Deep Learning with AMD GPU on Windows

First written：2022/09/26

Last updated：2023/07/15

In the past，using AMD GPU for DL，we need Linux with ROCm installed.

Now, Microsoft published DirectML, which makes any GPU supporting DirectX12 be able to be used for DL on Windows.

How to config

Choose PyTorch-DirectML as an example：

1. (Optional) Install MiniConda environment，and set environment variables：

   Path\To\Miniconda3
   Path\To\Miniconda3\Scripts
   Path\To\Miniconda3\Library\bin
   

2. Create a Python env（If without Conda, you also need Python3.8-3.10)

   conda create -n directML python=3.8
   

3. Activate env and install dependencies

   conda activate directML
   pip install torch-directml
   

4. Usage

   import torch_directml
   dml = torch_directml.device() # device=dml
   

Run an Opensource Project

Choose GFPGAN from Tencent as an example, which aims at recover faces, with Real-ESRGAN built in, which can enhance the background.

1. Follow its Guide：

   Mentioned: If you want to enhance the background, Real-ESRGAN installation is needed

   After Guide, download pre-trained model:

   wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models
   

   You can use other tools to download to specific directory.

2. How to use：

   Usage: python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2 [options]...
   
     -h                   show this help
     -i input             Input image or folder. Default: inputs/whole_imgs
     -o output            Output folder. Default: results
     -v version           GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3
     -s upscale           The final upsampling scale of the image. Default: 2
     -bg_upsampler        background upsampler. Default: realesrgan
     -bg_tile             Tile size for background sampler, 0 for no tile during testing. Default: 400
     -suffix              Suffix of the restored faces
     -only_center_face    Only restore the center face
     -aligned             Input are aligned faces
     -ext                 Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto
   

3. Because basicsr and facexlib both depend on torch, you need to uninstalltorch, and force to reinstall pytorch-directml: (Dependency relations will not be satisfied, but ok)

   pip uninstall torch
   pip install --force-reinstall pytorch-directml
   

4. About Real-ESRGAN

   The unoptimized RealESRGAN is slow on CPU. If you really want to use it on CPU, please modify the corresponding codes: (inference_gfpgan.py Line59)

   # ------------------------ set up background upsampler ------------------------
   if args.bg_upsampler == 'realesrgan':
       from basicsr.archs.rrdbnet_arch import RRDBNet
       from realesrgan import RealESRGANer
       model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
       bg_upsampler = RealESRGANer(
           scale=2,
           model_path='https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth',
           model=model,
           tile=args.bg_tile,
           tile_pad=10,
           pre_pad=0,
           half=False)  # use fp16, need to set False in CPU mode
   else:
       bg_upsampler = None
   

   Everything is OK here.

5. DirectML Support

   In inference_gfpgan.py: add device parameter to constructors of RealESRGANer and GFPGANer.

   import torch_directml
   dml = torch_directml.device()
   
   bg_upsampler = RealESRGANer(..., device=dml)
   restorer = GFPGANer(..., device=dml)
   

   But there are currently two problems:

   • upsampler triggers an assertion in basicsr\archs\arch_util.py after processing several tiles, which is not solved:

     Assert that hh % scale == 0 and hw % scale == 0
     

   • If realesrgan is disabled, the face can be processed normally, but result is very weird.

     In order to avoid discomfort, no results are shown here.

     The CPU output are normal, but the DirectML output is abnormal.

     After analysis, the problem may be in the bilinear interpolation calculation, I have raised an issue in the DirectML project, click here.