feat: support cpu transcoding

buganime/transcode.py

@@ -55,7 +55,10 @@ def __init__(self, input_path: str, output_path: str, height_out: int, width_out
         self.__width_out = width_out
         model = Transcoder.Module(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4)
         model.load_state_dict(torch.load(MODEL_PATH)['params'], strict=True)
-        self.__model = model.eval().cuda().half()
+        if torch.cuda.is_available():
+            self.__model = model.eval().cuda().half()
+        else:
+            self.__model = model.eval().half()
         self.__gpu_lock: Optional[asyncio.Lock] = None
         self.__frame_tasks_queue: Optional[asyncio.Queue[Optional[asyncio.Task[bytes]]]] = None
 
@@ -111,7 +114,9 @@ async def __write_output_frames(self, frames: AsyncIterator[bytes]) -> None:
     @retry.retry(RuntimeError, tries=10, delay=1)
     def __gpu_upscale(self, frame: torch.Tensor) -> torch.Tensor:
         with torch.no_grad():
-            frame_float = frame.cuda().permute(2, 0, 1).half() / 255
+            if torch.cuda.is_available():
+                frame = frame.cuda()
+            frame_float = frame.permute(2, 0, 1).half() / 255
             frame_upscaled_float = self.__model(frame_float.unsqueeze(0)).data.squeeze().clamp_(0, 1)
             return cast(torch.Tensor, (frame_upscaled_float * 255.0).round().byte().permute(1, 2, 0).cpu())