chore (test suite): skip or xfail random failures (kornia#2930)

Author: johnnv1

tests/augmentation/test_augmentation.py

@@ -1,4 +1,5 @@
 import os
+import sys
 from typing import Any, Dict, Optional, Tuple, Type
 from unittest.mock import patch
 
@@ -59,6 +60,7 @@
 from kornia.constants import Resample, pi
 from kornia.geometry import transform_points
 from kornia.utils import create_meshgrid
+from kornia.utils._compat import torch_version
 from kornia.utils.helpers import _torch_inverse_cast
 
 from testing.augmentation.datasets import DummyMPDataset
@@ -4459,6 +4461,10 @@ def test_mask_transform(self, device, dtype):
         )
 
     @pytest.mark.parametrize("batch_prob", [[True, True], [False, True], [False, False]])
+    @pytest.mark.skipif(
+        torch_version() in {"1.11.0", "1.12.1"} and sys.version_info.minor == 10,
+        reason="failing because no gaussian mean",
+    )
     def test_apply(self, batch_prob, device, dtype):
         torch.manual_seed(0)
 

tests/contrib/models/test_tiny_vit.py

@@ -1,3 +1,5 @@
+import sys
+
 import pytest
 import torch
 
@@ -35,6 +37,7 @@ def test_gradcheck(self): ...
     @pytest.mark.skip("not implemented")
     def test_module(self): ...
 
+    @pytest.mark.skipif(sys.version_info.major == 3 and sys.version_info.minor == 8, reason="not working for py3.8")
     def test_dynamo(self, device, dtype, torch_optimizer):
         op = TinyViT().to(device=device, dtype=dtype)
         img = torch.rand(1, 3, op.img_size, op.img_size, device=device, dtype=dtype)

tests/contrib/test_prompter.py

@@ -3,6 +3,7 @@
 
 from kornia.contrib.models.sam import SamConfig
 from kornia.contrib.visual_prompter import VisualPrompter
+from kornia.utils._compat import torch_version
 
 from testing.base import BaseTester
 
@@ -84,6 +85,7 @@ def test_gradcheck(self, device): ...
     @pytest.mark.skip(reason="Unnecessary test")
     def test_module(self): ...
 
+    @pytest.mark.skipif(torch_version() in {"2.1.2", "2.0.1"}, reason="Not working well")
     def test_dynamo(self, device, torch_optimizer):
         dtype = torch.float32
         batch_size = 1

tests/filters/test_canny.py

@@ -300,7 +300,7 @@ def test_module(self, device, dtype):
     @pytest.mark.skipif(torch_version() in {"2.0.0", "2.0.1"}, reason="Not working on 2.0")
     def test_dynamo(self, batch_size, kernel_size, device, dtype, torch_optimizer):
         if (
-            torch_version() in {"2.1.1", "2.1.2", "2.2.0"}
+            torch_version() in {"2.1.1", "2.1.2", "2.2.2", "2.3.1"}
             and dtype == torch.float64
             and (isinstance(kernel_size, int) or kernel_size[0] == kernel_size[1])
         ):

tests/filters/test_sobel.py

@@ -2,6 +2,7 @@
 import torch
 
 from kornia.filters import Sobel, SpatialGradient, SpatialGradient3d, sobel, spatial_gradient, spatial_gradient3d
+from kornia.utils._compat import torch_version
 
 from testing.base import BaseTester
 
@@ -242,6 +243,7 @@ def test_module(self, device, dtype):
     @pytest.mark.parametrize("mode", ["sobel", "diff"])
     @pytest.mark.parametrize("order", [1, 2])
     @pytest.mark.parametrize("batch_size", [1, 2])
+    @pytest.mark.xfail(torch_version() in {"2.0.1"}, reason="random failing")
     def test_dynamo(self, batch_size, order, mode, device, dtype, torch_optimizer):
         inpt = torch.ones(batch_size, 3, 10, 10, device=device, dtype=dtype)
         if order == 1 and dtype == torch.float64:

tests/geometry/epipolar/test_fundamental.py

@@ -247,6 +247,7 @@ def test_synthetic_sampson_7point(self, device, dtype):
                 error = epi.sampson_epipolar_distance(x1, x2, F)
                 self.assert_close(error, torch.zeros((F.shape[0], 7), device=device, dtype=dtype), atol=1e-4, rtol=1e-4)
 
+    @pytest.mark.xfail()
     def test_epipolar_constraint_7point(self, device, dtype):
         scene: Dict[str, torch.Tensor] = generate_two_view_random_scene(device, dtype)
         x1 = scene["x1"][:, :7, :]

tests/geometry/test_conversions.py

@@ -1,3 +1,4 @@
+import sys
 from functools import partial
 
 import numpy as np
@@ -19,6 +20,7 @@
     worldtocam_to_camtoworld_Rt,
 )
 from kornia.geometry.quaternion import Quaternion
+from kornia.utils._compat import torch_version
 from kornia.utils.misc import eye_like
 
 from testing.base import BaseTester, assert_close
@@ -1109,6 +1111,10 @@ def test_gradcheck(self, device):
         q = Quaternion.random(batch_size=1).to(device, torch.float64)
         self.gradcheck(euler_from_quaternion, (q.w, q.x, q.y, q.z))
 
+    @pytest.mark.skipif(
+        torch_version() in {"2.0.1", "2.1.2", "2.2.2", "2.3.1"} and sys.version_info.minor == 8,
+        reason="Not working on 2.0",
+    )
     def test_dynamo(self, device, dtype, torch_optimizer):
         q = Quaternion.random(batch_size=1)
         q = q.to(device, dtype)

tests/geometry/transform/test_image_registrator.py

@@ -64,6 +64,7 @@ def test_smoke(self, device, dtype, model_type):
         assert ir is not None
 
     @pytest.mark.slow
+    @pytest.mark.xfail(torch_version() in {"2.0.0", "2.0.1", "2.1.2", "2.2.2"}, reason="failing at some 2.x torch")
     def test_registration_toy(self, device, dtype):
         ch, height, width = 3, 16, 18
         homography = torch.eye(3, device=device, dtype=dtype)[None]

tests/losses/test_hd.py

@@ -46,6 +46,8 @@ def test_exception_3d(self):
         assert "Invalid target value" in str(errinf)
 
     def test_numeric(self, device, dtype):
+        if dtype == torch.float64:
+            pytest.xfail("Sometimes failing on float64")
         num_classes = 3
         shape = (50, 50)
         hd = kornia.losses.HausdorffERLoss