)\n", " plt.axis('off')\n", " plt.title(Path(res.path).name, fontsize=9)\n", "plt.tight_layout()\n", "plt.show()\n" ] }, { "cell_type": "markdown", "id": "71b0ec75", "metadata": {}, "source": [ "Архитектура.\n", "В работе использована YOLOv10n из статьи YOLOv10: Real-Time End-to-End Object Detection. Модель сохраняет схему backbone → neck → detection head, но ключевая новизна — end-to-end детекция без NMS за счёт механизма consistent dual assignments. Дополнительно применяются облегчённая классификационная голова, улучшенное понижение разрешения, крупные ядра свёртки и PSA-блоки для более эффективного извлечения признаков.\n", "\n", "Результаты на датасете Cars.\n", "После обучения получены:\n", "\n", "Precision = 0.858\n", "\n", "Recall = 0.718\n", "\n", "F1 = 0.782\n", "\n", "mAP@0.50 = 0.822\n", "\n", "mAP@0.50:0.95 = 0.571\n", "\n", "Средняя скорость инференса:\n", "\n", "8.74 мс на изображение\n", "\n", "около 114.4 FPS\n", "\n", "Вывод.\n", "YOLOv10 показала хорошее качество детекции при очень высокой скорости работы. По mAP@0.50 модель уже достаточно точна для практического применения, а по mAP@0.50:0.95 видно, что локализация не идеальна, но остаётся уверенной. В целом архитектура выглядит перспективной для задач обнаружения объектов в реальном времени, где важен компромисс между точностью и скоростью." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }