python: test the exact bfloat16 / float8 dtype conversions both ways

Signed-off-by: Abhishek Gola <abhishek.gola@bigvisionllc.com>

modules/python/test/test_ml_dtypes.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python
+from __future__ import print_function
+
+import numpy as np
+import cv2 as cv
+try:
+    import ml_dtypes
+except ImportError:
+    ml_dtypes = None
+
+from tests_common import NewOpenCVTests
+
+# Python does not publish these depths yet: CV_16BF, CV_8F_E4M3FN, CV_8F_E4M3FNUZ.
+CV_16BF, CV_8F_E4M3FN, CV_8F_E4M3FNUZ = 8, 13, 14
+
+
+class ml_dtypes_test(NewOpenCVTests):
+
+    def setUp(self):
+        super(ml_dtypes_test, self).setUp()
+        if ml_dtypes is None:
+            self.skipTest("ml_dtypes is not installed")
+        # (numpy dtype, OpenCV depth, typeToString prefix, relative half-ULP)
+        self.cases = [
+            (ml_dtypes.bfloat16, CV_16BF, "CV_16BF", 2.0 ** -8),
+            (ml_dtypes.float8_e4m3fn, CV_8F_E4M3FN, "CV_8F", 2.0 ** -4),
+            (ml_dtypes.float8_e4m3fnuz, CV_8F_E4M3FNUZ, "CV_8FNUZ", 2.0 ** -4),
+        ]
+
+    def assertSameBits(self, actual, expected):
+        self.assertEqual(actual.dtype, expected.dtype)
+        self.assertEqual(actual.shape, expected.shape)
+        self.assertTrue(np.array_equal(np.ascontiguousarray(actual).view(np.uint8),
+                                       np.ascontiguousarray(expected).view(np.uint8)))
+
+    def test_numpy_to_mat_keeps_exact_type(self):
+        for dt, _, name, _ in self.cases:
+            a = np.arange(6, dtype=np.float32).reshape(2, 3).astype(dt)
+            self.assertIn("type(-1)=%sC1" % name, cv.utils.dumpInputArray(a))
+            self.assertIn("type(-1)=%sC3" % name, cv.utils.dumpInputArray(np.stack([a, a, a], -1)))
+            self.assertIn("type(-1)=%sC1" % name, cv.utils.dumpInputArray(np.array(1.5, dtype=dt)))
+
+    def test_roundtrip_through_numpy_allocator(self):
+        src = np.array([[0.5, -1.25, 3.0, 7.0], [448.0, -448.0, 0.0, -0.5]], dtype=np.float32)
+        for dt, _, _, _ in self.cases:
+            a = src.astype(dt)
+            self.assertSameBits(cv.flip(a, 1), a[:, ::-1])
+            self.assertSameBits(cv.transpose(a), a.T)
+            strided = a[:, ::2]
+            self.assertSameBits(cv.flip(strided, 0), strided[::-1])
+
+    def test_mat_allocated_in_cpp_to_numpy(self):
+        ref = cv.utils.generateVectorOfMat(2, 3, 4, cv.CV_32F)
+        for dt, depth, _, rtol in self.cases:
+            mats = cv.utils.generateVectorOfMat(2, 3, 4, depth)
+            self.assertEqual(len(mats), len(ref))
+            for m, r in zip(mats, ref):
+                self.assertEqual(m.dtype, dt)
+                self.assertEqual(m.shape, (3, 4))
+                np.testing.assert_allclose(m.astype(np.float32), r, rtol=rtol, atol=2.0 ** -9)
+
+
+if __name__ == '__main__':
+    NewOpenCVTests.bootstrap()