can't inference and there are requirement conflicts

[seekingdeep]

can't inference and there are requirement conflicts

[Syzygianinfern0]

1. We are not sure how the dependency issue of torchvision and torch has crept up into the requirements file. Possibly the compatibility of them during 1.7.1's early releases could have been alright (when we installed it), and then incompatibility introduced during nightly versions and continues to exist. Anyways, thank you for pointing out the compatibility issue.
2. Although the cudatoolkit shipped with torch=1.7.1 is not of 10.0 version, the CUDA modules in our code do not depend on them. They rather compile on the system CUDA version (the one linked to nvcc).

[seekingdeep]

2. i have Cuda 10.1 installed on my system itself. and also using torch with cuda. I have tested both the conda and pip installation, still getting the attached error.

Basically, i cant inference so i can't test the quality of the output results, thus i can't decide whether to develop on the code.

[Syzygianinfern0]

CUDA is unfortunately not backward compatible (I have faced the same issue). To run code that depends on CUDA 10.0, we cannot use CUDA 10.1.

[seekingdeep]

what about the pytorch and torchvision versions required, which versions should i install?

[Syzygianinfern0]

Please go ahead with torch 1.7.1

[seekingdeep]

should i install the cpuonly version, because there is no cuda10.0 with pytorch 1.7.1

[Syzygianinfern0]

No. As I mentioned here, the CUDA modules in the code depend on the system CUDA. The python's cuda do not matter for it.

Let me lay out the complete configuration again.

• Python related
  • pytorch==1.7.1 torchvision==0.8.2
  • cudatoolkit (of your choice)
• System related
  • CUDA 10.0

[seekingdeep]

i have Cuda 10.0 and Cudnn 7.3 installed on my system
i installed pytorch 1.7.1 torchvision 0.8.2 using conda with cuda 10.1 in the conda environment
i built detectron2 0.4 from source
but still i am getting error when trying to inference:

(d2) home@home-lnx:~/programs/Palmira$ python demo.py --input ./input/1.jpg --output ./output --config configs/palmira/Palmira.yaml --opts MODEL.WEIGHTS ./Palmira_indiscapes.pth
Using /home/home/.cache/torch_extensions as PyTorch extensions root...
Detected CUDA files, patching ldflags
Emitting ninja build file /home/home/.cache/torch_extensions/check_condition_bbox/build.ninja...
Building extension module check_condition_bbox...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
ninja: no work to do.
Loading extension module check_condition_bbox...
Traceback (most recent call last):
  File "demo.py", line 14, in <module>
    from defgrid.config import add_defgrid_maskhead_config
  File "/home/home/programs/Palmira/defgrid/__init__.py", line 2, in <module>
    from .mask_head import DefGridHead
  File "/home/home/programs/Palmira/defgrid/mask_head.py", line 17, in <module>
    from defgrid.layers.DefGrid.diff_variance import LatticeVariance
  File "/home/home/programs/Palmira/defgrid/layers/DefGrid/diff_variance.py", line 7, in <module>
    from defgrid.layers.DefGrid.check_condition_lattice_bbox.utils import check_condition_f_bbox
  File "/home/home/programs/Palmira/defgrid/layers/DefGrid/check_condition_lattice_bbox/utils.py", line 22, in <module>
    verbose=True,
  File "/home/home/anaconda3/envs/d2/lib/python3.7/site-packages/torch/utils/cpp_extension.py", line 997, in load
    keep_intermediates=keep_intermediates)
  File "/home/home/anaconda3/envs/d2/lib/python3.7/site-packages/torch/utils/cpp_extension.py", line 1213, in _jit_compile
    return _import_module_from_library(name, build_directory, is_python_module)
  File "/home/home/anaconda3/envs/d2/lib/python3.7/site-packages/torch/utils/cpp_extension.py", line 1564, in _import_module_from_library
    return imp.load_module(module_name, file, path, description)
  File "/home/home/anaconda3/envs/d2/lib/python3.7/imp.py", line 242, in load_module
    return load_dynamic(name, filename, file)
  File "/home/home/anaconda3/envs/d2/lib/python3.7/imp.py", line 342, in load_dynamic
    return _load(spec)
ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory

[seekingdeep]

my main system, Cuda version

home@home-lnx:~$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130

my main system, Cudnn version

home@home-lnx:~$ cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2
#define CUDNN_MAJOR 7
#define CUDNN_MINOR 3
#define CUDNN_PATCHLEVEL 0
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

#include "driver_types.h"

[seekingdeep]

hmmm, this seems to be caused by having cuda 10.0 installed on the system, while pytorch 1.7.1 comes with cuda 10.1
as explained here facebookresearch/detectron2#588 (comment)
it seems that pytorch needs to be compiled from source.
Wouldn't it be better to make this repository code compatible with pytorch prebuilts

[seekingdeep]

Solved by building pytorch 1.7.1 and torchvision 0.8.2 from source within the environment.
build detectron 0.4 from source within the environment.
installed cuda 10.0 and cudnn 7.3 on main terminal

[Syzygianinfern0]

it seems that pytorch needs to be compiled from source.

We did not have to do that. Our torch and detectron were installed from pre-built binaries. Also, some other people from our group have reproduced the setup necessary to run this code.

Wouldn't it be better to make this repository code compatible with pytorch prebuilts

We believed it was. However, after your case we might need to investigate further. Sorry for your inconveniences. Glad to know its fixed now anyways.