Harden setup_server.sh against numpy/opencv/imgaug conflicts

- Install libgl1 and other OpenCV runtime deps via apt
- Skip paddle/blinker/opencv lines from PaddleOCR requirements (they conflict with the sm_120 wheel or system packages)
- Force opencv>=4.10 (numpy 2 ABI) then pin numpy<2 (imgaug uses removed np.sctypes)
- Smoke-test PGNet module imports before declaring success

Verified end-to-end on RTX 5090 + paddle 3.3.0.dev20251209.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

scripts/setup_server.sh

@@ -1,13 +1,12 @@
 #!/usr/bin/env bash
 # kr_lp_pgnet 전용 컨테이너 안에서 실행되는 환경 셋업 스크립트.
 #
-# 가정 (컨테이너 외부):
+# 컨테이너 가정:
-#   - Ubuntu 24.04 베이스 컨테이너 `kr_lp_pgnet` (--gpus all, --restart unless-stopped)
+#   - Ubuntu 24.04 + Python 3.10 + paddlepaddle-gpu sm_120 wheel 설치됨
-#   - 호스트 /home/cuuva/workspace ↔ 컨테이너 /workspace (bind mount)
+#   - 호스트 /home/cuuva/workspace ↔ 컨테이너 /workspace bind mount
-#   - 이 repo는 호스트 ~/workspace/kr_lp_pgnet/ 에 clone (= 컨테이너 /workspace/kr_lp_pgnet/)
+#   - 이 repo는 /workspace/kr_lp_pgnet/, paddle wheel은 /workspace/wheels/
-#   - paddle sm_120 wheel은 ~/workspace/wheels/ 안에 둠
 #
-# 실행 (호스트에서):
+# 실행 (호스트):
 #   docker exec kr_lp_pgnet bash /workspace/kr_lp_pgnet/scripts/setup_server.sh
 
 set -euo pipefail
@@ -20,43 +19,48 @@ PRETRAIN_DIR="$PADDLEOCR_DIR/pretrain_models"
 WHEEL_DIR="$WORKSPACE/wheels"
 
 PY=python3.10
+PIP="$PY -m pip"
+PIP_OPTS="--quiet --root-user-action=ignore"
 
-echo "[1/6] python / paddle 확인"
+echo "[1/7] python / paddle 확인"
 $PY --version
 if ! $PY -c 'import paddle; assert paddle.is_compiled_with_cuda()' 2>/dev/null; then
-    echo "  paddle 미설치 또는 CUDA 비호환. sm_120 wheel 재설치..."
+    echo "  paddle 미설치 또는 CUDA 비호환. sm_120 wheel 설치..."
     WHL=$(ls "$WHEEL_DIR"/paddlepaddle_gpu-*-cp310-*linux_x86_64.whl 2>/dev/null | head -1)
     if [ -z "$WHL" ]; then
-        echo "  wheel 파일을 $WHEEL_DIR/ 에 두고 다시 실행하세요." >&2
+        echo "  ERROR: $WHEEL_DIR/paddlepaddle_gpu-*-cp310-*.whl 없음" >&2
         exit 1
     fi
-    $PY -m pip install --quiet "$WHL"
+    $PIP install $PIP_OPTS "$WHL"
 fi
 $PY -c 'import paddle; print("  paddle:", paddle.__version__, "cuda:", paddle.is_compiled_with_cuda())'
 
-echo "[2/6] PaddleOCR clone (release/2.7)"
+echo "[2/7] OpenCV 시스템 의존성 (libgl 등)"
+export DEBIAN_FRONTEND=noninteractive
+apt-get install -y -q libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 wget git
+
+echo "[3/7] PaddleOCR clone (release/2.7)"
 if [ ! -d "$PADDLEOCR_DIR" ]; then
     git clone --depth 1 -b release/2.7 https://github.com/PaddlePaddle/PaddleOCR.git "$PADDLEOCR_DIR"
 fi
 cd "$PADDLEOCR_DIR"
 echo "  PaddleOCR @$(git rev-parse --short HEAD)"
 
-echo "[3/6] PaddleOCR requirements 설치 (paddlepaddle 라인은 제외)"
+echo "[4/7] PaddleOCR Python 의존성 (한 줄씩, 충돌 패키지는 skip)"
-if [ -f requirements.txt ]; then
+# paddle은 이미 wheel로, blinker는 system pkg, opencv는 numpy2 비호환 → 제외 후 별도 설치
-    grep -viE '^(paddlepaddle|paddleocr)' requirements.txt > /tmp/kr_lp_req.txt || true
+grep -viE '^(paddlepaddle|paddleocr|blinker|opencv-)' requirements.txt > /tmp/kr_lp_req.txt || true
-    $PY -m pip install --quiet -r /tmp/kr_lp_req.txt || true
+while IFS= read -r line; do
-fi
+    [[ -z "$line" || "$line" =~ ^# ]] && continue
+    $PIP install $PIP_OPTS --ignore-installed "$line" 2>/dev/null || echo "  skip: $line"
+done < /tmp/kr_lp_req.txt
 
-echo "[4/6] PaddleOCR import smoke test"
+echo "[5/7] OpenCV (numpy2 호환) + numpy<2 (PaddleOCR release/2.7 호환성)"
-cd "$PADDLEOCR_DIR"
+$PIP install $PIP_OPTS 'opencv-python>=4.10' 'opencv-contrib-python>=4.10'
-$PY -c "
+# imgaug 등이 numpy 1.x API(np.sctypes)에 의존하므로 numpy 1.x로 핀.
-import sys; sys.path.insert(0, '.')
+# paddle 3.3.0.dev는 numpy 1.x도 호환.
-import ppocr
+$PIP install $PIP_OPTS 'numpy<2' --force-reinstall
-from ppocr.modeling.architectures import build_model
-print('  ppocr import OK')
-" || { echo "  PaddleOCR import 실패 (paddle 3.3 호환성 이슈일 수 있음)"; exit 1; }
 
-echo "[5/6] PGNet step1 pretrain weight 다운로드"
+echo "[6/7] PGNet step1 pretrain weight 다운로드 + import smoke test"
 mkdir -p "$PRETRAIN_DIR"
 cd "$PRETRAIN_DIR"
 if [ ! -d train_step1 ]; then
@@ -64,8 +68,18 @@ if [ ! -d train_step1 ]; then
     tar xf train_step1.tar && rm train_step1.tar
 fi
 ls train_step1/
+cd "$PADDLEOCR_DIR"
+$PY -c "
+import sys; sys.path.insert(0, '.')
+from ppocr.modeling.architectures import build_model
+from ppocr.postprocess.pg_postprocess import PGPostProcess
+from ppocr.losses.e2e_pg_loss import PGLoss
+from ppocr.modeling.heads.e2e_pg_head import PGHead
+from ppocr.data.imaug.pg_process import PGProcessTrain
+print('  PGNet modules import OK')
+"
 
-echo "[6/6] dict / config symlink → PaddleOCR 트리"
+echo "[7/7] dict / config symlink → PaddleOCR 트리"
 ln -sf "$KR_LP_DIR/dict/kr_lp_dict.txt" "$PADDLEOCR_DIR/ppocr/utils/kr_lp_dict.txt"
 mkdir -p "$PADDLEOCR_DIR/configs/e2e"
 ln -sf "$KR_LP_DIR/configs/kr_lp_pgnet.yml" "$PADDLEOCR_DIR/configs/e2e/kr_lp_pgnet.yml"
@@ -74,6 +88,6 @@ ls -l "$PADDLEOCR_DIR/ppocr/utils/kr_lp_dict.txt" "$PADDLEOCR_DIR/configs/e2e/kr
 echo
 echo "==========================="
 echo "셋업 완료. 다음 단계:"
-echo "  bash $KR_LP_DIR/data_gen/setup_assets.sh        # 합성 자산 다운로드"
+echo "  bash $KR_LP_DIR/data_gen/setup_assets.sh           # 합성 자산 다운로드"
-echo "  python data_gen/generate_synthetic.py ...       # 합성 데이터 생성 (다음 단계)"
+echo "  python3.10 $KR_LP_DIR/data_gen/generate_synthetic.py ...  # 합성 데이터 생성"
 echo "==========================="