PyTorch

PyTorch is an alternative to TensorFlow.

If using a KVM, be sure CPU type is set to host.

Get Source

Get PyTorch source code:

# This is about 1 gig:
git clone --recursive https://github.com/pytorch/pytorch

The recursive git repos contain a mix of permissive licenses, mostly BSD, MIT, Apache style. No GPL. License owners are mostly Google and Facebook, with a mix of many others.

Build Py

Build from scratch with free software options. PyTorch has a CMakeLists.txt, so lets see how it goes with a cmake build....

Install build dependencies:

apt install cmake cmake-curses-gui g++ python-yaml python-typing
# Note, it uses python-yaml, not python3-yaml...

# Optional deps:
apt install doxygen
apt install libfftw3-dev
apt install libgmp3-dev
apt install libmpfr-dev
apt install libmkldnn-dev
apt install libnuma-dev # Nope, not right one
# Optional deps for BLAS OpenBLAS
apt install libopenblas-dev
# OpenCV
apt install libopencv-dev
# pybind?
apt install pybind11-dev
 pybind11_INCLUDE_DIR             /usr/include/pybind11                                                                                           

# BLAS Eigen
apt install libeigen3-dev # fail?
# ffmpeg
apt install libavcodec-dev libavdevice-dev libavfilter-dev libavformat-dev libavresample-dev libavutil-dev libpostproc-dev libswresample-dev libswscale-dev ffmpeg
# Thusly:
 USE_FFMPEG ON
 FFMPEG_AVCODEC_INCLUDE_DIR       /usr/include/x86_64-linux-gnu/libavcodec                                                                        
 FFMPEG_LIBAVCODEC                /usr/lib/x86_64-linux-gnu                                                                                       
 FFMPEG_LIBAVFORMAT               /usr/lib/x86_64-linux-gnu                                                                                       
 FFMPEG_LIBAVUTIL                 /usr/lib/x86_64-linux-gnu                                                                                       
 FFMPEG_LIBSWSCALE                /usr/lib/x86_64-linux-gnu                                                                                       
# or?
 FFMPEG_AVCODEC_INCLUDE_DIR       /usr/include/x86_64-linux-gnu/libavcodec                                                                        
 FFMPEG_LIBAVCODEC                /usr/lib/x86_64-linux-gnu/libavcodec.so                                                                         
 FFMPEG_LIBAVFORMAT               /usr/lib/x86_64-linux-gnu/libavformat.so                                                                        
 FFMPEG_LIBAVUTIL                 /usr/lib/x86_64-linux-gnu/libavutil.so                                                                          
 FFMPEG_LIBSWSCALE                /usr/lib/x86_64-linux-gnu/libswscale.so                                                                         


# XXX
TORCH_BUILD_VERSION  default is 1.1.0

mkdir build
cd build
cmake ..
ccmake ..
# build takes ~30 minutes
make -j8

#cmake note
  Generated cmake files are only fully tested if one builds with system glog,
   gflags, and protobuf.  Other settings may generate files that are not well
   tested.

# so maybe:
apt install libgflags-dev
apt install libprotobuf-dev
# glog?
apt install libgoogle-glog-dev # ???

Or:

python setup.py install

How docs say:

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py build --cmake-only
ccmake build  # or cmake-gui build

Notable build options:

BLAS --- has multiple options:
         MKL (non-free bits possible)
         vecLib (?)
         FLAME (?)
         Eigen (in Debian)
         ATLAS (in Debian?)
         OpenBLAS (in Debian)

More options:

BUILD_BINARY ON
BUILD_PYTHON ON
OPENMP_FOUND ON
USE_CUDA OFF
USE_CUDNN OFF
USE_DISTRIBUTED ON
USE_FFMPEG ON
USE_MKLDNN ON  ?? with Debian's packages ?
USE_MKLDNN_CBLAS NO ?
USE_LEVELDB ?? Available in debian.
USE_LMDB ?? Available in Debian.
USE_MPI ON heh
USE_NCCL OFF  nvidia
USE_NUMA ON  ?
USE_NUMPY ON
USE_OBSERVERS ON ?
USE_OPENCL ON ?
USE_OPENCV ON ?
USE_OPENMP ON ?
USE_REDIS ?
USE_ROCKSDB Available in Debian, like leveldb and lmdb
USE_ZMQ Available in Debian, messaging.
USE_ZSTD Available in Debian, compression.
WITH_BLAS ?
WITH_OPENMP ON
CAFFE2_USE_MSVC_STATIC_RUNTIME OFF ??
BUILD_CAFFE2_MOBILE OFF

More misc...

git checkout v1.4.0
apt install python3-dev python3-numpy python-numpy
# uh
apt install libcaffe-cpu-dev

Seems to be using not python3??

Misc

[E init_intrinsics_check.cc:43] CPU feature avx is present on your machine, but the Caffe2 binary is not compiled with it. It means you may not get the full speed of your CPU.

Build PyTorch `pip`

Lets get old Python 2 out of here:

apt autoremove --purge python2 python2-minimal python2.7 python2.7-minimal libpython2.7 libpython2.7-minimal libpython2.7-stdlib
git checkout v1.4.0
apt install python3-pip python3-setuptools
# Docs recommend this, but mkl is proprietary:
pip3 install --user numpy ninja pyyaml mkl mkl-include setuptools cmake cffi
# Try:
pip3 install --user --upgrade pip
# Set that new pip in your path, ~/.local/bin/ in ~/.bashrc:
export PATH="~/.local/bin:/usr/lib/ccache:$PATH"

pip3 install --user cmake
pip3 install --user numpy ninja pyyaml setuptools cmake cffi
# From upstream docs:
pip install torch==1.4.0+cpu torchvision==0.5.0+cpu -f https://download.pytorch.org/whl/torch_stable.html

Build with Python Setup

git submodule sync
git submodule update --init --recursive
python3 setup.py build --cmake-only
ccmake build
python3 setup.py install --user

Proprietary Bits

Parts to avoid in the PyTorch ecosystem:

Anaconda is a package manager for Python.The Anaconda repository contains non-free software, so don't use it.
"MKL" (Intel's non-free binaries).
Note, these MKL packages are in Debian: libmkldnn0 libmkldnn-dev libmkldnn-doc
Intel compiler.

5.8 KiB Raw Permalink Blame History