[fbsync] [proto] Performance improvements for equalize op (#6757)

Summary:
* [proto] Performance improvements for equalize op

* Added tests

Reviewed By: NicolasHug

Differential Revision: D40427459

fbshipit-source-id: 8cfba7a345b87fb56b7edcf58b7f9c7d526be813

Author: Yosua Michael Maranatha

test/test_prototype_transforms_functional.py

@@ -1037,3 +1037,14 @@ def test_to_image_pil(inpt, mode):
     assert isinstance(output, PIL.Image.Image)
 
     assert np.asarray(inpt).sum() == np.asarray(output).sum()
+
+
+def test_equalize_image_tensor_edge_cases():
+    inpt = torch.zeros(3, 200, 200, dtype=torch.uint8)
+    output = F.equalize_image_tensor(inpt)
+    torch.testing.assert_close(inpt, output)
+
+    inpt = torch.zeros(5, 3, 200, 200, dtype=torch.uint8)
+    inpt[..., 100:, 100:] = 1
+    output = F.equalize_image_tensor(inpt)
+    assert output.unique().tolist() == [0, 255]

torchvision/prototype/transforms/functional/_color.py

@@ -183,28 +183,37 @@ def autocontrast(inpt: features.InputTypeJIT) -> features.InputTypeJIT:
         return autocontrast_image_pil(inpt)
 
 
-def _scale_channel(img_chan: torch.Tensor) -> torch.Tensor:
-    # TODO: we should expect bincount to always be faster than histc, but this
-    # isn't always the case. Once
-    # https://github.com/pytorch/pytorch/issues/53194 is fixed, remove the if
-    # block and only use bincount.
-    if img_chan.is_cuda:
-        hist = torch.histc(img_chan.to(torch.float32), bins=256, min=0, max=255)
-    else:
-        hist = torch.bincount(img_chan.view(-1), minlength=256)
-
-    nonzero_hist = hist[hist != 0]
-    step = torch.div(nonzero_hist[:-1].sum(), 255, rounding_mode="floor")
-    if step == 0:
-        return img_chan
-
-    lut = torch.div(torch.cumsum(hist, 0) + torch.div(step, 2, rounding_mode="floor"), step, rounding_mode="floor")
-    # Doing inplace clamp and converting lut to uint8 improves perfs
-    lut.clamp_(0, 255)
-    lut = lut.to(torch.uint8)
-    lut = torch.nn.functional.pad(lut[:-1], [1, 0])
-
-    return lut[img_chan.to(torch.int64)]
+def _equalize_image_tensor_vec(img: torch.Tensor) -> torch.Tensor:
+    # input img shape should be [N, H, W]
+    shape = img.shape
+    # Compute image histogram:
+    flat_img = img.flatten(start_dim=1).to(torch.long)  # -> [N, H * W]
+    hist = flat_img.new_zeros(shape[0], 256)
+    hist.scatter_add_(dim=1, index=flat_img, src=flat_img.new_ones(1).expand_as(flat_img))
+
+    # Compute image cdf
+    chist = hist.cumsum_(dim=1)
+    # Compute steps, where step per channel is nonzero_hist[:-1].sum() // 255
+    # Trick: nonzero_hist[:-1].sum() == chist[idx - 1], where idx = chist.argmax()
+    idx = chist.argmax(dim=1).sub_(1)
+    # If histogram is degenerate (hist of zero image), index is -1
+    neg_idx_mask = idx < 0
+    idx.clamp_(min=0)
+    step = chist.gather(dim=1, index=idx.unsqueeze(1))
+    step[neg_idx_mask] = 0
+    step.div_(255, rounding_mode="floor")
+
+    # Compute batched Look-up-table:
+    # Necessary to avoid an integer division by zero, which raises
+    clamped_step = step.clamp(min=1)
+    chist.add_(torch.div(step, 2, rounding_mode="floor")).div_(clamped_step, rounding_mode="floor").clamp_(0, 255)
+    lut = chist.to(torch.uint8)  # [N, 256]
+
+    # Pad lut with zeros
+    zeros = lut.new_zeros((1, 1)).expand(shape[0], 1)
+    lut = torch.cat([zeros, lut[:, :-1]], dim=1)
+
+    return torch.where((step == 0).unsqueeze(-1), img, lut.gather(dim=1, index=flat_img).view_as(img))
 
 
 def equalize_image_tensor(image: torch.Tensor) -> torch.Tensor:
@@ -217,10 +226,8 @@ def equalize_image_tensor(image: torch.Tensor) -> torch.Tensor:
 
     if image.numel() == 0:
         return image
-    elif image.ndim == 2:
-        return _scale_channel(image)
-    else:
-        return torch.stack([_scale_channel(x) for x in image.view(-1, height, width)]).view(image.shape)
+
+    return _equalize_image_tensor_vec(image.view(-1, height, width)).view(image.shape)
 
 
 equalize_image_pil = _FP.equalize