Tensorflow编译

MacOS平台编译

1. clone tensorflow repositories

git clone https://github.com/tensorflow/tensorflow.git tensorflow_src

2. download all build dependencies

./tensorflow/lite/tools/make/download_dependencies.sh

3. build tensorflow-lite

cmake ../tensorflow_src/tensorflow/lite
cmake --build .

libtensorflow-lite.a需要的额外库路径tensorflow_src/_deps/xxx-build/xxx.a

Ubuntu平台编译

1. install necessary library

sudo apt-get install build-essential
sudo apt-get install zlib1g-dev
sudo apt install libgles2-mesa-dev 

2. download all build dependencies

./tensorflow/lite/tools/make/download_dependencies.sh

3. build tensorflow-lite

./tensorflow/lite/tools/make/build_lib.sh

Ubuntu环境编译动态库脚本

#!/bin/sh
set -e
#set -x

export TENSORFLOW_VER=r2.4
export TENSORFLOW_DIR=`pwd`/tensorflow_${TENSORFLOW_VER}

git clone -b ${TENSORFLOW_VER} https://github.com/tensorflow/tensorflow.git ${TENSORFLOW_DIR}

cd ${TENSORFLOW_DIR}


# install Bazel 3.1.0
wget https://github.com/bazelbuild/bazel/releases/download/3.1.0/bazel-3.1.0-installer-linux-x86_64.sh
chmod 755 bazel-3.1.0-installer-linux-x86_64.sh
sudo ./bazel-3.1.0-installer-linux-x86_64.sh

# clean up bazel cache, just in case.
bazel clean

echo "----------------------------------------------------"
echo " (configure) press ENTER-KEY several times.         "
echo "----------------------------------------------------"
./configure

# ---------------
#  Makefile build
# ---------------

# download all the build dependencies.
./tensorflow/lite/tools/make/download_dependencies.sh 2>&1 | tee -a log_download_dependencies.txt

# build TensorFlow Lite library (libtensorflow-lite.a)
./tensorflow/lite/tools/make/build_lib.sh EXTRA_CXXFLAGS="-march=native" 2>&1 | tee -a log_build_libtflite_make.txt


# ---------------
#  Bazel build
# ---------------
# build with Bazel (libtensorflowlite.so)
bazel build -s -c opt //tensorflow/lite:libtensorflowlite.so 2>&1 | tee -a log_build_libtflite_bazel.txt

# build GPU Delegate library (libdelegate.so)
bazel build -s -c opt --copt="-DMESA_EGL_NO_X11_HEADERS" --copt="-DEGL_NO_X11" tensorflow/lite/delegates/gpu:libtensorflowlite_gpu_delegate.so 2>&1 | tee -a log_build_delegate.txt

echo "----------------------------------------------------"
echo " build success."
echo "----------------------------------------------------"

cd ${TENSORFLOW_DIR}
#ls -l tensorflow/lite/tools/make/gen/linux_x86_64/lib/
ls -l bazel-bin/tensorflow/lite/
ls -l bazel-bin/tensorflow/lite/delegates/gpu/

遇到问题

virtual memory exhausted: Cannot allocate memory
解决：内存太小，用swap扩展内存的方法

[root@Byrd byrd]# free -m
             total       used       free     shared    buffers     cached
Mem:           512        108        403          0          0         28
-/+ buffers/cache:         79        432
Swap:            0          0          0
[root@Byrd ~]# mkdir /opt/images/
[root@Byrd ~]# rm -rf /opt/images/swap
[root@Byrd ~]# dd if=/dev/zero of=/opt/images/swap bs=1024 count=2048000
2048000+0 records in
2048000+0 records out
2097152000 bytes (2.1 GB) copied, 82.7509 s, 25.3 MB/s
[root@Byrd ~]# mkswap /opt/images/swap
mkswap: /opt/images/swap: warning: don't erase bootbits sectors
        on whole disk. Use -f to force.
Setting up swapspace version 1, size = 2047996 KiB
no label, UUID=59daeabb-d0c5-46b6-bf52-465e6b05eb0b
[root@hz mnt]# swapon /opt/images/swap
[root@hz mnt]# free -m
             total       used       free     shared    buffers     cached
Mem:           488        481          7          0          6        417
-/+ buffers/cache:         57        431
Swap:          999          0        999

使用完毕后可以关掉swap：

[root@hz mnt]# swapoff swap
[root@hz mnt]# rm -f /opt/images/swap

参考链接

https://tensorflow.google.cn/lite/guide/build_cmake