Detectron2训练模型导出ONNX/TorchScript失败求助

问题场景 #

我修改了Detectron2的export_model.py脚本适配Notebook，试图将训练好的Mask R-CNN模型导出为ONNX或TorchScript格式。模型配置与加载代码如下：

def setup_cfg():
    cfg = get_cfg()
    cfg_file = model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")
    cfg.merge_from_file(cfg_file)
    cfg.MODEL.ROI_HEADS.NUM_CLASSES = 2
    cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")
    return cfg

cfg = setup_cfg()
model_ = build_model(cfg) 
DetectionCheckpointer(model_).resume_or_load("./model/model_final.pth", resume=False)
model_.eval()

错误现象 #

调用export_tracing(model_, inputs, "./", "onnx")导出ONNX时，触发RuntimeError: object has no attribute nms，同时伴随TracerWarning；调用export_scripting(model_, "torchscript", "./")导出TorchScript时，出现完全相同的nms相关错误。

工作环境信息 #

sys.platform                     linux
Python                           3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0]
numpy                            1.24.1
detectron2                       0.6 @/home/xxx/.local/lib/python3.10/site-packages/detectron2
Compiler                         GCC 11.4
CUDA compiler                    not available
DETECTRON2_ENV_MODULE            <not set>
PyTorch                          2.1.0+cu121 @/home/xxx/.local/lib/python3.10/site-packages/torch
PyTorch debug build              False
torch._C._GLIBCXX_USE_CXX11_ABI  False
GPU available                    Yes
GPU 0                            NVIDIA GeForce RTX 3050 Laptop GPU (arch=8.6)
Driver version                   535.113.01
CUDA_HOME                        None - invalid!
Pillow                           9.0.1
torchvision                      0.16.0+cu118 @/home/xxx/.local/lib/python3.10/site-packages/torchvision
torchvision arch flags           /home/xxx/.local/lib/python3.10/site-packages/torchvision/_C.so
fvcore                           0.1.5.post20221221
iopath                           0.1.9
cv2                              4.8.1
-------------------------------  ---------------------------------------------------------------------------
PyTorch built with:
  - GCC 9.3
  - C++ Version: 201703
  - Intel(R) oneAPI Math Kernel Library Version 2022.2-Product Build 20220804 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v3.1.1 
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 12.1
  - NVCC architecture flags: -gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_90,code=sm_90
  - CuDNN 8.9.2
  - Magma 2.6.1
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=12.1, CUDNN_VERSION=8.9.2, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 -fabi-version=11 -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOROCTRACER -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=old-style-cast -Wno-invalid-partial-specialization -Wno-unused-private-field -Wno-aligned-allocation-unavailable -Wno-missing-braces -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_DISABLE_GPU_ASSERTS=ON, TORCH_VERSION=2.1.0, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=1, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF, 

排查与解决方案 #

问题原因 #

1. Detectron2默认模型的NMS操作采用了动态属性调用或自定义实现，PyTorch的追踪/脚本化机制无法正确识别这类动态调用，导致导出时找不到nms属性。
2. Detectron2 0.6版本与PyTorch 2.1存在兼容性差异，旧版本的导出逻辑未适配新版本PyTorch的特性。

正确导出方法 #

方法1：使用Detectron2官方原生导出流程

无需手动修改export_model.py，直接调用官方提供的导出函数，适配Notebook的代码如下：

import torch
from detectron2.export import export_torchscript_with_onnx

# 准备匹配模型输入尺寸的示例张量
inputs = [{"image": torch.randn(3, 800, 1200).to(model_.device)}]

# 同时导出ONNX与TorchScript
export_torchscript_with_onnx(cfg, model_, inputs, "./output_model.onnx")

# 单独导出TorchScript
scripted_model = torch.jit.script(model_)
scripted_model.save("./output_model.pt")

注意：输入尺寸需与训练时的配置保持一致，避免因尺寸不匹配触发错误。

方法2：替换NMS为PyTorch原生实现

将Detectron2自定义的NMS替换为PyTorch官方实现，确保导出时能被追踪器识别：

from torchvision.ops import nms

# 定义原生NMS包装函数
def custom_nms(boxes, scores, iou_threshold):
    return nms(boxes, scores, iou_threshold)

# 替换模型ROI Heads中的NMS调用
model_.roi_head.box_predictor.test_nms = custom_nms

修改完成后再执行导出操作即可。

方法3：升级Detectron2版本

Detectron2 0.6版本较旧，建议升级到适配PyTorch 2.1的最新稳定版：

pip install --upgrade detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu121/torch2.1/index.html

根据自身CUDA版本选择对应wheel链接，升级后官方修复的导出相关问题可直接解决当前报错。

内容的提问来源于stack exchange，提问作者rezahs