TX1上faster-Rcnn編譯測試與問題解決

部分借鑑http://blog.csdn.net/u011635764/article/details/52831167

根據之前發的文章，之前在TX1上已經安裝好caffe了 

基本環境是，Ubuntu16.04系統,cuda8.0, cudnn, opencv3.1.0等caffe必需的依賴

首先安裝pip用於安裝之後的Python依賴

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install python-pip

安裝Python依賴包

sudo pip install Cython
sudo pip install easydict

從github上下載程序源碼

git clone --recursive https://github.com/rbgirshick/py-faster-rcnn.git

修改並編譯Cython，打開py-faster-rcnn/lib找到setup.py

根據自己的顯卡計算能力修改第135行,查詢網址爲:https://developer.nvidia.com/cuda-gpus

TX1計算能力爲5.3，則修改爲

'nvcc': ['-arch=sm_53',

之後執行make進行編譯

之後打開caffe-faster-rcnn將安裝好的的caffe目錄下的Makefile.config和Makefile.config.example複製到該目錄下

然後對Makefile.config文件進行修改

USE_CUDNN := 1
OPENCV_VERSION := 3
CUSTOM_CXX := g++
WITH_PYTHON_LAYER := 1
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/lib/x86_64-Linux-gnu/hdf5/serial /usr/include/hdf5/serial 
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu /usr/lib/x86_64-linux-gnu/hdf5/serial
USE_PKG_CONFIG := 1

以上是在網上查閱各種資料彙總而得，可能有多餘的修改，但是能順利跑通

因爲作者的caffe版本不支持新的cudnn加速,我們需要將

py-faster-rcnn/caffe-fast-rcnn/include/caffe/util/cudnn.hpp  
py-faster-rcnn/caffe-fast-rcnn/src/caffe/util/cudnn.cpp
py-faster-rcnn/caffe-fast-rcnn/src/caffe/layers/ 目錄下16個cudnn_開頭的文件
py-faster-rcnn/caffe-fast-rcnn/include/caffe/layers/目錄下8個cudnn_開頭的文件

用已安裝好的caffe相應目錄下的對應文件進行替換

修改一個include錯誤，打開

py-faster-rcnn/caffe-fast-rcnn/src/caffe/test/test_smooth_L1_loss_layer.cpp

刪除第十一行

#include "caffe/vision_layers.hpp"

下面終於開始caffe編譯了，此處出現一個大坑

如果直接通過網上很多教程說的方法直接用make編譯，會報錯如下

AR -o .build_release/lib/libcaffe.a
LD -o .build_release/lib/libcaffe.so.1.0.0-rc3
/usr/bin/ld: cannot find -lhdf5_hl
/usr/bin/ld: cannot find -lhdf5
collect2: error: ld returned 1 exit status
Makefile:554: recipe for target '.build_release/lib/libcaffe.so.1.0.0-rc3' failed
make: *** [.build_release/lib/libcaffe.so.1.0.0-rc3] Error 1

此處必須通過cmake進行編譯

在caffe-fast-rcnn路徑下，在終端執行

make clean
mkdir build
cd build
cmake ..
make all -j4

就能通過編譯，原因貌似是make只能在當前目錄下尋找動態鏈接庫，具體原因還希望有大神進行解答
之後在caffe-fast-rcnn路徑下編譯

sudo make -j8 test

立刻完成

之後在caffe-fast-rcnn路徑下執行

make runtest 

發現問題

F0511 14:58:00.916214  7960 db_lmdb.hpp:14] Check failed: mdb_status == 0 (12 vs. 0) Cannot allocate memory
*** Check failure stack trace: ***
    @       0x7f8401e718  google::LogMessage::Fail()
    @       0x7f84020614  google::LogMessage::SendToLog()
    @       0x7f8401e290  google::LogMessage::Flush()
    @       0x7f84020eb4  google::LogMessageFatal::~LogMessageFatal()
    @       0x7f843b37ec  caffe::db::LMDB::Open()
    @           0x93b264  caffe::DBTest<>::SetUp()
    @           0xcb8f04  testing::internal::HandleExceptionsInMethodIfSupported<>()
    @           0xcb1dd8  testing::Test::Run()
    @           0xcb1f98  testing::TestInfo::Run()
    @           0xcb20a4  testing::TestCase::Run()
    @           0xcb3b08  testing::internal::UnitTestImpl::RunAllTests()
    @           0xcb3e1c  testing::UnitTest::Run()
    @           0x84f570  main
    @       0x7f7fc008a0  __libc_start_main
Aborted
src/caffe/test/CMakeFiles/runtest.dir/build.make:57: recipe for target 'src/caffe/test/CMakeFiles/runtest' failed
make[3]: *** [src/caffe/test/CMakeFiles/runtest] Error 134
CMakeFiles/Makefile2:328: recipe for target 'src/caffe/test/CMakeFiles/runtest.dir/all' failed
make[2]: *** [src/caffe/test/CMakeFiles/runtest.dir/all] Error 2
CMakeFiles/Makefile2:335: recipe for target 'src/caffe/test/CMakeFiles/runtest.dir/rule' failed
make[1]: *** [src/caffe/test/CMakeFiles/runtest.dir/rule] Error 2
Makefile:240: recipe for target 'runtest' failed
make: *** [runtest] Error 2

此處是因爲分配內存不夠，解決方法爲：

cd examples/mnist

將convert_mnist_data.cpp中

CHECK_EQ(mdb_env_set_mapsize(mdb_env, 1099511627776), MDB_SUCCESS) // 1TB 

修改爲

CHECK_EQ(mdb_env_set_mapsize(mdb_env, 1073741824), MDB_SUCCESS) // 1GB

同樣的，在/src/caffe/util/db_lmdb.cpp中
將LMDB_MAP_SIZE修改爲4294967296
之後在caffe-fast-rcnn路徑下執行

sudo make pycaffe

就可以完成編譯

之後下載已經訓練好的模型，網上大多數方法爲

./data/scripts/fetch_faster_rcnn_models.sh

我發現網站無響應，大家可以直接用瀏覽器打開https://dl.dropboxusercontent.com/s/o6ii098bu51d139/faster_rcnn_models.tgz進行下載

下載完成後，解壓文件夾至py-faster-rcnn\data目錄下

最後，執行demo程序

python /tools/demo.py

發現報錯

ImportError: No module named skimage.io

解決方法爲，執行

sudo apt-get install python-skimage

再次執行demo又出現問題

ImportError: No module named google.protobuf.internal

解決方法爲，執行

sudo pip install protobuf

呼，這樣就可以執行demo了，希望這篇文章能幫助到大家

但是，執行demo時又出現了問題，由於這是在嵌入式環境下進行，默認載入的模型是VGG16網絡，VGG16網絡比較大，可能會出現內存不足，導致killed

Loaded network /mnt/sd1/faster_rcnn_models/VGG16_faster_rcnn_final.caffemodel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000456.jpg
Detection took 1.879s for 300 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000542.jpg
Killed

解決方法是打開tools/demo.py文件將

parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16]',
                        choices=NETS.keys(), default='vgg16')

修改爲

parser.add_argument('--net', dest='demo_net', help='Network to use [zf]',
                        choices=NETS.keys(), default='zf')

即可跑通demo

Loaded network /mnt/sd1/faster_rcnn_models/ZF_faster_rcnn_final.caffemodel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000456.jpg
Detection took 0.510s for 300 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000542.jpg
Detection took 0.552s for 135 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/001150.jpg
Detection took 1.014s for 231 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/001763.jpg
Detection took 0.714s for 200 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/004545.jpg
Detection took 0.577s for 300 object proposals

上文可以看到，faster-Rcnn在TX1運行比較快，但跟PC端還是有相當大差距的，得到的結果是