bugfix aspect ratio sampling in transforms.RandomErasing (#3344)

* aspect ratio must be a sampling from log scale.
reference from: https://github.com/pytorch/vision/blob/8317295c1d272e0ba7b2ce31e3fd2c048235fc73/torchvision/transforms/transforms.py#L833-L836



* add random erasing unittest code

* Increased threshold for difference in sampling rate

* move outside of the loop

* log_ratio move outside of the loop. RandomResizedCrop also
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

test/test_transforms.py

@@ -1955,6 +1955,25 @@ class Tester(unittest.TestCase):
                     img = transform(img)
                 transform.__repr__()
 
+    def test_random_erasing(self):
+        img = torch.ones(3, 128, 128)
+
+        t = transforms.RandomErasing(scale=(0.1, 0.1), ratio=(1. / 3., 3. / 1.))
+        y, x, h, w, v = t.get_params(img, t.scale, t.ratio, [t.value, ])
+        aspect_ratio = h / w
+        self.assertTrue(aspect_ratio > 1. / 3. and aspect_ratio < 3. / 1.)
+
+        aspect_ratios = []
+        random.seed(42)
+        trial = 1000
+        for _ in range(trial):
+            y, x, h, w, v = t.get_params(img, t.scale, t.ratio, [t.value, ])
+            aspect_ratios.append(h / w)
+
+        count_bigger_then_ones = len([1 for aspect_ratio in aspect_ratios if aspect_ratio > 1])
+        count_smaller_then_ones = len([1 for aspect_ratio in aspect_ratios if aspect_ratio < 1])
+        self.assertAlmostEqual(count_bigger_then_ones / trial, count_smaller_then_ones / trial, 1)
+
 
 if __name__ == '__main__':
     unittest.main()

torchvision/transforms/transforms.py

@@ -828,9 +828,9 @@ class RandomResizedCrop(torch.nn.Module):
         width, height = F._get_image_size(img)
         area = height * width
 
+        log_ratio = torch.log(torch.tensor(ratio))
         for _ in range(10):
             target_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item()
-            log_ratio = torch.log(torch.tensor(ratio))
             aspect_ratio = torch.exp(
                 torch.empty(1).uniform_(log_ratio[0], log_ratio[1])
             ).item()
@@ -1576,9 +1576,12 @@ class RandomErasing(torch.nn.Module):
         img_c, img_h, img_w = img.shape[-3], img.shape[-2], img.shape[-1]
         area = img_h * img_w
 
+        log_ratio = torch.log(torch.tensor(ratio))
         for _ in range(10):
             erase_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item()
-            aspect_ratio = torch.empty(1).uniform_(ratio[0], ratio[1]).item()
+            aspect_ratio = torch.exp(
+                torch.empty(1).uniform_(log_ratio[0], log_ratio[1])
+            ).item()
 
             h = int(round(math.sqrt(erase_area * aspect_ratio)))
             w = int(round(math.sqrt(erase_area / aspect_ratio)))