Fix training lifecycle and dataset splitting bugs

.agents/PROJECT_ISSUES.md

@@ -0,0 +1,73 @@
+# Исправленные проблемы проекта YOLO Train TUI
+
+Дата исправления и повторной проверки: 2026-07-17
+
+## Итог
+
+Все 11 ранее зафиксированных дефектов исправлены и покрыты регрессионными
+проверками.
+
+| ID | Приоритет | Статус | Исправление |
+|---|---|---|---|
+| BUG-001 | Критический | Исправлено | `subprocess_runner.main()` возвращает код, а `SystemExit` создаётся только снаружи обрабатывающего блока |
+| BUG-002 | Высокий | Исправлено | Перед каждым запуском `prepare_run()` сбрасывает состояние остановки |
+| BUG-003 | Высокий | Исправлено | Родитель отправляет кооперативный сигнал; принудительный `kill()` используется только после таймаута |
+| BUG-004 | Высокий | Исправлено | Запрос, сделанный до готовности subprocess, сохраняется и доставляется после маркера `READY` |
+| BUG-005 | Высокий | Исправлено | Detection-style авторазбиение запрещено для `classify` в UI и конфигурации |
+| BUG-006 | Средний | Исправлено | `.yaml`/`.yml` разбираются через `yaml.safe_load()`, поле `names` валидируется |
+| BUG-007 | Средний | Исправлено | Датасет с одним изображением отклоняется с понятной ошибкой |
+| BUG-008 | Средний | Исправлено | Изображения и метки ищутся рекурсивно с сохранением вложенных путей |
+| BUG-009 | Средний | Исправлено | Каждый результат создаётся в уникальном `.yolo-tui/splits/<id>` без перезаписи пользовательского `split/` |
+| BUG-010 | Низкий | Исправлено | Traceback выводится в журнал TUI; абсолютный путь другого пользователя удалён |
+| BUG-011 | Низкий | Исправлено | Явно указанный отсутствующий или некорректный файл классов вызывает точную ошибку без fallback |
+
+## Жизненный цикл обучения
+
+- Дочерний процесс устанавливает обработчики остановки и только затем печатает
+  `__YOLO_TUI_READY__`.
+- Если пользователь нажал «Остановить» раньше, родитель запоминает запрос и
+  отправляет его после получения маркера готовности.
+- Дочерний `TrainingRunner` устанавливает `trainer.stop = True`; Ultralytics
+  останавливается между пакетами данных, затем выполняет штатную финализацию и
+  завершающие callbacks.
+- Если процесс не завершился за 30 секунд, используется принудительный fallback.
+- После завершения ссылка на subprocess и таймер очищаются; перед следующим
+  запуском флаг остановки сбрасывается.
+
+## Работа с датасетами
+
+- Текущий splitter предназначен для задач `detect`, `segment`, `pose` и `obb`
+  со структурой `images/` + `labels/`.
+- Для `classify` требуется готовый каталог с `train`/`val` и подкаталогами
+  классов; несовместимый переключатель в UI отключён.
+- Поддерживаются `classes.txt`, `.yaml` и `.yml`; YAML может хранить `names` как
+  список или словарь с последовательными ID от 0.
+- Явный путь к классам считается обязательным и не заменяется автопоиском при
+  опечатке или ошибке формата.
+- Split требует минимум два изображения и рекурсивно обрабатывает вложенные
+  каталоги.
+- Файлы каждого запуска создаются эксклюзивно в отдельном управляемом каталоге.
+
+## Проверка
+
+Выполнены команды:
+
+```text
+uv run pytest -q
+uv run python -m compileall -q src tests
+git diff --check
+```
+
+Результат: `38 passed`; ошибок компиляции и форматирования diff нет.
+
+Регрессионные тесты проверяют:
+
+1. успешный и ошибочный коды `subprocess_runner.main()`;
+2. сброс остановки между запусками;
+3. кооперативный сигнал вместо немедленного `terminate()`;
+4. доставку раннего запроса после готовности subprocess;
+5. запрет авторазбиения для `classify`;
+6. пользовательские YAML-файлы классов и ошибочный явный путь;
+7. датасеты из одного и двух изображений;
+8. вложенные изображения и метки;
+9. сохранность пользовательского каталога `split/` и уникальность результатов.

src/yolo_tui/app.py

@@ -329,9 +329,26 @@ class YoloTrainApp(App[None]):
 
     @on(Switch.Changed, "#split-enabled")
     def toggle_split(self, event: Switch.Changed) -> None:
+        if self.query_one("#task", Select).value == "classify" and event.value:
+            self.query_one("#split-enabled", Switch).value = False
+            self.notify(
+                "Для classify укажите готовый каталог с train/val по классам.",
+                severity="warning",
+            )
+            return
         for control in self.query(".split-field Input"):
             control.disabled = not event.value
 
+    @on(Select.Changed, "#task")
+    def task_changed(self, event: Select.Changed) -> None:
+        split_switch = self.query_one("#split-enabled", Switch)
+        is_classify = event.value == "classify"
+        if is_classify and split_switch.value:
+            split_switch.value = False
+        split_switch.disabled = is_classify
+        for control in self.query(".split-field Input"):
+            control.disabled = is_classify or not split_switch.value
+
     @on(Switch.Changed, "#mlflow-enabled")
     def toggle_mlflow(self, event: Switch.Changed) -> None:
         for widget_id in ("tracking-uri", "experiment-name", "mlflow-run-name"):
@@ -363,6 +380,7 @@ class YoloTrainApp(App[None]):
             self.notify(str(exc), title="Проверьте параметры", severity="error")
             return
 
+        self.runner.prepare_run()
         self._set_running(True)
         progress = self.query_one("#progress", ProgressBar)
         progress.update(total=config.epochs, progress=0)
@@ -387,7 +405,7 @@ class YoloTrainApp(App[None]):
             return
         self.runner.request_stop()
         self.query_one("#status-title", Static).update("ОСТАНОВКА")
-        self.query_one("#status-text", Static).update("Завершаю текущую эпоху…")
+        self.query_one("#status-text", Static).update("Корректно останавливаю обучение…")
         self.query_one("#stop-button", Button).disabled = True
         self._write_log("[yellow]Запрошена остановка обучения.[/yellow]")
 
@@ -415,7 +433,7 @@ class YoloTrainApp(App[None]):
                 text=True,
                 bufsize=1,
             )
-            self.runner.set_subprocess(process)
+            self.runner.set_subprocess(process, ready=False)
 
             output_dir = None
 
@@ -427,7 +445,9 @@ class YoloTrainApp(App[None]):
                 if not line_str:
                     continue
 
-                if line_str.startswith("__YOLO_TUI_EVENT__:"):
+                if line_str == "__YOLO_TUI_READY__":
+                    self.runner.mark_subprocess_ready()
+                elif line_str.startswith("__YOLO_TUI_EVENT__:"):
                     try:
                         event_data = json.loads(line_str[len("__YOLO_TUI_EVENT__:"):])
                         event = TrainingEvent(
@@ -446,6 +466,7 @@ class YoloTrainApp(App[None]):
 
             process.wait()
             rc = process.returncode
+            self.runner.clear_subprocess()
 
             if rc == 0:
                 self.app.call_from_thread(self._training_finished, output_dir)
@@ -458,13 +479,21 @@ class YoloTrainApp(App[None]):
                         Exception("Процесс обучения завершился с ошибкой. Проверьте логи выше."),
                     )
 
-        except BaseException as exc:
+        except Exception as exc:
             import traceback
-            try:
-                with open("/Users/vadim/.gemini/antigravity/brain/ab216120-71db-4a55-a2bf-2eff6a9caf26/error.log", "w", encoding="utf-8") as f:
-                    traceback.print_exc(file=f)
-            except Exception:
-                pass
+            if process is not None:
+                try:
+                    if process.poll() is None:
+                        process.kill()
+                    process.wait(timeout=5)
+                except Exception:
+                    pass
+            self.runner.clear_subprocess()
+            details = escape(traceback.format_exc())
+            self.app.call_from_thread(
+                self._write_log,
+                f"[red]{details}[/red]",
+            )
             self.app.call_from_thread(self._training_failed, exc)
         finally:
             if temp_config_path and os.path.exists(temp_config_path):

src/yolo_tui/config.py

@@ -157,6 +157,11 @@ class TrainingConfig:
             raise ValueError("Укажите путь или имя модели.")
         if self.task not in SUPPORTED_TASKS:
             raise ValueError(f"Неизвестный тип задачи: {self.task}.")
+        if self.task == "classify" and self.split.enabled:
+            raise ValueError(
+                "Автоматическое разделение доступно только для YOLO-датасетов "
+                "с папками images/labels и не поддерживает задачу classify."
+            )
         if self.epochs < 1:
             raise ValueError("Количество эпох должно быть не меньше 1.")
         if self.image_size < 32:

src/yolo_tui/dataset_splitter.py

@@ -2,138 +2,217 @@ from __future__ import annotations
 
 import random
 from pathlib import Path
+from typing import Any
+from uuid import uuid4
+
 import yaml
 
 
+def _normalize_names(names: Any, source: Path) -> dict[int, str]:
+    if isinstance(names, list):
+        items = enumerate(names)
+    elif isinstance(names, dict):
+        normalized_items: list[tuple[int, Any]] = []
+        seen: set[int] = set()
+        for raw_key, value in names.items():
+            if isinstance(raw_key, bool):
+                raise ValueError(f"Некорректный ID класса в '{source}': {raw_key!r}.")
+            try:
+                class_id = int(raw_key)
+            except (TypeError, ValueError) as exc:
+                raise ValueError(
+                    f"Некорректный ID класса в '{source}': {raw_key!r}."
+                ) from exc
+            if class_id in seen:
+                raise ValueError(f"Повторяющийся ID класса {class_id} в '{source}'.")
+            seen.add(class_id)
+            normalized_items.append((class_id, value))
+        items = normalized_items
+    else:
+        raise ValueError(
+            f"Поле names в '{source}' должно быть списком или словарём."
+        )
+
+    result: dict[int, str] = {}
+    for class_id, raw_name in items:
+        if isinstance(raw_name, (dict, list)):
+            name = ""
+        else:
+            name = str(raw_name).strip()
+        if not name:
+            raise ValueError(f"Пустое имя класса {class_id} в '{source}'.")
+        result[class_id] = name
+
+    if not result:
+        raise ValueError(f"Список классов в '{source}' пуст.")
+    expected_ids = list(range(len(result)))
+    if sorted(result) != expected_ids:
+        raise ValueError(
+            f"ID классов в '{source}' должны идти подряд, начиная с 0."
+        )
+    return dict(sorted(result.items()))
+
+
+def _parse_text_classes(path: Path) -> dict[int, str]:
+    with path.open("r", encoding="utf-8") as classes_file:
+        lines = [line.strip() for line in classes_file if line.strip()]
+    return _normalize_names(lines, path)
+
+
+def _load_yaml(path: Path) -> Any:
+    try:
+        with path.open("r", encoding="utf-8") as yaml_file:
+            return yaml.safe_load(yaml_file)
+    except yaml.YAMLError as exc:
+        raise ValueError(f"Не удалось разобрать YAML классов '{path}': {exc}.") from exc
+
+
+def _parse_yaml_classes(path: Path) -> dict[int, str]:
+    data = _load_yaml(path)
+    if not isinstance(data, dict) or "names" not in data:
+        raise ValueError(f"В YAML-файле '{path}' отсутствует поле names.")
+    return _normalize_names(data["names"], path)
+
+
 def _parse_classes_file(path: Path) -> dict[int, str]:
-    with open(path, "r", encoding="utf-8") as f:
-        lines = [line.strip() for line in f if line.strip()]
-    return {i: name for i, name in enumerate(lines)}
+    suffix = path.suffix.lower()
+    if suffix == ".txt":
+        return _parse_text_classes(path)
+    if suffix in {".yaml", ".yml"}:
+        return _parse_yaml_classes(path)
+    raise ValueError(
+        f"Неподдерживаемый формат файла классов '{path}'. "
+        "Используйте .txt, .yaml или .yml."
+    )
 
 
 def read_classes(dataset_dir: Path, custom_classes_path: str) -> dict[int, str]:
-    # 1. Custom path specified by user
+    # An explicit path is authoritative: typos and malformed files must not fall back.
     if custom_classes_path.strip():
-        path = Path(custom_classes_path.strip())
-        if path.exists():
-            return _parse_classes_file(path)
-
-    # 2. classes.txt in dataset_dir root
-    path = dataset_dir / "classes.txt"
-    if path.exists():
-        return _parse_classes_file(path)
-
-    # 3. classes.txt inside dataset_dir/labels
-    path = dataset_dir / "labels" / "classes.txt"
-    if path.exists():
-        return _parse_classes_file(path)
-
-    # 4. Any .yaml file in dataset_dir (excluding split/dataset.yaml)
-    yaml_files = list(dataset_dir.glob("*.yaml"))
-    yaml_files = [f for f in yaml_files if f.name != "dataset.yaml"]
-    if yaml_files:
+        path = Path(custom_classes_path.strip()).expanduser()
+        if not path.exists():
+            raise ValueError(f"Указанный файл классов '{path}' не существует.")
+        if not path.is_file():
+            raise ValueError(f"Путь к классам '{path}' не является файлом.")
         try:
-            with open(yaml_files[0], "r", encoding="utf-8") as f:
-                data = yaml.safe_load(f)
-                if isinstance(data, dict) and "names" in data:
-                    names = data["names"]
-                    if isinstance(names, dict):
-                        return {int(k): str(v) for k, v in names.items()}
-                    elif isinstance(names, list):
-                        return {i: str(v) for i, v in enumerate(names)}
-        except Exception:
-            pass
+            return _parse_classes_file(path)
+        except OSError as exc:
+            raise ValueError(f"Не удалось прочитать файл классов '{path}': {exc}.") from exc
 
-    # 5. Fallback: Scan label files to determine number of classes and use class_i names
-    class_ids = set()
+    for path in (dataset_dir / "classes.txt", dataset_dir / "labels" / "classes.txt"):
+        if path.is_file():
+            return _parse_text_classes(path)
+
+    yaml_files = sorted(
+        (*dataset_dir.glob("*.yaml"), *dataset_dir.glob("*.yml")),
+        key=lambda item: item.name,
+    )
+    for path in yaml_files:
+        try:
+            data = _load_yaml(path)
+        except (OSError, ValueError):
+            continue
+        if isinstance(data, dict) and "names" in data:
+            return _normalize_names(data["names"], path)
+
+    # Infer the range from every nested label file when no class list is available.
+    class_ids: set[int] = set()
     labels_dir = dataset_dir / "labels"
     if labels_dir.exists():
-        for txt_file in labels_dir.glob("*.txt"):
+        for txt_file in labels_dir.rglob("*.txt"):
             if txt_file.name == "classes.txt":
                 continue
             try:
-                with open(txt_file, "r", encoding="utf-8") as f:
-                    for line in f:
+                with txt_file.open("r", encoding="utf-8") as label_file:
+                    for line in label_file:
                         parts = line.strip().split()
                         if parts:
                             class_ids.add(int(parts[0]))
-            except Exception:
-                pass
+            except (OSError, ValueError):
+                continue
 
     if class_ids:
         max_id = max(class_ids)
-        return {i: f"class_{i}" for i in range(max_id + 1)}
+        return {class_id: f"class_{class_id}" for class_id in range(max_id + 1)}
 
     raise ValueError(
-        "Не удалось найти список классов. Пожалуйста, создайте файл classes.txt "
-        "в корневой папке датасета или укажите путь к нему."
+        "Не удалось найти список классов. Создайте classes.txt или YAML с полем "
+        "names в корневой папке датасета либо укажите путь к нему."
     )
 
 
+def _write_new(path: Path, content: str) -> None:
+    with path.open("x", encoding="utf-8") as output_file:
+        output_file.write(content)
+
+
 def split_dataset(
     dataset_dir: str, train_ratio: float, classes_path: str
 ) -> tuple[int, int, str]:
-    base_dir = Path(dataset_dir.strip()).absolute()
+    if not dataset_dir.strip():
+        raise ValueError("Укажите каталог датасета.")
+    if not 0.1 <= train_ratio <= 0.95:
+        raise ValueError("Доля обучающей выборки должна быть от 0.1 до 0.95.")
+
+    base_dir = Path(dataset_dir.strip()).expanduser().absolute()
     images_dir = base_dir / "images"
     labels_dir = base_dir / "labels"
 
-    if not base_dir.exists():
+    if not base_dir.is_dir():
         raise ValueError(f"Каталог датасета '{base_dir}' не существует.")
-    if not images_dir.exists():
+    if not images_dir.is_dir():
         raise ValueError(f"Папка с изображениями '{images_dir}' не найдена.")
-    if not labels_dir.exists():
+    if not labels_dir.is_dir():
         raise ValueError(f"Папка с разметкой '{labels_dir}' не найдена.")
 
-    # Find images
     valid_extensions = {".jpg", ".jpeg", ".png", ".bmp", ".webp", ".tif", ".tiff"}
-    image_files = [
-        f for f in images_dir.iterdir()
-        if f.is_file() and f.suffix.lower() in valid_extensions
-    ]
+    image_files = sorted(
+        (
+            path
+            for path in images_dir.rglob("*")
+            if path.is_file() and path.suffix.lower() in valid_extensions
+        ),
+        key=lambda item: item.as_posix(),
+    )
 
     if not image_files:
         raise ValueError(f"В папке '{images_dir}' не найдено изображений.")
+    if len(image_files) == 1:
+        raise ValueError(
+            "Для разделения нужно минимум 2 изображения; найдено: 1."
+        )
 
-    # Shuffle deterministically using a fixed seed
     rng = random.Random(42)
     rng.shuffle(image_files)
 
     split_idx = int(len(image_files) * train_ratio)
-    if split_idx == 0:
-        split_idx = 1
-    if split_idx >= len(image_files):
-        split_idx = len(image_files) - 1
-
+    split_idx = max(1, min(split_idx, len(image_files) - 1))
     train_images = image_files[:split_idx]
     val_images = image_files[split_idx:]
 
-    split_dir = base_dir / "split"
-    split_dir.mkdir(parents=True, exist_ok=True)
+    # Resolve classes before creating output so invalid input leaves no partial split.
+    classes = read_classes(base_dir, classes_path)
+
+    relative_split_dir = Path(".yolo-tui") / "splits" / uuid4().hex
+    split_dir = base_dir / relative_split_dir
+    split_dir.mkdir(parents=True, exist_ok=False)
 
     train_txt_path = split_dir / "train.txt"
     val_txt_path = split_dir / "val.txt"
-
-    with open(train_txt_path, "w", encoding="utf-8") as f:
-        for img in train_images:
-            f.write(f"{img}\n")
-
-    with open(val_txt_path, "w", encoding="utf-8") as f:
-        for img in val_images:
-            f.write(f"{img}\n")
-
-    # Load class mapping
-    classes = read_classes(base_dir, classes_path)
-
-    # Write data.yaml equivalent
     dataset_yaml_path = split_dir / "dataset.yaml"
+
+    _write_new(train_txt_path, "".join(f"{image}\n" for image in train_images))
+    _write_new(val_txt_path, "".join(f"{image}\n" for image in val_images))
+
     dataset_data = {
         "path": str(base_dir),
-        "train": f"split/{train_txt_path.name}",
-        "val": f"split/{val_txt_path.name}",
+        "train": (relative_split_dir / train_txt_path.name).as_posix(),
+        "val": (relative_split_dir / val_txt_path.name).as_posix(),
         "names": classes,
     }
-
-    with open(dataset_yaml_path, "w", encoding="utf-8") as f:
-        yaml.safe_dump(dataset_data, f, allow_unicode=True, sort_keys=False)
+    _write_new(
+        dataset_yaml_path,
+        yaml.safe_dump(dataset_data, allow_unicode=True, sort_keys=False),
+    )
 
     return len(train_images), len(val_images), str(dataset_yaml_path)

src/yolo_tui/subprocess_runner.py

@@ -1,6 +1,10 @@
 import json
 import os
+import signal
 import sys
+import traceback
+from collections.abc import Iterator, Sequence
+from contextlib import contextmanager
 from pathlib import Path
 
 # Force headless Matplotlib to avoid any thread/process GUI issues
@@ -9,41 +13,66 @@ os.environ["MPLBACKEND"] = "Agg"
 from yolo_tui.config import TrainingConfig
 from yolo_tui.trainer import TrainingEvent, TrainingRunner
 
-def main():
-    if len(sys.argv) < 2:
-        print("Usage: python -m yolo_tui.subprocess_runner <config_json_path>", file=sys.stderr)
-        sys.exit(1)
 
-    config_path = sys.argv[1]
+@contextmanager
+def _stop_signal_handlers(runner: TrainingRunner) -> Iterator[None]:
+    previous: dict[signal.Signals, signal.Handlers] = {}
+
+    def request_stop(_signum: int, _frame: object) -> None:
+        runner.request_stop()
+
+    for signum in (signal.SIGTERM, signal.SIGINT):
+        previous[signum] = signal.getsignal(signum)
+        signal.signal(signum, request_stop)
     try:
-        with open(config_path, "r", encoding="utf-8") as f:
-            config_dict = json.load(f)
-        config = TrainingConfig.from_dict(config_dict)
-    except Exception as e:
-        print(f"Error loading config: {e}", file=sys.stderr)
-        sys.exit(1)
+        yield
+    finally:
+        for signum, handler in previous.items():
+            signal.signal(signum, handler)
 
+
+def main(argv: Sequence[str] | None = None) -> int:
+    args = list(sys.argv[1:] if argv is None else argv)
+    if not args:
+        print(
+            "Usage: python -m yolo_tui.subprocess_runner <config_json_path>",
+            file=sys.stderr,
+        )
+        return 1
+
+    config_path = Path(args[0])
     runner = TrainingRunner()
+    runner.prepare_run()
 
-    def handle_event(event: TrainingEvent) -> None:
-        event_dict = {
-            "kind": event.kind,
-            "message": event.message,
-            "epoch": event.epoch,
-            "total_epochs": event.total_epochs,
-        }
-        # Print structured JSON event so the parent process can parse it
-        print(f"__YOLO_TUI_EVENT__:{json.dumps(event_dict)}", flush=True)
+    with _stop_signal_handlers(runner):
+        # The parent waits for this marker before sending a cooperative signal.
+        print("__YOLO_TUI_READY__", flush=True)
+        try:
+            with config_path.open("r", encoding="utf-8") as config_file:
+                config_dict = json.load(config_file)
+            config = TrainingConfig.from_dict(config_dict)
+        except Exception as exc:
+            print(f"Error loading config: {exc}", file=sys.stderr)
+            return 1
 
-    try:
-        output_dir = runner.train(config, handle_event)
-        if output_dir:
-            print(f"__YOLO_TUI_RESULT__:{output_dir}", flush=True)
-        sys.exit(0)
-    except BaseException as e:
-        import traceback
-        traceback.print_exc()
-        sys.exit(1)
+        def handle_event(event: TrainingEvent) -> None:
+            event_dict = {
+                "kind": event.kind,
+                "message": event.message,
+                "epoch": event.epoch,
+                "total_epochs": event.total_epochs,
+            }
+            # Print structured JSON event so the parent process can parse it
+            print(f"__YOLO_TUI_EVENT__:{json.dumps(event_dict)}", flush=True)
+
+        try:
+            output_dir = runner.train(config, handle_event)
+            if output_dir:
+                print(f"__YOLO_TUI_RESULT__:{output_dir}", flush=True)
+            return 0
+        except Exception:
+            traceback.print_exc()
+            return 1
 
 if __name__ == "__main__":
-    main()
+    raise SystemExit(main())

src/yolo_tui/trainer.py

@@ -1,11 +1,12 @@
 from __future__ import annotations
 
 import os
+import signal
 from collections.abc import Callable, Iterator
 from contextlib import contextmanager
 from dataclasses import dataclass
 from pathlib import Path
-from threading import Event, Lock
+from threading import Event, RLock, Timer
 from typing import Any
 
 from .config import MlflowConfig, TrainingConfig
@@ -51,39 +52,108 @@ def mlflow_environment(config: MlflowConfig) -> Iterator[None]:
 class TrainingRunner:
     """Owns a single YOLO training run and exposes cooperative cancellation."""
 
+    FORCE_STOP_TIMEOUT_SECONDS = 30.0
+
     def __init__(self) -> None:
         self._model: Any | None = None
-        self._state_lock = Lock()
+        self._state_lock = RLock()
         self._stop_requested = Event()
         self._subprocess: Any | None = None
+        self._subprocess_ready = False
+        self._force_stop_timer: Timer | None = None
 
-    def set_subprocess(self, process: Any) -> None:
+    def prepare_run(self) -> None:
+        """Reset cancellation state before starting a new training run."""
+        with self._state_lock:
+            timer = self._force_stop_timer
+            self._force_stop_timer = None
+            self._subprocess_ready = False
+            self._stop_requested.clear()
+        if timer is not None:
+            timer.cancel()
+
+    def set_subprocess(self, process: Any, *, ready: bool = False) -> None:
+        """Register the child process without losing an earlier stop request."""
         with self._state_lock:
             self._subprocess = process
+            self._subprocess_ready = ready
+            stop_requested = self._stop_requested.is_set()
+        if stop_requested:
+            if ready:
+                self._send_cooperative_stop(process)
+            self._schedule_force_stop(process)
+
+    def mark_subprocess_ready(self) -> None:
+        """Mark the child signal handler as ready and deliver any pending stop."""
+        with self._state_lock:
+            process = self._subprocess
+            self._subprocess_ready = process is not None
+            stop_requested = self._stop_requested.is_set()
+        if process is not None and stop_requested:
+            self._send_cooperative_stop(process)
+            self._schedule_force_stop(process)
 
     def clear_subprocess(self) -> None:
         with self._state_lock:
             self._subprocess = None
+            self._subprocess_ready = False
+            timer = self._force_stop_timer
+            self._force_stop_timer = None
+        if timer is not None:
+            timer.cancel()
 
     def request_stop(self) -> None:
         self._stop_requested.set()
         with self._state_lock:
-            if self._subprocess is not None:
-                try:
-                    self._subprocess.terminate()
-                except Exception:
-                    pass
+            process = self._subprocess
+            process_ready = self._subprocess_ready
             trainer = getattr(self._model, "trainer", None)
-            if trainer is not None:
-                trainer.stop = True
+        if process is not None:
+            if process_ready:
+                self._send_cooperative_stop(process)
+            self._schedule_force_stop(process)
+        if trainer is not None:
+            trainer.stop = True
 
     @property
     def stop_requested(self) -> bool:
         return self._stop_requested.is_set()
 
+    @staticmethod
+    def _send_cooperative_stop(process: Any) -> None:
+        try:
+            process.send_signal(signal.SIGTERM)
+        except (AttributeError, OSError, ProcessLookupError):
+            try:
+                process.terminate()
+            except (AttributeError, OSError, ProcessLookupError):
+                pass
+
+    def _schedule_force_stop(self, process: Any) -> None:
+        with self._state_lock:
+            if self._subprocess is not process or self._force_stop_timer is not None:
+                return
+            timer = Timer(
+                self.FORCE_STOP_TIMEOUT_SECONDS,
+                self._force_stop,
+                args=(process,),
+            )
+            timer.daemon = True
+            self._force_stop_timer = timer
+        timer.start()
+
+    def _force_stop(self, process: Any) -> None:
+        with self._state_lock:
+            if self._subprocess is not process:
+                return
+        try:
+            if process.poll() is None:
+                process.kill()
+        except (AttributeError, OSError, ProcessLookupError):
+            pass
+
     def train(self, config: TrainingConfig, on_event: EventHandler) -> Path | None:
         config.validate()
-        self._stop_requested.clear()
 
         train_args = config.train_kwargs()
 

tests/test_app.py

@@ -98,3 +98,21 @@ def test_read_config_ignores_disabled_split() -> None:
             assert config.split.train_ratio == DatasetSplitConfig(enabled=False).train_ratio
 
     asyncio.run(exercise())
+
+
+def test_classify_disables_detection_style_split() -> None:
+    async def exercise() -> None:
+        app = YoloTrainApp()
+        async with app.run_test(size=(140, 45)) as pilot:
+            app.query_one("#split-enabled", Switch).value = True
+            await pilot.pause()
+
+            app.query_one("#task", Select).value = "classify"
+            await pilot.pause()
+
+            assert app.query_one("#split-enabled", Switch).value is False
+            assert app.query_one("#split-enabled", Switch).disabled is True
+            assert app.query_one("#split-ratio", Input).disabled is True
+            assert app.query_one("#split-classes", Input).disabled is True
+
+    asyncio.run(exercise())

tests/test_config.py

@@ -96,3 +96,15 @@ def test_dataset_split_ratio_is_validated(ratio: float) -> None:
     config = DatasetSplitConfig(enabled=True, train_ratio=ratio)
     with pytest.raises(ValueError, match="Доля обучающей выборки"):
         config.validate()
+
+
+def test_classification_rejects_detection_style_auto_split() -> None:
+    config = TrainingConfig(
+        dataset="classification-dataset",
+        model="model.pt",
+        task="classify",
+        split=DatasetSplitConfig(enabled=True),
+    )
+
+    with pytest.raises(ValueError, match="classify"):
+        config.validate()

tests/test_splitter.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
-import os
 from pathlib import Path
+
 import pytest
 import yaml
 
@@ -16,6 +16,32 @@ def test_read_classes_custom_path(tmp_path: Path) -> None:
     assert classes == {0: "classA", 1: "classB"}
 
 
+@pytest.mark.parametrize(
+    ("names", "expected"),
+    [
+        (["cat", "dog"], {0: "cat", 1: "dog"}),
+        ({0: "cat", 1: "dog"}, {0: "cat", 1: "dog"}),
+    ],
+)
+def test_read_classes_custom_yaml(
+    tmp_path: Path, names: object, expected: dict[int, str]
+) -> None:
+    custom_file = tmp_path / "custom_classes.yaml"
+    custom_file.write_text(
+        yaml.safe_dump({"names": names}, allow_unicode=True),
+        encoding="utf-8",
+    )
+
+    assert read_classes(tmp_path, str(custom_file)) == expected
+
+
+def test_missing_custom_classes_path_does_not_fall_back(tmp_path: Path) -> None:
+    (tmp_path / "classes.txt").write_text("fallback\n", encoding="utf-8")
+
+    with pytest.raises(ValueError, match="не существует"):
+        read_classes(tmp_path, str(tmp_path / "typo.yaml"))
+
+
 def test_read_classes_root_classes_txt(tmp_path: Path) -> None:
     classes_file = tmp_path / "classes.txt"
     classes_file.write_text("class0\nclass1\nclass2\n", encoding="utf-8")
@@ -84,13 +110,15 @@ def test_split_dataset_flow(tmp_path: Path) -> None:
     with open(yaml_path, "r", encoding="utf-8") as f:
         data = yaml.safe_load(f)
         assert data["path"] == str(tmp_path)
-        assert data["train"] == "split/train.txt"
-        assert data["val"] == "split/val.txt"
+        relative_output_dir = Path(yaml_path).parent.relative_to(tmp_path)
+        assert data["train"] == (relative_output_dir / "train.txt").as_posix()
+        assert data["val"] == (relative_output_dir / "val.txt").as_posix()
         assert data["names"] == {0: "dummy_class"}
 
     # Verify lists content
-    train_list = (tmp_path / "split" / "train.txt").read_text(encoding="utf-8").strip().split("\n")
-    val_list = (tmp_path / "split" / "val.txt").read_text(encoding="utf-8").strip().split("\n")
+    output_dir = Path(yaml_path).parent
+    train_list = (output_dir / "train.txt").read_text(encoding="utf-8").strip().split("\n")
+    val_list = (output_dir / "val.txt").read_text(encoding="utf-8").strip().split("\n")
 
     assert len(train_list) == 3
     assert len(val_list) == 1
@@ -98,3 +126,69 @@ def test_split_dataset_flow(tmp_path: Path) -> None:
     # Verify paths are absolute
     assert Path(train_list[0]).is_absolute()
     assert Path(val_list[0]).is_absolute()
+
+
+def test_split_dataset_rejects_single_image_without_writing_output(
+    tmp_path: Path,
+) -> None:
+    images_dir = tmp_path / "images"
+    labels_dir = tmp_path / "labels"
+    images_dir.mkdir()
+    labels_dir.mkdir()
+    (images_dir / "only.jpg").write_bytes(b"")
+    (labels_dir / "only.txt").write_text("0 0.5 0.5 1 1\n", encoding="utf-8")
+    (tmp_path / "classes.txt").write_text("item\n", encoding="utf-8")
+
+    with pytest.raises(ValueError, match="минимум 2"):
+        split_dataset(str(tmp_path), 0.8, "")
+
+    assert not (tmp_path / ".yolo-tui").exists()
+
+
+def test_split_dataset_finds_nested_images_and_labels(tmp_path: Path) -> None:
+    images_dir = tmp_path / "images" / "day"
+    labels_dir = tmp_path / "labels" / "day"
+    images_dir.mkdir(parents=True)
+    labels_dir.mkdir(parents=True)
+    for index in range(2):
+        (images_dir / f"nested-{index}.jpg").write_bytes(b"")
+        (labels_dir / f"nested-{index}.txt").write_text(
+            "2 0.5 0.5 0.2 0.2\n",
+            encoding="utf-8",
+        )
+
+    train_count, val_count, yaml_path = split_dataset(str(tmp_path), 0.8, "")
+
+    assert (train_count, val_count) == (1, 1)
+    data = yaml.safe_load(Path(yaml_path).read_text(encoding="utf-8"))
+    assert data["names"] == {0: "class_0", 1: "class_1", 2: "class_2"}
+    listed_images = "".join(
+        (Path(yaml_path).parent / filename).read_text(encoding="utf-8")
+        for filename in ("train.txt", "val.txt")
+    )
+    assert "images/day/nested-0.jpg" in listed_images
+    assert "images/day/nested-1.jpg" in listed_images
+
+
+def test_split_dataset_preserves_existing_split_files(tmp_path: Path) -> None:
+    images_dir = tmp_path / "images"
+    labels_dir = tmp_path / "labels"
+    images_dir.mkdir()
+    labels_dir.mkdir()
+    for index in range(2):
+        (images_dir / f"image-{index}.jpg").write_bytes(b"")
+        (labels_dir / f"image-{index}.txt").write_text(
+            "0 0.5 0.5 0.2 0.2\n",
+            encoding="utf-8",
+        )
+    (tmp_path / "classes.txt").write_text("item\n", encoding="utf-8")
+    user_split = tmp_path / "split"
+    user_split.mkdir()
+    (user_split / "train.txt").write_text("user data\n", encoding="utf-8")
+
+    first_yaml = Path(split_dataset(str(tmp_path), 0.5, "")[2])
+    second_yaml = Path(split_dataset(str(tmp_path), 0.5, "")[2])
+
+    assert (user_split / "train.txt").read_text(encoding="utf-8") == "user data\n"
+    assert first_yaml.parent != second_yaml.parent
+    assert user_split not in first_yaml.parents

tests/test_subprocess_runner.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from yolo_tui import subprocess_runner
+
+
+class FakeRunner:
+    def __init__(self, *, error: Exception | None = None) -> None:
+        self.error = error
+        self.prepared = False
+        self.stop_requested = False
+
+    def prepare_run(self) -> None:
+        self.prepared = True
+
+    def request_stop(self) -> None:
+        self.stop_requested = True
+
+    def train(self, config: Any, on_event: Any) -> Path:
+        if self.error is not None:
+            raise self.error
+        return Path("/tmp/successful-run")
+
+
+def _write_config(tmp_path: Path) -> Path:
+    config_path = tmp_path / "config.json"
+    config_path.write_text(
+        json.dumps({"dataset": "dataset.yaml", "model": "model.pt"}),
+        encoding="utf-8",
+    )
+    return config_path
+
+
+def test_main_returns_zero_after_successful_training(
+    monkeypatch: Any, tmp_path: Path, capsys: Any
+) -> None:
+    runner = FakeRunner()
+    monkeypatch.setattr(subprocess_runner, "TrainingRunner", lambda: runner)
+
+    return_code = subprocess_runner.main([str(_write_config(tmp_path))])
+
+    output = capsys.readouterr()
+    assert return_code == 0
+    assert runner.prepared is True
+    assert "__YOLO_TUI_READY__" in output.out
+    assert "__YOLO_TUI_RESULT__:/tmp/successful-run" in output.out
+    assert "Traceback" not in output.err
+
+
+def test_main_returns_one_when_training_raises(
+    monkeypatch: Any, tmp_path: Path, capsys: Any
+) -> None:
+    runner = FakeRunner(error=RuntimeError("training failed"))
+    monkeypatch.setattr(subprocess_runner, "TrainingRunner", lambda: runner)
+
+    return_code = subprocess_runner.main([str(_write_config(tmp_path))])
+
+    output = capsys.readouterr()
+    assert return_code == 1
+    assert "RuntimeError: training failed" in output.err

tests/test_trainer.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import sys
+import signal
 from pathlib import Path
 from types import ModuleType, SimpleNamespace
 from typing import Any
@@ -9,6 +10,25 @@ from yolo_tui.config import MlflowConfig, TrainingConfig
 from yolo_tui.trainer import TrainingEvent, TrainingRunner
 
 
+class FakeProcess:
+    def __init__(self) -> None:
+        self.signals: list[int] = []
+        self.terminate_calls = 0
+        self.kill_calls = 0
+
+    def send_signal(self, signum: int) -> None:
+        self.signals.append(signum)
+
+    def terminate(self) -> None:
+        self.terminate_calls += 1
+
+    def kill(self) -> None:
+        self.kill_calls += 1
+
+    def poll(self) -> None:
+        return None
+
+
 def test_runner_wires_yolo_callbacks_and_returns_output(
     monkeypatch: Any, tmp_path: Path
 ) -> None:
@@ -66,3 +86,40 @@ def test_runner_wires_yolo_callbacks_and_returns_output(
     assert train_arguments[0]["data"] == "dataset.yaml"
     assert train_arguments[0]["verbose"] is False
     assert [event.kind for event in events] == ["info", "started", "epoch", "epoch", "success"]
+
+
+def test_prepare_run_clears_previous_stop_request() -> None:
+    runner = TrainingRunner()
+    runner.request_stop()
+    assert runner.stop_requested is True
+
+    runner.prepare_run()
+
+    assert runner.stop_requested is False
+
+
+def test_early_stop_is_delivered_after_subprocess_ready() -> None:
+    runner = TrainingRunner()
+    process = FakeProcess()
+    runner.request_stop()
+
+    runner.set_subprocess(process, ready=False)
+    assert process.signals == []
+
+    runner.mark_subprocess_ready()
+
+    assert process.signals == [signal.SIGTERM]
+    assert process.terminate_calls == 0
+    runner.clear_subprocess()
+
+
+def test_ready_subprocess_receives_cooperative_signal_not_terminate() -> None:
+    runner = TrainingRunner()
+    process = FakeProcess()
+    runner.set_subprocess(process, ready=True)
+
+    runner.request_stop()
+
+    assert process.signals == [signal.SIGTERM]
+    assert process.terminate_calls == 0
+    runner.clear_subprocess()