Scripts for installing tensorflow-2.6.0 on JetPack-4.6

install_tensorflow-2.6.0.sh

s#!/bin/bash

set -e

# tensorflow version
version=2.6.0

if [[ ! $(head -1 /etc/nv_tegra_release) =~ R32.*6\.1 ]] ; then
  echo "ERROR: not JetPack-4.6"
  exit 1
fi

case $(cat /sys/module/tegra_fuse/parameters/tegra_chip_id) in
  "33" )  # Nano and TX1
    cuda_compute=5.3
    ;;
  "24" )  # TX2
    cuda_compute=6.2
    ;;
  "25" )  # Xavier NX and AGX Xavier
    cuda_compute=7.2
    ;;
  * )     # default
    cuda_compute=5.3,6.2,7.2
    ;;
esac

script_path=$(realpath $0)
patch_path=$(dirname $script_path)/tensorflow/tensorflow-${version}.patch
trt_version=$(echo /usr/lib/aarch64-linux-gnu/libnvinfer.so.? | cut -d '.' -f 3)

src_folder=${HOME}/src
mkdir -p $src_folder

if pip3 list | grep tensorflow > /dev/null; then
  echo "ERROR: tensorflow is installed already"
  exit 1
fi

if [[ -z $(pip3 list | grep numpy) ]]; then
  echo "ERROR: missing numpy"
  exit 1
fi

if ! which bazel > /dev/null; then
  echo "ERROR: bazel has not been installled"
  exit 1
fi

echo "** Install requirements"
sudo apt install -y llvm-10* clang-10*
sudo apt install -y libhdf5-serial-dev hdf5-tools libhdf5-dev zlib1g-dev \
                    zip libjpeg8-dev liblapack-dev libblas-dev gfortran
sudo pip3 install -U six wheel setuptools typing_extensions
sudo pip3 install pkgconfig
sudo env H5PY_SETUP_REQUIRES=0 pip3 install -U h5py==3.1.0
sudo pip3 install -U future==0.18.2 mock==3.0.5 \
                     keras_preprocessing==1.1.2 keras_applications==1.0.8 \
                     gast==0.4.0 futures pybind11

echo "** Download and patch tensorflow-${version}"
pushd $src_folder
if [ ! -f tensorflow-${version}.tar.gz ]; then
  wget https://github.com/tensorflow/tensorflow/archive/v${version}.tar.gz -O tensorflow-${version}.tar.gz
fi
tar xzvf tensorflow-${version}.tar.gz
cd tensorflow-${version}

patch -N -p1 < $patch_path && \
  echo "tensorflow-${version} source tree appears to be patched already.  Continue..."

echo "** Configure and build tensorflow-${version}"
export TMP=/tmp
export CC=/usr/bin/clang-10
export CXX=/usr/bin/clang++-10
export CXXFLAGS="-stdlib=libc++"

PYTHON_BIN_PATH=$(which python3) \
PYTHON_LIB_PATH=$(python3 -c 'import site; print(site.getsitepackages()[0])') \
TF_CUDA_COMPUTE_CAPABILITIES=${cuda_compute} \
TF_CUDA_VERSION=10.2 \
TF_CUDA_CLANG=1 \
TF_DOWNLOAD_CLANG=0 \
CLANG_CUDA_COMPILER_PATH=/usr/bin/clang-10 \
TF_CUDNN_VERSION=8 \
TF_TENSORRT_VERSION=${trt_version} \
CUDA_TOOLKIT_PATH=/usr/local/cuda \
CUDNN_INSTALL_PATH=/usr/lib/aarch64-linux-gnu \
TENSORRT_INSTALL_PATH=/usr/lib/aarch64-linux-gnu \
TF_NEED_IGNITE=0 \
TF_ENABLE_XLA=0 \
TF_NEED_OPENCL_SYCL=0 \
TF_NEED_COMPUTECPP=0 \
TF_NEED_ROCM=0 \
TF_NEED_CUDA=1 \
TF_NEED_TENSORRT=1 \
TF_NEED_OPENCL=0 \
TF_NEED_MPI=0 \
GCC_HOST_COMPILER_PATH=$(which clang-10) \
CC_OPT_FLAGS="-Wno-sign-compare" \
TF_SET_ANDROID_WORKSPACE=0 \
  ./configure

bazel build --config=opt \
            --config=cuda \
            --config=noaws \
            --local_cpu_resources=HOST_CPUS*0.25 \
            --local_ram_resources=HOST_RAM*0.5 \
            //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package wheel/tensorflow_pkg

echo "** Install tensorflow-${version}"
sudo pip3 install wheel/tensorflow_pkg/tensorflow-${version}-*.whl

popd

TF_CPP_MIN_LOG_LEVEL=3 \
  python3 -c "import tensorflow as tf; tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR); print('tensorflow version: %s' % tf.__version__); print('tensorflow.test.is_built_with_cuda(): %s' % tf.test.is_built_with_cuda()); print('tensorflow.test.is_gpu_available(): %s' % tf.test.is_gpu_available(cuda_only=False, min_cuda_compute_capability=None))"

echo "** Build and install tensorflow-${version} successfully"

protobuf-3.9.2-error.log

make[3]: Entering directory '/home/nvidia/src/protobuf-3.9.2/src'
../test-driver: line 107: 12829 Segmentation fault      (core dumped) "$@" > $log_file 2>&1
FAIL: protobuf-test
PASS: protobuf-lazy-descriptor-test
PASS: protobuf-lite-test
PASS: google/protobuf/compiler/zip_output_unittest.sh
PASS: google/protobuf/io/gzip_stream_unittest.sh
PASS: protobuf-lite-arena-test
PASS: no-warning-test
============================================================================
Testsuite summary for Protocol Buffers 3.9.2
============================================================================
# TOTAL: 7
# PASS:  6
# SKIP:  0
# XFAIL: 0
# FAIL:  1
# XPASS: 0
# ERROR: 0
============================================================================
See src/test-suite.log
Please report to protobuf@googlegroups.com
============================================================================
Makefile:7751: recipe for target 'test-suite.log' failed
make[3]: *** [test-suite.log] Error 1
make[3]: Leaving directory '/home/nvidia/src/protobuf-3.9.2/src'
Makefile:7857: recipe for target 'check-TESTS' failed
make[2]: *** [check-TESTS] Error 2
make[2]: Leaving directory '/home/nvidia/src/protobuf-3.9.2/src'
Makefile:7972: recipe for target 'check-am' failed
make[1]: *** [check-am] Error 2
make[1]: Leaving directory '/home/nvidia/src/protobuf-3.9.2/src'
Makefile:1715: recipe for target 'check-recursive' failed
make: *** [check-recursive] Error 1

tensorflow-2.6.0.patch

diff -Naur a/third_party/nccl/build_defs.bzl.tpl b/third_party/nccl/build_defs.bzl.tpl
--- a/third_party/nccl/build_defs.bzl.tpl	2021-08-10 03:10:27.000000000 +0800
+++ b/third_party/nccl/build_defs.bzl.tpl	2021-10-08 10:26:33.536077745 +0800
@@ -43,7 +43,7 @@
     # The global functions can not have a lower register count than the
     # device functions. This is enforced by setting a fixed register count.
     # https://github.com/NVIDIA/nccl/blob/f93fe9bfd94884cec2ba711897222e0df5569a53/makefiles/common.mk#L48
-    maxrregcount = "-maxrregcount=96"
+    maxrregcount = "-maxrregcount=80"
 
     return cuda_default_copts() + select({
         "@local_config_cuda//:is_cuda_compiler_nvcc": [

Hi There,
I am trying to use your valuable sh script to install tensorflow 2.6.0 in my ubuntu 18.04. But I got error below:
You have bazel 4.2.0 installed.
Please downgrade your bazel installation to version 3.99.0 or lower to build TensorFlow! To downgrade: download the installer for the old version (from https://github.com/bazelbuild/bazel/releases) then run the installer.

Could you please advise if you were facing similar issue before?
Much appreciate.

Have you tried to downgrade your bezel installation (bezel 3.7.2 for tensorflow 2.6.0)? You have to use the matching version of bezel otherwise the compilation process might likely fail.

Please refer to https://www.tensorflow.org/install/source

GPU

Version	Python version	Compiler	Build tools	cuDNN	CUDA
......	......	......	......	...	...
tensorflow-2.6.0	3.6-3.9	GCC 7.3.1	Bazel 3.7.2	8.1	11.2

I am not so familiar with bazel so not sure what's the best way to downgrade my bazel( it was bazel 4.2.0) to 3.7.2.
what i did:

1. Download bazel-3.7.2-linux-arm64 from https://github.com/bazelbuild/bazel/releases?expanded=true&page=1&q=3.7 to my robot(Jetson Nano Jetpack 4.6, Python 3.6). This looks exe file but i don't know how to run it(tried double click etc);
2. Download bazel-3.7.2-installer-linux-x86_64.sh from same side. to run it as : sudo bash bazel-3.7.2-installer-linux-x86_64.sh.

Then run install_tensorflow-2.6.0.sh again. I got below error.
Configuration finished
/usr/local/bin/bazel: line 163: /usr/local/lib/bazel/bin/bazel-real: cannot execute binary file: Exec format error
ERROR: The project you're trying to build requires Bazel 3.7.2 (specified in /home/hiwonder/src/tensorflow-2.6.0/.bazelversion), but it wasn't found in /usr/local/lib/bazel/bin.

And I can see bazel
hiwonder@JetMax:~/tensorflow$ ls -trl /usr/local/lib/bazel/bin
total 46420
-rwxr-xr-x 1 root root 47154615 Jan 1 1980 bazel-real
-rw-r--r-- 1 root root 5305 Jan 1 1980 bazel.fish
-rwxr-xr-x 1 root root 339462 Jan 1 1980 bazel-complete.bash
-rw-r--r-- 1 root root 10238 Jan 1 1980 _bazel
-rwxr-xr-x 1 root root 8678 Jan 1 1980 bazel

I don't know where to go now. would you please advise?