问题
I'm using anaconda to regulate my environment, for a project i have to use my GPU for network training. I use pytorch for my project and i'm trying to get CUDA working.
I installed cudatoolkit, numba, cudnn
still, when i try this command:
torch.cuda.is_available()
I get "False" as output. This is my environment:
# Name Version Build Channel
blas 1.0 mkl
bzip2 1.0.6 h470a237_2 conda-forge
ca-certificates 2018.03.07 0
cairo 1.14.12 he6fea26_5 conda-forge
certifi 2018.8.24 py35_1
cffi 1.11.5 py35he75722e_1
cloudpickle 0.5.5 py35_0
cudatoolkit 9.2 0 anaconda
cudnn 7.2.1 cuda9.2_0 anaconda
cycler 0.10.0 py_1 conda-forge
cython 0.28.5 py35hf484d3e_0 anaconda
dask-core 0.19.2 py35_0
dbus 1.13.0 h3a4f0e9_0 conda-forge
decorator 4.3.0 py35_0
expat 2.2.5 hfc679d8_2 conda-forge
ffmpeg 4.0.2 ha0c5888_1 conda-forge
fontconfig 2.13.1 h65d0f4c_0 conda-forge
freetype 2.9.1 h6debe1e_4 conda-forge
gettext 0.19.8.1 h5e8e0c9_1 conda-forge
giflib 5.1.4 h470a237_1 conda-forge
glib 2.55.0 h464dc38_2 conda-forge
gmp 6.1.2 hfc679d8_0 conda-forge
gnutls 3.5.19 h2a4e5f8_1 conda-forge
graphite2 1.3.12 hfc679d8_1 conda-forge
gst-plugins-base 1.12.5 hde13a9d_0 conda-forge
gstreamer 1.12.5 h61a6719_0 conda-forge
harfbuzz 1.9.0 h08d66d9_0 conda-forge
hdf5 1.10.2 hc401514_2 conda-forge
icu 58.2 hfc679d8_0 conda-forge
imageio 2.4.1 py35_0
intel-openmp 2019.0 118
jasper 1.900.1 hff1ad4c_5 conda-forge
jpeg 9c h470a237_1 conda-forge
kiwisolver 1.0.1 py35h2d50403_2 conda-forge
libedit 3.1.20170329 h6b74fdf_2
libffi 3.2.1 hd88cf55_4
libgcc-ng 8.2.0 hdf63c60_1
libgfortran 3.0.0 1 conda-forge
libgfortran-ng 7.3.0 hdf63c60_0
libiconv 1.15 h470a237_3 conda-forge
libopenblas 0.3.3 h5a2b251_3
libpng 1.6.35 ha92aebf_2 conda-forge
libstdcxx-ng 8.2.0 hdf63c60_1
libtiff 4.0.9 he6b73bb_2 conda-forge
libuuid 2.32.1 h470a237_2 conda-forge
libwebp 0.5.2 7 conda-forge
libxcb 1.13 h470a237_2 conda-forge
libxml2 2.9.8 h422b904_5 conda-forge
llvmlite 0.24.0 py35hdbcaa40_0
matplotlib 3.0.0 py35h0b34cb6_1 conda-forge
mkl 2019.0 118
mkl_fft 1.0.6 py35_0 conda-forge
mkl_random 1.0.1 py35_0 conda-forge
ncurses 6.1 hf484d3e_0
nettle 3.3 0 conda-forge
networkx 2.1 py35_0
ninja 1.8.2 py35h6bb024c_1
numba 0.39.0 py35h04863e7_0
numpy 1.15.2 py35h1d66e8a_0
numpy-base 1.15.2 py35h81de0dd_0
olefile 0.46 py35_0
openblas 0.2.20 8 conda-forge
opencv 3.4.1 py35h6fd60c2_1
opencv-python 3.4.3.18 <pip>
openh264 1.7.0 0 conda-forge
openssl 1.0.2p h14c3975_0
pandas 0.23.4 py35h04863e7_0
pcre 8.41 hfc679d8_3 conda-forge
pillow 5.2.0 py35heded4f4_0
Pillow 5.3.0 <pip>
pip 10.0.1 py35_0
pixman 0.34.0 h470a237_3 conda-forge
pthread-stubs 0.4 h470a237_1 conda-forge
pycparser 2.19 py35_0
pyparsing 2.2.2 py_0 conda-forge
pyqt 5.6.0 py35h8210e8a_7 conda-forge
python 3.5.6 hc3d631a_0
python-dateutil 2.7.3 py_0 conda-forge
pytorch 0.4.1 py35_py27__9.0.176_7.1.2_2 pytorch
pytz 2018.5 py35_0
pywavelets 1.0.0 py35hdd07704_0
qt 5.6.2 hf70d934_9 conda-forge
readline 7.0 h7b6447c_5
scikit-image 0.14.0 py35hf484d3e_1
scipy 1.1.0 py35hfa4b5c9_1
setuptools 40.2.0 py35_0
sip 4.18.1 py35hfc679d8_0 conda-forge
six 1.11.0 py35_1 conda-forge
sqlite 3.25.2 h7b6447c_0
tk 8.6.8 hbc83047_0
toolz 0.9.0 py35_0
torchvision 0.1.9 py35h72e4c6f_1 soumith
tornado 5.1.1 py35h470a237_0 conda-forge
wheel 0.31.1 py35_0
x264 1!152.20180717 h470a237_1 conda-forge
xorg-kbproto 1.0.7 h470a237_2 conda-forge
xorg-libice 1.0.9 h470a237_4 conda-forge
xorg-libsm 1.2.2 h8c8a85c_6 conda-forge
xorg-libx11 1.6.6 h470a237_0 conda-forge
xorg-libxau 1.0.8 h470a237_6 conda-forge
xorg-libxdmcp 1.1.2 h470a237_7 conda-forge
xorg-libxext 1.3.3 h470a237_4 conda-forge
xorg-libxrender 0.9.10 h470a237_2 conda-forge
xorg-renderproto 0.11.1 h470a237_2 conda-forge
xorg-xextproto 7.3.0 h470a237_2 conda-forge
xorg-xproto 7.0.31 h470a237_7 conda-forge
xz 5.2.4 h14c3975_4
zlib 1.2.11 ha838bed_2
My desktop has a NVIDIA GeForce GTX 970 (so it is cuda available) Also for some reason, as you can see here:
My graphics card doesn't show, however when using the
lspci -v
command, i can see my graphics card there. Don't know if that has something to do with it. Anyone knows how i can fix this?
回答1:
FIXED: I installed CUDA seperately, not through anaconda and now it works. If anyone knows why it doesn't work when installing cuda in anaconda, feel free to answer
回答2:
You need to install pytorch "in one go" using https://pytorch.org/get-started/locally/ to construct the anaconda command.
With the standard configuration in anaconda, you get:
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
(Please always check on https://pytorch.org/get-started/locally/ whether this command is still up to date.)
You seem to need the right cuda version 10.2 package to be aligned with what pytorch can handle. This is what is meant with @RussellGallop's helpful message.
We can see that installing pytorch and cuda separately is not recommended, and that Anaconda installation is recommended, against your answer:
Anaconda is our recommended package manager since it installs all dependencies.
(Note: If in case of problems, you better uninstall all covered packages and apply https://pytorch.org/get-started/locally/ to get the command again, instead of trying to fix it with separate installations. And if you then want to uninstall, you need to use the exactly same command of the installation, but with "conda uninstall" instead. This needed uninstall "in one go" again is another hint at the sensitive installation of pytorch, and that separate installation is risky. See How can l uninstall PyTorch?)
回答3:
Your CUDA install needs to be compatible with your NVidia drivers (which aren't installed by Anaconda). I had a similar thing with Anaconda on Windows. I ignored the torch.cuda.is_available() check and tried to send a model to the GPU with:
device = torch.device("cuda")
model.to(device)
This gave a more helpful message:
The NVIDIA driver on your system is too old (found version 10010).
Please update your GPU driver by downloading and installing a new
version from the URL: http://www.nvidia.com/Download/index.aspx
Alternatively, go to: https://pytorch.org to install
a PyTorch version that has been compiled with your version
of the CUDA driver.
来源：https://stackoverflow.com/questions/53102436/pytorch-cuda-with-anaconda-not-available