How to run the demo on pytorch 1.4.0 or 1.1.0? about gfpgan HOT 2 CLOSED

I managed to run it with Pytorch 1.3 since it's the latest build that supports my GPU. Windows 8.1, Python 3.7, CUDA 10.2
You need to:

1. Install basicsr, I already had it installed. If I remember correctly, I cloned the repo and modified requirements.txt to change pytorch version, and executed pip install -e /path/to/basicsr
   If you have version error with basicsr, just remove the line with version from basicsr/__init__.py
2. Run pip install ninja
3. Go to your python installation path, lib/site-packages/ and fix torch/hub.py , add

def get_dir():
    return hub_dir 
#required by facexlib,  implemented only in Pytorch 1.6+

4. Run pip install facexlib
5. Edit requirements.txt to exclude fresh pytorch, then run pip install -r requirements.txt
6. Similar to step 3, fix torch/utils/cpp_extension.py

• if you have utf-8 errors, change all decode() to decode('encoding-for-your-spoken-language')
• in function _prepare_ldflags add extra_ldflags.append('c10_cuda.lib') near the other similar lines

7. Run set BASICSR_JIT=True
8. Run the py script from the Command Line Prompt of VS[YEAR] x64 Native Tools (Visual Studio Build Tools)

from gfpgan.

I have not tested it on pytorch1.4 or 1.1.
You may adapt to that version.

from gfpgan.