Fixed reparametrization for shear X/Y in autoaugment ops

test/common_utils.py

@@ -187,7 +187,7 @@ def _assert_approx_equal_tensor_to_pil(
     tensor = tensor.to(torch.float)
     pil_tensor = pil_tensor.to(torch.float)
     err = getattr(torch, agg_method)(torch.abs(tensor - pil_tensor)).item()
-    assert err < tol
+    assert err < tol, f"{err} vs {tol}"
 
 
 def _test_fn_on_batch(batch_tensors, fn, scripted_fn_atol=1e-8, **fn_kwargs):

test/test_transforms_tensor.py

@@ -14,9 +14,11 @@
     cpu_and_gpu,
     assert_equal,
 )
+from PIL import Image
 from torchvision import transforms as T
 from torchvision.transforms import InterpolationMode
 from torchvision.transforms import functional as F
+from torchvision.transforms.autoaugment import _apply_op
 
 NEAREST, BILINEAR, BICUBIC = InterpolationMode.NEAREST, InterpolationMode.BILINEAR, InterpolationMode.BICUBIC
 
@@ -725,6 +727,48 @@ def test_autoaugment_save(augmentation, tmpdir):
     s_transform.save(os.path.join(tmpdir, "t_autoaugment.pt"))
 
 
+@pytest.mark.parametrize("interpolation", [F.InterpolationMode.NEAREST, F.InterpolationMode.BILINEAR])
+@pytest.mark.parametrize("mode", ["X", "Y"])
+def test_autoaugment__op_apply_shear(interpolation, mode):
+    # We check that torchvision's implementation of shear is equivalent
+    # to official CIFAR10 autoaugment implementation:
+    # https://github.com/tensorflow/models/blob/885fda091c46c59d6c7bb5c7e760935eacc229da/research/autoaugment/augmentation_transforms.py#L273-L290
+    image_size = 32
+
+    def shear(pil_img, level, mode, resample):
+        if mode == "X":
+            matrix = (1, level, 0, 0, 1, 0)
+        elif mode == "Y":
+            matrix = (1, 0, 0, level, 1, 0)
+        return pil_img.transform((image_size, image_size), Image.AFFINE, matrix, resample=resample)
+
+    t_img, pil_img = _create_data(image_size, image_size)
+
+    resample_pil = {
+        F.InterpolationMode.NEAREST: Image.NEAREST,
+        F.InterpolationMode.BILINEAR: Image.BILINEAR,
+    }[interpolation]
+
+    level = 0.3
+    expected_out = shear(pil_img, level, mode=mode, resample=resample_pil)
+
+    # Check pil output vs expected pil
+    out = _apply_op(pil_img, op_name=f"Shear{mode}", magnitude=level, interpolation=interpolation, fill=0)
+    assert out == expected_out
+
+    if interpolation == F.InterpolationMode.BILINEAR:
+        # We skip bilinear mode for tensors as
+        # affine transformation results are not exactly the same
+        # between tensors and pil images
+        # MAE as around 1.40
+        # Max Abs error can be 163 or 170
+        return
+
+    # Check tensor output vs expected pil
+    out = _apply_op(t_img, op_name=f"Shear{mode}", magnitude=level, interpolation=interpolation, fill=0)
+    _assert_approx_equal_tensor_to_pil(out, expected_out)
+
+
 @pytest.mark.parametrize("device", cpu_and_gpu())
 @pytest.mark.parametrize(
     "config",

torchvision/transforms/autoaugment.py

@@ -14,23 +14,31 @@ def _apply_op(
     img: Tensor, op_name: str, magnitude: float, interpolation: InterpolationMode, fill: Optional[List[float]]
 ):
     if op_name == "ShearX":
+        # magnitude should be arctan(magnitude)
+        # official autoaug: (1, level, 0, 0, 1, 0)
+        # https://github.com/tensorflow/models/blob/dd02069717128186b88afa8d857ce57d17957f03/research/autoaugment/augmentation_transforms.py#L290
+        # compared to
+        # torchvision:      (1, tan(level), 0, 0, 1, 0)
+        # https://github.com/pytorch/vision/blob/0c2373d0bba3499e95776e7936e207d8a1676e65/torchvision/transforms/functional.py#L976
         img = F.affine(
             img,
             angle=0.0,
             translate=[0, 0],
             scale=1.0,
-            shear=[math.degrees(magnitude), 0.0],
+            shear=[math.degrees(math.atan(magnitude)), 0.0],
             interpolation=interpolation,
             fill=fill,
             center=[0, 0],
         )
     elif op_name == "ShearY":
+        # magnitude should be arctan(magnitude)
+        # See above
         img = F.affine(
             img,
             angle=0.0,
             translate=[0, 0],
             scale=1.0,
-            shear=[0.0, math.degrees(magnitude)],
+            shear=[0.0, math.degrees(math.atan(magnitude))],
             interpolation=interpolation,
             fill=fill,
             center=[0, 0],