331 lines
11 KiB
Plaintext
331 lines
11 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"# 02 — Trénování YOLOv8 na VisDrone\n",
|
||
"\n",
|
||
"Trénujeme YOLOv8s (small) fine-tuned z ImageNet vah na VisDrone dataset.\n",
|
||
"Model detekuje 4 třídy vozidel z leteckých snímků.\n",
|
||
"\n",
|
||
"**GPU doporučeno** (trénink na CPU trvá ~10× déle)."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-8d392d",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Kontrola prostředí\n",
|
||
"\n",
|
||
"Zjistíme, zda je k dispozici GPU (CUDA) nebo Apple Silicon (MPS), a vypíšeme verzi PyTorche. Trénink na GPU je přibližně 10× rychlejší než na CPU."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable.It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning.\u001b[0m\u001b[33m\n",
|
||
"\u001b[0m\n",
|
||
"\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m A new release of pip is available: \u001b[0m\u001b[31;49m24.2\u001b[0m\u001b[39;49m -> \u001b[0m\u001b[32;49m26.0.1\u001b[0m\n",
|
||
"\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m To update, run: \u001b[0m\u001b[32;49mpython -m pip install --upgrade pip\u001b[0m\n",
|
||
"PyTorch: 2.4.1+cu124\n",
|
||
"CUDA dostupná: True\n",
|
||
"GPU: NVIDIA GeForce RTX 4090\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"%pip install ultralytics --quiet\n",
|
||
"\n",
|
||
"import torch\n",
|
||
"print(f\"PyTorch: {torch.__version__}\")\n",
|
||
"print(f\"CUDA dostupná: {torch.cuda.is_available()}\")\n",
|
||
"if torch.cuda.is_available():\n",
|
||
" print(f\"GPU: {torch.cuda.get_device_name(0)}\")\n",
|
||
"elif torch.backends.mps.is_available():\n",
|
||
" print(\"MPS (Apple Silicon) dostupné\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-af9159",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Načtení předtrénovaného modelu\n",
|
||
"\n",
|
||
"Použijeme YOLOv8s (small) s váhami předtrénovanými na ImageNet. Tento model budeme fine-tunovat na VisDrone dataset."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Creating new Ultralytics Settings v0.0.6 file ✅ \n",
|
||
"View Ultralytics Settings with 'yolo settings' or at '/root/.config/Ultralytics/settings.json'\n",
|
||
"Update Settings with 'yolo settings key=value', i.e. 'yolo settings runs_dir=path/to/dir'. For help see https://docs.ultralytics.com/quickstart/#ultralytics-settings.\n",
|
||
"Downloading https://github.com/ultralytics/assets/releases/download/v8.4.0/yolov8s.pt to 'yolov8s.pt': 100% ━━━━━━━━━━━━ 21.5MB 60.0MB/s 0.4s\n",
|
||
"YOLOv8s summary: 129 layers, 11,166,560 parameters, 0 gradients, 28.8 GFLOPs\n",
|
||
"Model načten: (129, 11166560, 0, 28.816844800000002)\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"from ultralytics import YOLO\n",
|
||
"from pathlib import Path\n",
|
||
"\n",
|
||
"YAML = Path(\"data/yolo_visdrone/dataset.yaml\")\n",
|
||
"assert YAML.exists(), f\"Nejprve spusť 01_dataset_prep.ipynb! Chybí: {YAML}\"\n",
|
||
"\n",
|
||
"model = YOLO(\"models/yolov8s.pt\") \n",
|
||
"print(\"Model načten:\", model.info())"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-58715d",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Trénování modelu\n",
|
||
"\n",
|
||
"Spustíme trénink na 50 epoch s early stoppingem (patience=10). Augmentace (rotace, překlápění, mosaic) pomáhají modelu generalizovat na různé letecké pohledy a světelné podmínky."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import torch\n",
|
||
"from pathlib import Path\n",
|
||
"\n",
|
||
"\n",
|
||
"if torch.cuda.is_available():\n",
|
||
" DEVICE = \"cuda\"\n",
|
||
"elif torch.backends.mps.is_available():\n",
|
||
" DEVICE = \"mps\"\n",
|
||
"else:\n",
|
||
" DEVICE = \"cpu\"\n",
|
||
"\n",
|
||
"print(f\"Trénink na: {DEVICE}\")\n",
|
||
"\n",
|
||
"results = model.train(\n",
|
||
" data=str(YAML),\n",
|
||
" epochs=50,\n",
|
||
" imgsz=640,\n",
|
||
" batch=16,\n",
|
||
" device=DEVICE,\n",
|
||
" name=\"visdrone_vehicles\",\n",
|
||
" project=str(Path(\"runs/train\").resolve()),\n",
|
||
" patience=10, # early stopping\n",
|
||
" save=True,\n",
|
||
" save_period=10,\n",
|
||
" val=True,\n",
|
||
" plots=True,\n",
|
||
" degrees=15.0,\n",
|
||
" flipud=0.5,\n",
|
||
" fliplr=0.5,\n",
|
||
" mosaic=1.0,\n",
|
||
" scale=0.5,\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-1fc179",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Výsledky tréninku\n",
|
||
"\n",
|
||
"Zobrazíme grafy průběhu trénování (loss, mAP), matici záměn a příklady predikcí na validační sadě."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import matplotlib.pyplot as plt\n",
|
||
"import matplotlib.image as mpimg\n",
|
||
"from pathlib import Path\n",
|
||
"import glob\n",
|
||
"\n",
|
||
"run_dir = Path(\"runs/train/visdrone_vehicles\")\n",
|
||
"\n",
|
||
"for plot_name in [\"results.png\", \"confusion_matrix.png\", \"PR_curve.png\", \"val_batch0_pred.jpg\"]:\n",
|
||
" p = run_dir / plot_name\n",
|
||
" if p.exists():\n",
|
||
" fig, ax = plt.subplots(figsize=(12, 6))\n",
|
||
" ax.imshow(mpimg.imread(p))\n",
|
||
" ax.axis(\"off\")\n",
|
||
" ax.set_title(plot_name)\n",
|
||
" plt.tight_layout()\n",
|
||
" plt.show()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-021cf0",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Validace nejlepšího modelu\n",
|
||
"\n",
|
||
"Načteme nejlepší checkpoint z tréninku a spustíme plnou validaci na validační sadě. Sledujeme metriku mAP50, která je standardem pro detekci objektů."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# Validace nejlepšího modelu\n",
|
||
"best_weights = Path(\"runs/train/visdrone_vehicles/weights/best.pt\")\n",
|
||
"assert best_weights.exists(), \"Trénink ještě neproběhl nebo selhal\"\n",
|
||
"\n",
|
||
"model_best = YOLO(str(best_weights))\n",
|
||
"val_results = model_best.val(data=str(YAML), imgsz=640, split=\"val\")\n",
|
||
"\n",
|
||
"print(f\"\\nmAP50: {val_results.box.map50:.4f}\")\n",
|
||
"print(f\"mAP50-95: {val_results.box.map:.4f}\")\n",
|
||
"for i, cls in enumerate([\"car\", \"van\", \"truck\", \"bus\"]):\n",
|
||
" ap = val_results.box.ap50[i] if i < len(val_results.box.ap50) else float('nan')\n",
|
||
" print(f\" AP50[{cls}]: {ap:.4f}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-f05d8c",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Fine-tuning na vlastních datech (Vrchlabí)\n",
|
||
"\n",
|
||
"Pokud je k dispozici lokálně anotovaný dataset z Vrchlabí, provedeme fine-tuning modelu natrénovaného na VisDrone. Tímto krokem model přizpůsobíme konkrétním podmínkám českých měst."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import zipfile\n",
|
||
"import shutil\n",
|
||
"from pathlib import Path\n",
|
||
"\n",
|
||
"VRCHLABI_ZIP = Path('data/dataset_vrchlabi.zip')\n",
|
||
"VRCHLABI_YAML = Path('data/vrchlabi/dataset.yaml')\n",
|
||
"\n",
|
||
"if VRCHLABI_ZIP.exists() and not VRCHLABI_YAML.exists():\n",
|
||
" print('Rozbaluji dataset_vrchlabi.zip ...')\n",
|
||
" with zipfile.ZipFile(VRCHLABI_ZIP) as zf:\n",
|
||
" zf.extractall('.data/')\n",
|
||
" print('Hotovo.')\n",
|
||
"\n",
|
||
"if not VRCHLABI_YAML.exists():\n",
|
||
" print('Vrchlabí dataset nenalezen — fine-tuning přeskočen.')\n",
|
||
" print('Spusť scripts/prepare_dataset.py nebo nahraj dataset_vrchlabi.zip.')\n",
|
||
"else:\n",
|
||
" print(f'Fine-tuning na: {VRCHLABI_YAML}')\n",
|
||
"\n",
|
||
" backup = best_weights.parent / 'best_before_finetune.pt'\n",
|
||
" shutil.copy(best_weights, backup)\n",
|
||
" print(f'Záloha uložena: {backup}')\n",
|
||
"\n",
|
||
" vrchlabi_model = YOLO(str(best_weights))\n",
|
||
"\n",
|
||
" vrchlabi_results = vrchlabi_model.train(\n",
|
||
" data=str(VRCHLABI_YAML),\n",
|
||
" epochs=30,\n",
|
||
" imgsz=256,\n",
|
||
" batch=32,\n",
|
||
" device=DEVICE,\n",
|
||
" name='vrchlabi_finetune',\n",
|
||
" project=str(Path(\"runs/train\").resolve()),\n",
|
||
" patience=10,\n",
|
||
" save=True,\n",
|
||
" val=True,\n",
|
||
" plots=True,\n",
|
||
" degrees=10.0,\n",
|
||
" flipud=0.5,\n",
|
||
" fliplr=0.5,\n",
|
||
" mosaic=0.5,\n",
|
||
" lr0=0.001,\n",
|
||
" )\n",
|
||
"\n",
|
||
" best_weights = Path(vrchlabi_results.save_dir) / 'weights' / 'best.pt'\n",
|
||
" print(f'Vrchlabí fine-tuning hotov. Nejlepší váhy: {best_weights}')"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "md-805d37",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Uložení konfigurace modelu\n",
|
||
"\n",
|
||
"Cestu k nejlepšímu modelu zapíšeme do `models/model_config.json`, odkud ji načtou následující notebooky."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Nejlepší model uložen v: /workspace/runs/detect/runs/detect/runs/train/vrchlabi_finetune-2/weights/best.pt\n",
|
||
"Konfigurace zapsána do: model_config.json\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"import json\n",
|
||
"\n",
|
||
"config = {\"best_model\": str(best_weights.resolve())}\n",
|
||
"with open(\"models/model_config.json\", \"w\") as f:\n",
|
||
" json.dump(config, f)\n",
|
||
"\n",
|
||
"print(f\"Nejlepší model uložen v: {best_weights}\")\n",
|
||
"print(f\"Konfigurace zapsána do: models/model_config.json\")"
|
||
]
|
||
}
|
||
],
|
||
"metadata": {
|
||
"kernelspec": {
|
||
"display_name": ".venv (3.14.4)",
|
||
"language": "python",
|
||
"name": "python3"
|
||
},
|
||
"language_info": {
|
||
"codemirror_mode": {
|
||
"name": "ipython",
|
||
"version": 3
|
||
},
|
||
"file_extension": ".py",
|
||
"mimetype": "text/x-python",
|
||
"name": "python",
|
||
"nbconvert_exporter": "python",
|
||
"pygments_lexer": "ipython3",
|
||
"version": "3.14.4"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 4
|
||
}
|