Merge pull request #1 from gremlinflat/master

xinntao#584 - custom backend device (auto, cuda, m1, cpu)

.gitignore

@@ -138,3 +138,7 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# VSCode
+.vscode/
+.vscode/*

inference_realesrgan.py

@@ -8,6 +8,8 @@
 from realesrgan import RealESRGANer
 from realesrgan.archs.srvgg_arch import SRVGGNetCompact
 
+from torch.cuda import is_available as cudaIsAvailable
+from torch.backends.mps import is_available as mpsIsAvailable
 
 def main():
     """Inference demo for Real-ESRGAN.
@@ -52,6 +54,8 @@ def main():
     parser.add_argument(
         '-g', '--gpu-id', type=int, default=None, help='gpu device to use (default=None) can be 0,1,2 for multi-gpu')
 
+    parser.add_argument('--backend_type', type=str, default='auto', choices=['auto', 'cuda', 'cpu', 'mps'], help='backend type. Options: auto(cuda-cpu) | cuda | cpu | mps')
+
     args = parser.parse_args()
 
     # determine models according to model names
@@ -103,6 +107,21 @@ def main():
         model_path = [model_path, wdn_model_path]
         dni_weight = [args.denoise_strength, 1 - args.denoise_strength]
 
+    # deternime backend type (cpu, cuda, mps)
+    if args.backend_type == 'auto':
+        if cudaIsAvailable():
+            backend_type = 'cuda'
+        elif mpsIsAvailable():
+            backend_type = 'mps'
+        else:
+            backend_type = 'cpu'
+    elif args.backend_type == 'cuda' and cudaIsAvailable():
+        backend_type = 'cuda'
+    elif args.backend_type == 'mps' and mpsIsAvailable():
+        backend_type = 'mps'
+    else:
+        backend_type = 'cpu'
+
     # restorer
     upsampler = RealESRGANer(
         scale=netscale,
@@ -113,6 +132,7 @@ def main():
         tile_pad=args.tile_pad,
         pre_pad=args.pre_pad,
         half=not args.fp32,
+        device=backend_type,
         gpu_id=args.gpu_id)
 
     if args.face_enhance:  # Use GFPGAN for face enhancement
@@ -122,6 +142,7 @@ def main():
             upscale=args.outscale,
             arch='clean',
             channel_multiplier=2,
+            device='cpu', # <--- MPS is not supported yet, crash pas runtime. TODO: FIX THIS
             bg_upsampler=upsampler)
     os.makedirs(args.output, exist_ok=True)
 

realesrgan/utils.py

@@ -89,7 +89,7 @@ def pre_process(self, img):
         """Pre-process, such as pre-pad and mod pad, so that the images can be divisible
         """
         img = torch.from_numpy(np.transpose(img, (2, 0, 1))).float()
-        self.img = img.unsqueeze(0).to(self.device)
+        self.img = img.unsqueeze(0).contiguous().to(self.device)
         if self.half:
             self.img = self.img.half()