Usage instructions/inaccurates with CUDA support
toynbeeidea opened this issue · comments
I am not using Ubuntu, but a different distro. Ubuntu instructions say, if you have CUDA 10.1, to run pip install onnxruntime-gpu. However, running python demo.py --debug yields the following:
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory
(its actually a MUCH longer error message, but that's the gist at the end)
This is because tensorflow-gpu, or at least the required version of tensorflow-gpu in the GPU requirements file (1.12), does not support CUDA 10.x. But there is a further line in the instructions: "If you have CUDA 10.1, pip install onnxruntime-gpu" Well, I have done that, but it doesn't seem like demo.py knows what to do with that.
I could downgrade to CUDA 10.1 (current installed version is libcublas.so.11.3.1.68), but that seems pointless if Tensorflow requires 9.0. So... what's missing here?
I have attempted to upgrade to the current version of tensowflow-gpu (2.4.1), but running demo.py then yields the following:
2021-01-26 16:42:23.851549: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0
Traceback (most recent call last):
File "demo.py", line 192, in <module>
main()
File "demo.py", line 42, in main
import face_alignment
File "/home/erik/.conda/envs/vtuber/lib/python3.6/site-packages/face_alignment/__init__.py", line 7, in <module>
from .api import FaceAlignment, LandmarksType, NetworkSize
File "/home/erik/.conda/envs/vtuber/lib/python3.6/site-packages/face_alignment/api.py", line 7, in <module>
from .utils import *
File "/home/erik/.conda/envs/vtuber/lib/python3.6/site-packages/face_alignment/utils.py", line 13, in <module>
from torch.hub import download_url_to_file, HASH_REGEX
ImportError: cannot import name 'download_url_to_file'