Files
ant-vision-inspector/ai/trainer.py
2026-06-10 16:18:41 +09:00

270 lines
9.1 KiB
Python

# AI 학습 — YOLOv8 재학습 및 모델 저장
import multiprocessing
import os
import random
import shutil
from utils.path_helper import get_path
import yaml
from PyQt5.QtCore import QThread, pyqtSignal
class Trainer:
def __init__(self):
self.model = None
self.is_training = False
# ------------------------------------------------------------------ #
def prepare_dataset(self, image_folder: str) -> str:
dataset_dir = get_path("ai", "dataset")
if os.path.exists(dataset_dir):
shutil.rmtree(dataset_dir)
for split in ("train", "val"):
os.makedirs(os.path.join(dataset_dir, "images", split), exist_ok=True)
os.makedirs(os.path.join(dataset_dir, "labels", split), exist_ok=True)
pairs = []
for f in os.listdir(image_folder):
if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
img_path = os.path.join(image_folder, f)
txt_path = os.path.join(
image_folder, os.path.splitext(f)[0] + ".txt"
)
if os.path.exists(txt_path):
pairs.append((img_path, txt_path))
random.shuffle(pairs)
split_idx = int(len(pairs) * 0.8)
train_pairs = pairs[:split_idx]
val_pairs = pairs[split_idx:]
for img, lbl in train_pairs:
shutil.copy(img, os.path.join(dataset_dir, "images", "train"))
shutil.copy(lbl, os.path.join(dataset_dir, "labels", "train"))
for img, lbl in val_pairs:
shutil.copy(img, os.path.join(dataset_dir, "images", "val"))
shutil.copy(lbl, os.path.join(dataset_dir, "labels", "val"))
yaml_content = {
"path": dataset_dir,
"train": "images/train",
"val": "images/val",
"nc": 4,
"names": ["스크래치", "이물", "흑점", "변형"],
}
yaml_path = os.path.join(dataset_dir, "data.yaml")
with open(yaml_path, "w", encoding="utf-8") as fh:
yaml.dump(yaml_content, fh, allow_unicode=True)
return yaml_path
# ------------------------------------------------------------------ #
def train(
self,
image_folder: str,
epochs: int,
batch: int,
save_path: str,
log_callback=None,
progress_callback=None,
):
self.is_training = True
try:
if log_callback:
log_callback("데이터셋 준비 중...")
yaml_path = self.prepare_dataset(image_folder)
if log_callback:
log_callback(f"데이터셋 준비 완료: {yaml_path}")
if log_callback:
log_callback("YOLOv8 모델 로드 중...")
from ultralytics import YOLO # 지연 로딩 — 앱 시작 시 torch DLL 오류 방지
self.model = YOLO("yolov8n.pt")
if log_callback:
log_callback(f"학습 시작 (epoch={epochs}, batch={batch})")
def _on_epoch_end(trainer):
ep = trainer.epoch + 1
try:
loss_val = float(trainer.loss)
loss_str = f"{loss_val:.4f}"
except Exception:
loss_str = "?"
if log_callback:
log_callback(f"Epoch {ep}/{epochs} loss={loss_str}")
if progress_callback:
progress_callback(int(ep / epochs * 100))
self.model.add_callback("on_train_epoch_end", _on_epoch_end)
self.model.train(
data=yaml_path,
epochs=epochs,
batch=batch,
imgsz=640,
project=get_path("ai", "runs"),
name="train",
exist_ok=True,
verbose=True,
workers=0, # disable DataLoader multiprocessing inside subprocess
amp=False, # AMP check also spawns a subprocess on Windows
plots=False, # matplotlib can interfere with Qt event loop
)
# best.pt 복사
best_pt = get_path("ai", "runs", "train", "weights", "best.pt")
if os.path.exists(best_pt):
os.makedirs(os.path.dirname(save_path), exist_ok=True)
shutil.copy(best_pt, save_path)
if log_callback:
log_callback(f"모델 저장 완료: {save_path}")
if progress_callback:
progress_callback(100)
if log_callback:
log_callback("학습 완료!")
except BaseException as e:
import traceback
if log_callback:
try:
log_callback(f"학습 오류: {e}")
log_callback(traceback.format_exc())
except Exception:
pass
finally:
self.is_training = False
# ------------------------------------------------------------------ #
def stop(self):
self.is_training = False
# ====================================================================== #
# Subprocess entry point — defined at module level so it is picklable.
# ultralytics training can call sys.exit() internally; running it in a
# separate process completely isolates the Qt application from that.
# ====================================================================== #
def _train_subprocess_main(queue, image_folder, epochs, batch, save_path):
"""Entry point for the isolated training subprocess."""
try:
trainer = Trainer()
def _log(msg):
try:
queue.put(("log", msg))
except Exception:
pass
def _progress(pct):
try:
queue.put(("progress", int(pct)))
except Exception:
pass
trainer.train(
image_folder=image_folder,
epochs=epochs,
batch=batch,
save_path=save_path,
log_callback=_log,
progress_callback=_progress,
)
queue.put(("done", True))
except BaseException as e:
import traceback
try:
queue.put(("log", f"학습 오류: {e}"))
queue.put(("log", traceback.format_exc()))
except Exception:
pass
try:
queue.put(("done", False))
except Exception:
pass
# ====================================================================== #
class TrainWorker(QThread):
log_signal = pyqtSignal(str)
progress_signal = pyqtSignal(int)
finished_signal = pyqtSignal(bool)
def __init__(self, trainer, image_folder, epochs, batch, save_path):
super().__init__()
self.trainer = trainer
self.image_folder = image_folder
self.epochs = epochs
self.batch = batch
self.save_path = save_path
self._proc = None # training subprocess handle
def run(self):
# Spawn an isolated subprocess so that any sys.exit() call inside
# ultralytics does not reach PyQt5's QThread handler and trigger
# QApplication.exit() in the main process.
ctx = multiprocessing.get_context("spawn")
q = ctx.Queue()
self._proc = ctx.Process(
target=_train_subprocess_main,
args=(q, self.image_folder, self.epochs, self.batch, self.save_path),
daemon=True,
)
self._proc.start()
success = False
while True:
proc_alive = self._proc.is_alive()
try:
msg_type, msg_data = q.get(timeout=0.3)
except Exception:
# queue.Empty — check if subprocess died unexpectedly
if not proc_alive:
# Give one last chance to read remaining messages
while True:
try:
msg_type, msg_data = q.get_nowait()
except Exception:
break
if msg_type == "log":
self.log_signal.emit(str(msg_data))
elif msg_type == "progress":
self.progress_signal.emit(int(msg_data))
elif msg_type == "done":
success = bool(msg_data)
break
continue
if msg_type == "log":
self.log_signal.emit(str(msg_data))
elif msg_type == "progress":
self.progress_signal.emit(int(msg_data))
elif msg_type == "done":
success = bool(msg_data)
break
self._proc.join(timeout=30)
if self._proc.is_alive():
self._proc.terminate()
self._proc.join(timeout=5)
self.finished_signal.emit(success)
def stop_subprocess(self):
"""Call from the main thread to forcefully stop training."""
if self._proc and self._proc.is_alive():
self._proc.terminate()