Change image default dtype from float32 to uint8 (openvinotoolkit#1175)

- Currently, there is discrepancy between the return image data types:
`ImageFromBytes.data` (`np.float32`), `ImageFromNumpy.data`
(`np.float32`), and `ImageFromFile.data` (`np.uint8`).
- This makes the data loader based on the Arrow data format (using
`ImageFromBytes.data`) slower since the image preprocessing will be
conducted on the `np.float32` data (4x larger than `np.uint8`).
- This PR forces `np.uint8` data to be returned for all `Image` classes.

Signed-off-by: Kim, Vinnam <[email protected]>

CHANGELOG.md

@@ -17,6 +17,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Enhancements
 - Enhance Datumaro data format stream importer performance
   (<https://github.com/openvinotoolkit/datumaro/pull/1153>)
+- Change image default dtype from float32 to uint8
+  (<https://github.com/openvinotoolkit/datumaro/pull/1175>)
 
 ### Bug fixes
 - Fix errata in the voc document. Color values in the labelmap.txt should be separated by commas, not colons.

src/datumaro/components/media.py

@@ -309,7 +309,7 @@ def __init__(
 
     @property
     def data(self) -> Optional[np.ndarray]:
-        """Image data in BGRA HWC [0; 255] (float) format"""
+        """Image data in BGRA HWC [0; 255] (uint8) format"""
 
         if not self.has_data:
             return None
@@ -375,12 +375,12 @@ def __init__(
 
     @property
     def data(self) -> Optional[np.ndarray]:
-        """Image data in BGRA HWC [0; 255] (float) format"""
+        """Image data in BGRA HWC [0; 255] (uint8) format"""
 
         data = super().data
 
-        if isinstance(data, np.ndarray):
-            data = data.astype(np.float32)
+        if isinstance(data, np.ndarray) and data.dtype != np.uint8:
+            data = np.clip(data, 0.0, 255.0).astype(np.uint8)
         if self._size is None and data is not None:
             if not 2 <= data.ndim <= 3:
                 raise MediaShapeError("An image should have 2 (gray) or 3 (bgra) dims.")
@@ -420,14 +420,12 @@ def _guess_ext(cls, data: bytes) -> Optional[str]:
 
     @property
     def data(self) -> Optional[np.ndarray]:
-        """Image data in BGRA HWC [0; 255] (float) format"""
+        """Image data in BGRA HWC [0; 255] (uint8) format"""
 
         data = super().data
 
         if isinstance(data, bytes):
-            data = decode_image(data)
-        if isinstance(data, np.ndarray):
-            data = data.astype(np.float32)
+            data = decode_image(data, dtype=np.uint8)
         if self._size is None and data is not None:
             if not 2 <= data.ndim <= 3:
                 raise MediaShapeError("An image should have 2 (gray) or 3 (bgra) dims.")

src/datumaro/plugins/framework_converter.py

@@ -112,9 +112,6 @@ def __init__(
         def __getitem__(self, idx):
             image, label = self._gen_item(idx)
 
-            if image.dtype == np.uint8 or image.max() > 1:
-                image = image.astype(np.float32) / 255
-
             if len(image.shape) == 2:
                 image = np.expand_dims(image, axis=-1)
 

tests/unit/operations/test_statistics.py

@@ -19,8 +19,9 @@
 
 @pytest.fixture
 def fxt_image_dataset_expected_mean_std():
+    np.random.seed(3003)
     expected_mean = [100, 50, 150]
-    expected_std = [20, 50, 10]
+    expected_std = [2, 1, 3]
 
     return expected_mean, expected_std
 
@@ -90,9 +91,9 @@ def test_image_stats(
         actual_std = actual["subsets"]["default"]["image std"][::-1]
 
         for em, am in zip(expected_mean, actual_mean):
-            assert am == pytest.approx(em, 1e-2)
+            assert am == pytest.approx(em, 5e-1)
         for estd, astd in zip(expected_std, actual_std):
-            assert astd == pytest.approx(estd, 1e-2)
+            assert astd == pytest.approx(estd, 1e-1)
 
     @mark_requirement(Requirements.DATUM_BUG_873)
     def test_invalid_media_type(

tests/unit/test_framework_converter.py

@@ -296,7 +296,8 @@ def test_can_convert_torch_framework(
                 label = np.sum(masks, axis=0, dtype=np.uint8)
 
             if fxt_convert_kwargs.get("transform", None):
-                assert np.array_equal(image, dm_torch_item[0].reshape(5, 5, 3).numpy())
+                actual = dm_torch_item[0].permute(1, 2, 0).mul(255.0).to(torch.uint8).numpy()
+                assert np.array_equal(image, actual)
             else:
                 assert np.array_equal(image, dm_torch_item[0])
 

tests/unit/test_ops.py

@@ -44,8 +44,9 @@
 class TestOperations(TestCase):
     @mark_requirement(Requirements.DATUM_GENERAL_REQ)
     def test_mean_std(self):
+        np.random.seed(3000)
         expected_mean = [100, 50, 150]
-        expected_std = [20, 50, 10]
+        expected_std = [2, 1, 3]
 
         dataset = Dataset.from_iterable(
             [
@@ -62,9 +63,9 @@ def test_mean_std(self):
         actual_mean, actual_std = mean_std(dataset)
 
         for em, am in zip(expected_mean, actual_mean):
-            self.assertAlmostEqual(em, am, places=0)
+            assert np.allclose(em, am, atol=0.6)
         for estd, astd in zip(expected_std, actual_std):
-            self.assertAlmostEqual(estd, astd, places=0)
+            assert np.allclose(estd, astd, atol=0.1)
 
     @mark_requirement(Requirements.DATUM_GENERAL_REQ)
     def test_stats(self):