bev-project/scripts/check_env_docker.sh

#!/bin/bash
# BEVFusion Docker环境检查

export PATH=/opt/conda/bin:$PATH

echo "=========================================="
echo "BEVFusion Docker环境检查"
echo "=========================================="
echo ""

PASS=0
FAIL=0

check_pass() {
    echo "  ✓ $1"
    ((PASS++))
}

check_fail() {
    echo "  ✗ $1"
    ((FAIL++))
}

# 1. Python和PyTorch
echo "[1/10] Python和PyTorch"
PYTHON_VERSION=$(python --version 2>&1)
check_pass "$PYTHON_VERSION"

PYTORCH_INFO=$(python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA: {torch.cuda.is_available()}, GPUs: {torch.cuda.device_count()}')" 2>&1)
check_pass "$PYTORCH_INFO"

# 2. torchpack
echo ""
echo "[2/10] torchpack"
python -c "import torchpack; print('torchpack installed')" 2>/dev/null
if [ $? -eq 0 ]; then
    check_pass "torchpack 已安装"
else
    check_fail "torchpack 未安装"
fi

# 3. mmcv
echo ""
echo "[3/10] mmcv"
MMCV_VERSION=$(python -c "import mmcv; print(mmcv.__version__)" 2>/dev/null)
if [ $? -eq 0 ]; then
    check_pass "mmcv $MMCV_VERSION"
else
    check_fail "mmcv 未安装"
fi

# 4. mmdet
echo ""
echo "[4/10] mmdetection"
MMDET_VERSION=$(python -c "import mmdet; print(mmdet.__version__)" 2>/dev/null)
if [ $? -eq 0 ]; then
    check_pass "mmdetection $MMDET_VERSION"
else
    check_fail "mmdetection 未安装"
fi

# 5. nuscenes-devkit
echo ""
echo "[5/10] nuscenes-devkit"
python -c "from nuscenes import NuScenes" 2>/dev/null
if [ $? -eq 0 ]; then
    check_pass "nuscenes-devkit 已安装"
else
    check_fail "nuscenes-devkit 未安装"
fi

# 6. 其他依赖
echo ""
echo "[6/10] 其他Python依赖"
python -c "import tqdm" 2>/dev/null && check_pass "tqdm" || check_fail "tqdm"
python -c "from PIL import Image" 2>/dev/null && check_pass "Pillow" || check_fail "Pillow"
python -c "import numpy" 2>/dev/null && check_pass "numpy" || check_fail "numpy"

# 7. 自定义算子
echo ""
echo "[7/10] 自定义CUDA算子"
python -c "from mmdet3d.ops import bev_pool_v2" 2>/dev/null && check_pass "BEV Pool" || check_fail "BEV Pool (需要: python setup.py develop)"
python -c "from mmdet3d.ops import Voxelization" 2>/dev/null && check_pass "Voxelization" || check_fail "Voxelization"
python -c "from mmdet3d.ops.spconv import SparseConv3d" 2>/dev/null && check_pass "Sparse Conv" || check_fail "Sparse Conv"

# 8. GPU信息
echo ""
echo "[8/10] GPU信息"
nvidia-smi --query-gpu=index,name,memory.total,memory.free --format=csv,noheader | while read line; do
    check_pass "GPU $line"
done

# 9. 数据集
echo ""
echo "[9/10] nuScenes数据集"
if [ -d "data/nuscenes" ]; then
    check_pass "数据集目录存在"
    [ -f "data/nuscenes/nuscenes_infos_train.pkl" ] && check_pass "训练info文件" || check_fail "训练info文件缺失"
    [ -f "data/nuscenes/nuscenes_infos_val.pkl" ] && check_pass "验证info文件" || check_fail "验证info文件缺失"
else
    check_fail "数据集目录不存在"
fi

# 10. 预训练模型
echo ""
echo "[10/10] 预训练模型"
if [ -d "pretrained" ]; then
    check_pass "预训练目录存在"
    [ -f "pretrained/swint-nuimages-pretrained.pth" ] && check_pass "SwinT预训练模型" || check_fail "SwinT预训练模型缺失"
    [ -f "pretrained/lidar-only-det.pth" ] && check_pass "LiDAR检测模型" || check_fail "LiDAR检测模型缺失"
else
    check_fail "预训练目录不存在"
fi

# 总结
echo ""
echo "=========================================="
echo "检查完成: 通过 $PASS 项, 失败 $FAIL 项"
echo "=========================================="

if [ $FAIL -eq 0 ]; then
    echo ""
    echo "✅ 环境完整，可以开始训练！"
    echo ""
    echo "推荐训练命令："
    echo ""
    echo "1. 3D检测训练（8 GPU）："
    echo "   torchpack dist-run -np 8 python tools/train.py \\"
    echo "     configs/nuscenes/det/transfusion/secfpn/camera+lidar/swint_v0p075/convfuser.yaml \\"
    echo "     --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth \\"
    echo "     --load_from pretrained/lidar-only-det.pth"
    echo ""
    echo "2. BEV分割训练（8 GPU）："
    echo "   torchpack dist-run -np 8 python tools/train.py \\"
    echo "     configs/nuscenes/seg/fusion-bev256d2-lss.yaml \\"
    echo "     --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth"
    echo ""
    echo "3. 多任务训练（检测+分割，8 GPU）："
    echo "   torchpack dist-run -np 8 python tools/train.py \\"
    echo "     configs/nuscenes/multitask/fusion-det-seg-swint.yaml \\"
    echo "     --model.encoders.camera.backbone.init_cfg.checkpoint pretrained/swint-nuimages-pretrained.pth \\"
    echo "     --load_from pretrained/lidar-only-det.pth"
    exit 0
else
    echo ""
    echo "⚠️  发现 $FAIL 个问题需要解决"
    exit 1
fi