Presvedcime sa ze mame 64 bitovy hardware a bezime na nom 64 bitove windowsy (win10, win8.1 alebo win7): Napriklad tak, ze stlacenim windows-I (vo Win10) sa dostaneme ku System/About - a vidime tam 64 bit operating system Dalej sa uistime ze mame graficku kartu od NVIDIA, napr. GeForce. Zistime si aku cudu pouziva aktualna verzia tensorflow Tento navod je pre verziu tensorflow 1.4.0 ktora pouziva CUDA 8 Novsia tensorflow 1.5.0 pouziva CUDA 9.0 (ale nedokaze pouzit CUDA 9.1) Nainstalujeme cuda 8 z https://developer.nvidia.com/cuda-80-ga2-download-archive cuda_8.0.61_win10.exe - express install + service pack cuda_8.0.61.2_windows.exe Po instalacii c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\bin pribudla do path v env (Control Panel / System / Advanced system settings / Environmental Variables -> edit path v System variables) a mame k dispozicii cudnn64_80.dll (pokial mate cygwin, mozete sa o tom presvedcit z cmd cez 'which cudnn64_80.dll') Pokial je mozne ze je na vasom pocitaci viac verzii cuda, presvedcte sa ze v path mate ako prvu tu spravnu: ked z cmd spustite nasledovny prikaz, vypise nasledovny vystup nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2016 NVIDIA Corporation Built on Mon_Jan__9_17:32:33_CST_2017 Cuda compilation tools, release 8.0, V8.0.60 nainstalujeme cudnn 6 z https://developer.nvidia.com/compute/machine-learning/cudnn/secure/v6/prod/8.0_20170307/cudnn-8.0-windows10-x64-v6.0-zip tak, ze bin lib a include zo tohto zipu vybalime do c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0 nasledne mame k dispozicii cudnn64_6.dll Presvedcime sa, ze nam cuda funguje, pustime z cmd niektore demo, napriklad: c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\extras\demo_suite>oceanFFT.exe Pokial nebezi: spravidla hlasi chybu nedostatocneho drivera, ktora sa riesit povysenim driverov: - v Control Panel / System / Device Manager / Display adapters tam su spravidla dva napr Intel HD Graphics a NVIDIA N16E-GT - na NVIDIA dame Update driver (trva to minuty) vysledkom je aj zmena nazvu, napr na NVIDIA GeForce GTX 970M - restartneme cele windowsy nainstalujeme Visual C++ Redistributable for Visual Studio 2015 Update 3 https://www.microsoft.com/en-us/download/details.aspx?id=53587 vc_redist.x64.exe Pozorne sledujme ci sa ho podari nainstalovat, ak nie, v Control Panel / Program and Features odinstalujeme vyssie verzie a zopakujeme instalaciu s uspechom. ziskame tak k dispozicii vcomp140.dll a msvcp140.dll Keras je zviazany s jazykom Python a vyzaduje verzie 3.5.x alebo 3.6.x Opiera sa pritom o rad inych softwarov, je mimoriadne citlivy na ich verzie a lahko sa pobije s inymi aktivitami ktore vykonavame s pythonom. Preto je odporucane zriadit si v pythone pre neho virtualne prostredie, ktore nas od tychto problemov ochrani. (Pokial si zelate keras nainstalovat rovno do Pythonu, vynechajte nasledujuci krok a potom pracujte pod C:\> nie pod (keras) C:\> a baliky kerasu su instalovane priamo v instalacnom adresari Pythonu Virtualne prostredie mozeme zriadit dvomi sposobmi - doplnit tuto moznost do standardnej instacie Pythonu alebo pouzit software podporujuci virtualne prostredia, ktory ma v sebe aj Python - vola sa Anaconda. Ak zatial nepouzivate Python, je lepsie odinstalovat vsetky verzie pythonu a pouzit Anaconda3. - variant s anacondou: nainstalujeme https://repo.continuum.io/archive/Anaconda3-4.2.0-Windows-x86_64.exe destination folder C:\Anaconda3 pokial nemate v system iny pouzivany python, dajte si pri instalacii anakondu zaregistrovat ako python3.5 do registrov, inak tento check box vypnite. pridajte do add to path v env C:\Anaconda3; C:\Anaconda3\Scripts spustite cmd vytvorime virtualne prostredie C:\> conda create -n keras python=3.5 fyzicky sa zjavi v C:\Anaconda3\envs\keras z cmd ho potom pustame cez C:\>activate keras (keras) C:\ - variant so standardnou distribuciou Pythonu nainstalujeme https://www.python.org/ftp/python/3.5.2/python-3.5.2-amd64.exe (volime Customized installation aby sme vedeli zadat vlastnu cestu c:\Python35) pridame c:\Python35 a c:\Python35\Scripts do path v env alebo (na tento ucel menej preferovanu vyssiu verziu) nainstalujeme https://www.python.org/ftp/python/3.6.3/python-3.6.3-amd64.exe do c:\Python36 (volime Customized installation aby sme vedeli zadat vlastnu cestu c:\Python36) a pridame c:\Python36 a c:\Python36\Scripts do path v env alebo niektoru z tychto a podobnych 64 bit verzii mame uz nainstalovaju a analogicke dve cesty mame v path pustime cmd C:\> pip install virtualenvwrapper-win Collecting virtualenvwrapper-win Downloading virtualenvwrapper-win-1.2.5.tar.gz Collecting virtualenv (from virtualenvwrapper-win) Using cached virtualenv-15.1.0-py2.py3-none-any.whl Installing collected packages: virtualenv, virtualenvwrapper-win Running setup.py install for virtualenvwrapper-win ... done Successfully installed virtualenv-15.1.0 virtualenvwrapper-win-1.2.5 v c:\Python35\Scripts\ pribudli mkvirtualenv.bat a workon.bat C:\>mkvirtualenv keras -p c:\Python35\python.exe Running virtualenv with interpreter c:\Python35\python.exe Using base prefix 'c:\\Python35' New python executable in C:\Users\24000031\Envs\keras\Scripts\python.exe Installing setuptools, pip, wheel...done. (keras) C:\> exit tymto v c:\Users\...\Envs\ pribudol adresar keras v ktorom sa budu zhromazdovat potrebne baliky kerasu virtualne prostredie budeme pouzivat tak, ze pustime cmd a dame C:\> workon keras (keras) C:\> nasledne - ci uz vo virtualnom prostredi alebo nie instalujeme do pythonu podporne baliky a baliky kerasu Keras je abstraktny frontend ktora dokaze bezat na dvomi roznymi backendami: Theano (rozirenie numphy) a tensorflow. Oba backendy rovnocenne implementuju numericke vypocty potrebne na realizaciu umelych neuronovych sieti. Default je tensorflow (da sa zmenit v c:\Users\...\.keras) a ten aj budeme pouzivat. nainstalujeme teda Theano (keras) C:\>pip install Theano to viac menej nebudeme pouzivat, ale musime ho nainstalovat nainstalujeme tensorflow (keras) C:\>pip install tensorflow-gpu keby robil drahoty, ze ziadna verzia mu nesedi, je mozne ho presviedcat (keras) C:\>pip install --ignore-installed --upgrade tensorflow-gpu pripadne zadat konkretnu a nizsiu verziu (aktualna je 1.4.0) (keras) C:\>pip install --ignore-installed --upgrade "tensorflow-gpu==1.3.0" (keras) C:\>pip install --ignore-installed --upgrade "tensorflow-gpu<1.4.0" ale to skor indikuje problem s niektorym vyssie uvedenym krokom po instalacii vyskusame: (keras) C:\>python >>> import tensorflow as tf nesmie hodit fail - spravidla hlasi ze nevie loadnut nejaku konkretnu DLL z vyssie uvedenych (to sme nezvladli krok ktory ju spomina) alebo nemenovanu DLL (vtedy nefunguje CUDA ako taka a pomoze napr updatnut drivere grafickej karty NVIDIA) >>> hello = tf.constant('Hello, TensorFlow!') >>> sess = tf.Session() >>> print(sess.run(hello)) z pythonu odideme cez >>> quit() (pouzite vzdy ked v navode skoncia >>>) stiahneme do napr C:\Downloads z https://gist.githubusercontent.com/mrry/ee5dbcfdd045fa48a27d56664411d41c/raw/5966d71d4356b91d96f3e4c3f18b657848110a23/tensorflow_self_check.py (keby sa link zmenil, hladajte cez internetovy vyhladavac tensorflow_self_check.py na github-e) (keras) C:\Downloads>python tensorflow_self_check.py TensorFlow successfully installed. The installed version of TensorFlow includes GPU support. keby sa nepodarilo tensorflow rozbehat na gpu, da sa (ale to uz nie je celkom ono) pouzit len CPU: (keras) C:\>pip uninstall tensorflow-gpu y (keras) C:\>pip install tensorflow nainstalujeme keras (keras) C:\>pip install keras vyskusame (keras) C:\>python >>> from keras.layers import Dense, Activation Using TensorFlow backend. >>> from keras.models import Sequential >>> nesmu hodit error nainstalujeme zopar toolov: (keras) C:\>pip install h5py pydot matplotlib (pri pouziti anakondy su uz nainstalovane) nainstalujme opencv: ak pouzivame odporucanu verziu Python 3.5: stiahneme opencv_python-3.4.0+contrib-cp35-cp35m-win_amd64.whl z https://www.lfd.uci.edu/~gohlke/pythonlibs/#opencv do napr. C:\Downloads (keras) C:\Downloads>python -m pip install opencv_python-3.4.0+contrib-cp35-cp35m-win_amd64.whl Processing C:\Downloads\opencv_python-3.4.0+contrib-cp35-cp35m-win_amd64.whl Installing collected packages: opencv-python Successfully installed opencv-python-3.4.0+contrib C:\Downloads> respektive ak pouzivame Python 3.6 tak analogicy s opencv_python-3.4.0+contrib-cp36-cp36m-win_amd64.whl vyskusame: (keras) C:\>python >>> import cv2 >>> nesmie hodit chybu Dalej (do windowsov) nainstalujeme https://graphviz.gitlab.io/_pages/Download/windows/graphviz-2.38.msi a do path v env pridame cestu c:\Program Files (x86)\Graphviz2.38\bin\ do env pridame premennu TF_CPP_MIN_LOG_LEVEL s hodnotou 2 zabrani to neprijemnym vypisom, ktore na prvy pohlad vyzeraju ako chyby, pritom vsak len oznamuju nevinne veci ako ze vasa graficka karta umoznuje este pokrocilejsie nastavenia nez dokaze keras vyuzit a pod. Vysledok instalacie je: (keras) C:\>pip list backports.weakref (1.0rc1) bleach (1.5.0) certifi (2016.2.28) cycler (0.10.0) enum34 (1.1.6) h5py (2.7.1) html5lib (0.9999999)Keras (2.1.2) Markdown (2.6.10) matplotlib (2.1.1) numpy (1.13.3) opencv-python (3.4.0+contrib) pip (9.0.1) protobuf (3.5.1) pydot (1.2.4) pyparsing (2.2.0) python-dateutil (2.6.1) pytz (2017.3) PyYAML (3.12) scipy (1.0.0) setuptools (38.2.5) six (1.11.0) tensorflow-gpu (1.4.0) tensorflow-tensorboard (0.4.0rc3) Theano (1.0.1) Werkzeug (0.13) wheel (0.30.0) wincertstore (0.2)