The solution of problem about: no module named 'enclib_cpu' or no module named 'enclib_gpu'

[Chrisqyf]

If you want to solve this problem, you can follow this reference: https://zhuanlan.zhihu.com/p/191314560
To be specific, you can get 'enclib_cpu.so' or 'enclib_gpu.so' by 'build.ninja' in '/site-packages/encoding/lib/cpu' or '/site-packages/encoding/lib/gpu'.
But in this process, I met a new problem: nvcc fatal: Unsupported gpu architecture 'compute_86'
One widely-used solution is to set environment variables to reduce computing power requirements.
e.g. export TORCH_CUDA_ARCH_LIST="7.5"
CUDA 10.1 does not support compute_86, while CUDA 10.1 supports compute_75 at most.
I failed again. This solution was not worked.
My final solution is edit the 'build.ninja': change '-gencode arch=compute_86, code=sm_86' into '-gencode arch=compute_75, code=sm_75'
I got it!
Hope my experience is helpful for ones struggling the same problem.

note:
My environment settings: Ubuntu 20.04 CUDA 10.1 python=3.6 torch=1.4.1
There are many problems for configuring pytorch-encoding in Windows system, in my experience. So the successful configuration may be easier in Linux or mac.

[Chrisqyf]

There is a new problem!
RuntimeError: CUDA error: no kernel image is available for execution on the device
The reason is cuda or pytorch do not match the computing power of my graphics card.
This is essentially the same problem as the previous 'nvcc fatal'. The previous edit in 'build.ninja' just haven't solved the essential problem.
And I have two ways to deal with this:

1. Use torch cpu
   
   After installing torch, you should modify the parameter 'map_location' in torch.load(), like this:
   
   The code is in the '/site-packages/encoding/models/deepten.py'
   It's worked!
   

2. Change the version of CUDA and pytorch
   I'm trying

[Chrisqyf]

2. Change the version of CUDA and pytorch
   Current environment settings: Ubuntu 20.04 CUDA 11.1 python=3.8 torch=1.8.1
   This method worked, too.

But in the testing process, there was a 'RuntimeError':

We should click the error link, and add '.contiguous()' before '.view()'