运行Faster RCNN总结

运行Faster RCNN总结

在科研期间，做目标检测的任务，运行基于TensorFlow的Faster RCNN，进行了总结：

环境

1. tensorflow 1.6.0
2. gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6 ubuntu1~16.04.9)

运行期间遇到的错误

1. ./lib/roi_pooling_layer/roi_pooling.so: undefined symbol: _ZTIN10tensorflow8OpKernelE
   make.sh添加TF_LIB=$(python -c ‘import tensorflow as tf; print(tf.sysconfig.get_lib())’)
   在g++配置后添加-D_GLIBCXX_USE_CXX11_ABI=0
2. Loaded network ./data/VGGnet_fast_rcnn_iter_70000.ckpt
   段错误：I also encountered ‘Segmentation Fault(core dumped)’ problem as @serhannn was, while i found i had made a stupid mistake: I run gpu mode with cpu version of tensorflow.
   After update tensorflow-gpu & run make one more time it worked well. With Ubuntu16.04 & CUDA8 & cudnn5 & tensorflow-gpu-1.3.0
3. 测试模型时，设置device_id为非0会报错，
   Loading model weights from ./data/VGGnet_fast_rcnn_iter_70000.ckpt cudaCheckError() failed : invalid resource handle