Refactor upscale_2 helper out of ScuNET/SwinIR; make sure devices are right

1 files changed, 69 insertions, 20 deletions

diff --git a/modules/upscaler_utils.py b/modules/upscaler_utils.py
index 9379f512..e4c63f09 100644
--- a/modules/upscaler_utils.py
+++ b/modules/upscaler_utils.py
@@ -11,23 +11,40 @@ from modules import images, shared, torch_utils
 logger = logging.getLogger(__name__)
 
 
-def upscale_without_tiling(model, img: Image.Image):
-    img = np.array(img)
-    img = img[:, :, ::-1]
-    img = np.ascontiguousarray(np.transpose(img, (2, 0, 1))) / 255
-    img = torch.from_numpy(img).float()
-
+def pil_image_to_torch_bgr(img: Image.Image) -> torch.Tensor:
+    img = np.array(img.convert("RGB"))
+    img = img[:, :, ::-1]  # flip RGB to BGR
+    img = np.transpose(img, (2, 0, 1))  # HWC to CHW
+    img = np.ascontiguousarray(img) / 255  # Rescale to [0, 1]
+    return torch.from_numpy(img)
+
+
+def torch_bgr_to_pil_image(tensor: torch.Tensor) -> Image.Image:
+    if tensor.ndim == 4:
+        # If we're given a tensor with a batch dimension, squeeze it out
+        # (but only if it's a batch of size 1).
+        if tensor.shape[0] != 1:
+            raise ValueError(f"{tensor.shape} does not describe a BCHW tensor")
+        tensor = tensor.squeeze(0)
+    assert tensor.ndim == 3, f"{tensor.shape} does not describe a CHW tensor"
+    # TODO: is `tensor.float().cpu()...numpy()` the most efficient idiom?
+    arr = tensor.float().cpu().clamp_(0, 1).numpy()  # clamp
+    arr = 255.0 * np.moveaxis(arr, 0, 2)  # CHW to HWC, rescale
+    arr = arr.astype(np.uint8)
+    arr = arr[:, :, ::-1]  # flip BGR to RGB
+    return Image.fromarray(arr, "RGB")
+
+
+def upscale_pil_patch(model, img: Image.Image) -> Image.Image:
+    """
+    Upscale a given PIL image using the given model.
+    """
     param = torch_utils.get_param(model)
-    img = img.unsqueeze(0).to(device=param.device, dtype=param.dtype)
 
     with torch.no_grad():
-        output = model(img)
-
-    output = output.squeeze().float().cpu().clamp_(0, 1).numpy()
-    output = 255. * np.moveaxis(output, 0, 2)
-    output = output.astype(np.uint8)
-    output = output[:, :, ::-1]
-    return Image.fromarray(output, 'RGB')
+        tensor = pil_image_to_torch_bgr(img).unsqueeze(0)  # add batch dimension
+        tensor = tensor.to(device=param.device, dtype=param.dtype)
+        return torch_bgr_to_pil_image(model(tensor))
 
 
 def upscale_with_model(
@@ -40,7 +57,7 @@ def upscale_with_model(
 ) -> Image.Image:
     if tile_size <= 0:
         logger.debug("Upscaling %s without tiling", img)
-        output = upscale_without_tiling(model, img)
+        output = upscale_pil_patch(model, img)
         logger.debug("=> %s", output)
         return output
 
@@ -52,7 +69,7 @@ def upscale_with_model(
             newrow = []
             for x, w, tile in row:
                 logger.debug("Tile (%d, %d) %s...", x, y, tile)
-                output = upscale_without_tiling(model, tile)
+                output = upscale_pil_patch(model, tile)
                 scale_factor = output.width // tile.width
                 logger.debug("=> %s (scale factor %s)", output, scale_factor)
                 newrow.append([x * scale_factor, w * scale_factor, output])
@@ -71,19 +88,22 @@ def upscale_with_model(
 
 
 def tiled_upscale_2(
-    img,
+    img: torch.Tensor,
     model,
     *,
     tile_size: int,
     tile_overlap: int,
     scale: int,
-    device,
     desc="Tiled upscale",
 ):
     # Alternative implementation of `upscale_with_model` originally used by
     # SwinIR and ScuNET.  It differs from `upscale_with_model` in that tiling and
     # weighting is done in PyTorch space, as opposed to `images.Grid` doing it in
     # Pillow space without weighting.
+
+    # Grab the device the model is on, and use it.
+    device = torch_utils.get_param(model).device
+
     b, c, h, w = img.size()
     tile_size = min(tile_size, h, w)
 
@@ -100,7 +120,8 @@ def tiled_upscale_2(
         h * scale,
         w * scale,
         device=device,
-    ).type_as(img)
+        dtype=img.dtype,
+    )
     weights = torch.zeros_like(result)
     logger.debug("Upscaling %s to %s with tiles", img.shape, result.shape)
     with tqdm.tqdm(total=len(h_idx_list) * len(w_idx_list), desc=desc, disable=not shared.opts.enable_upscale_progressbar) as pbar:
@@ -112,11 +133,13 @@ def tiled_upscale_2(
                 if shared.state.interrupted or shared.state.skipped:
                     break
 
+                # Only move this patch to the device if it's not already there.
                 in_patch = img[
                     ...,
                     h_idx : h_idx + tile_size,
                     w_idx : w_idx + tile_size,
-                ]
+                ].to(device=device)
+
                 out_patch = model(in_patch)
 
                 result[
@@ -138,3 +161,29 @@ def tiled_upscale_2(
     output = result.div_(weights)
 
     return output
+
+
+def upscale_2(
+    img: Image.Image,
+    model,
+    *,
+    tile_size: int,
+    tile_overlap: int,
+    scale: int,
+    desc: str,
+):
+    """
+    Convenience wrapper around `tiled_upscale_2` that handles PIL images.
+    """
+    tensor = pil_image_to_torch_bgr(img).float().unsqueeze(0)  # add batch dimension
+
+    with torch.no_grad():
+        output = tiled_upscale_2(
+            tensor,
+            model,
+            tile_size=tile_size,
+            tile_overlap=tile_overlap,
+            scale=scale,
+            desc=desc,
+        )
+    return torch_bgr_to_pil_image(output)