一、背景:之前一篇博文中已经实现了maskrcnn_tf1.15.0环境的win10cpu模型训练但cpu训练实在是非常的耗时据说tf1.x是支持RTX1060的本人未测试但不支持最新的RTX3090查阅了很多资料原因应该是tf1.x与tf2.x的差别比较大必须升级到tf2.x才可以正常使用rtx3090。下面是maskrcnn_tf1.15.0的开发案例本人亲测可用。mask_rcnn 训练自定义数据集本地win10系统cpu已调通采坑无数均已列出解决方法_君临天下tjm的博客-CSDN博客源代码地址https://github.com/junlintianxiatjm/Mask_RCNN-master0071.版本信息python 3.6.9Tensorflow 1.15.0keras 2.2.5Pillow 5.3.0必须否则labelme执行json_to_dataset会出错Cv2必须安装训练模型时用到Wraptopt_einsumGastscikit-imageIPython虚拟环境完整配置库如下..https://blog.csdn.net/shanxiderenheni/article/details/118832615二、maskrcnn_tf1升级到maskrcnn_tf2的解决方法1.安装必要的包用镜像安装速度更快格式如pip install -i https://pypi.douban.com/simple/ tensorflow主要的pip listtensorflow 2.6.0keras 2.6.0matplotlib 3.2.2h5py 3.1.0numpy 1.19.5scikit-image 0.16.2tensorflow-gpu 2.6.2opencv-python 4.5.4.60详细pip list如下(py36_maskrcnn_env_bak) C:\Users\dellpip list Package Version ------------------------ ------------------- absl-py 0.15.0 astor 0.8.1 astunparse 1.6.3 backcall 0.2.0 bleach 1.5.0 cached-property 1.5.2 cachetools 4.2.4 certifi 2020.6.20 charset-normalizer 2.0.9 clang 5.0 colorama 0.4.4 cycler 0.11.0 Cython 0.29.28 dataclasses 0.8 decorator 4.4.2 flatbuffers 1.12 gast 0.4.0 google-auth 1.35.0 google-auth-oauthlib 0.4.6 google-pasta 0.2.0 grpcio 1.44.0 h5py 3.1.0 html5lib 0.9999999 idna 3.3 imageio 2.13.5 imgaug 0.4.0 imgviz 1.4.1 importlib-metadata 4.8.3 ipython 7.16.2 ipython-genutils 0.2.0 jedi 0.17.2 keras 2.6.0 Keras-Applications 1.0.8 Keras-Preprocessing 1.1.2 kiwisolver 1.3.1 labelme 3.16.2 libclang 11.1.0 Markdown 3.3.6 matplotlib 3.2.2 mock 4.0.3 networkx 2.5.1 numpy 1.19.5 nvidia-pyindex 1.0.9 oauthlib 3.1.1 opencv-python 4.5.4.60 opt-einsum 3.3.0 packaging 21.3 parso 0.7.1 pickleshare 0.7.5 Pillow 8.3.2 pip 21.2.2 prompt-toolkit 3.0.24 protobuf 3.17.3 pyasn1 0.4.8 pyasn1-modules 0.2.8 Pygments 2.10.0 pyparsing 3.0.6 PyQt5 5.15.6 PyQt5-Qt5 5.15.2 PyQt5-sip 12.9.0 python-dateutil 2.8.2 PyWavelets 1.1.1 PyYAML 6.0 QtPy 1.9.0 requests 2.26.0 requests-oauthlib 1.3.0 rsa 4.8 scikit-image 0.16.2 scipy 1.4.1 setuptools 58.0.4 Shapely 1.8.1.post1 six 1.15.0 tb-nightly 2.6.0a20210806 tensorboard 2.6.0 tensorboard-data-server 0.6.1 tensorboard-plugin-wit 1.8.1 tensorflow 2.6.0 tensorflow-estimator 2.6.0 tensorflow-gpu 2.6.2 tensorflow-gpu-estimator 2.2.0 termcolor 1.1.0 tf-estimator-nightly 2.7.0.dev2021092408 tifffile 2020.9.3 traitlets 4.3.3 typing-extensions 3.7.4.3 urllib3 1.26.7 wcwidth 0.2.5 Werkzeug 2.0.2 wheel 0.37.0 wincertstore 0.2 wrapt 1.12.1 zipp 3.6.02.用新mrcnn替换原mrcnn路径下的全部代码文件mrcnn 替换掉官方的mrcnn文件夹即可开箱使用新mrcnn代码路径如下https://github.com/junlintianxiatjm/MaskRCNN_TF2https://github.com/junlintianxiatjm/MaskRCNN_TF2注意调小Batch size即调小IMAGES_PER_GPU,本项目测试是8太大会提示内存耗尽。错误信息如下2022-03-11 11:02:03.235518: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8200 2022-03-11 11:02:09.947906: W tensorflow/core/common_runtime/bfc_allocator.cc:272] Allocator (GPU_0_bfc) ran out of memory trying to allocate 2.40GiB with freed_by_count0. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.设置 train.py里面的 “IMAGES_PER_GPU 8”或者更新值即可解决。gpu rtx3090训练效果如下