fix MNIST byte flipping (#7081)

* fix MNIST byte flipping

* add test

* move to utils

* remove lazy import

test/test_datasets_utils.py

@@ -7,7 +7,9 @@ import tarfile
 import zipfile
 
 import pytest
+import torch
 import torchvision.datasets.utils as utils
+from common_utils import assert_equal
 from torch._utils_internal import get_file_path_2
 from torchvision.datasets.folder import make_dataset
 from torchvision.datasets.utils import _COMPRESSED_FILE_OPENERS
@@ -215,6 +217,24 @@ class TestDatasetsUtils:
         pytest.raises(ValueError, utils.verify_str_arg, 0, ("a",), "arg")
         pytest.raises(ValueError, utils.verify_str_arg, "b", ("a",), "arg")
 
+    @pytest.mark.parametrize(
+        ("dtype", "actual_hex", "expected_hex"),
+        [
+            (torch.uint8, "01 23 45 67 89 AB CD EF", "01 23 45 67 89 AB CD EF"),
+            (torch.float16, "01 23 45 67 89 AB CD EF", "23 01 67 45 AB 89 EF CD"),
+            (torch.int32, "01 23 45 67 89 AB CD EF", "67 45 23 01 EF CD AB 89"),
+            (torch.float64, "01 23 45 67 89 AB CD EF", "EF CD AB 89 67 45 23 01"),
+        ],
+    )
+    def test_flip_byte_order(self, dtype, actual_hex, expected_hex):
+        def to_tensor(hex):
+            return torch.frombuffer(bytes.fromhex(hex), dtype=dtype)
+
+        assert_equal(
+            utils._flip_byte_order(to_tensor(actual_hex)),
+            to_tensor(expected_hex),
+        )
+
 
 @pytest.mark.parametrize(
     ("kwargs", "expected_error_msg"),

torchvision/datasets/mnist.py

@@ -12,7 +12,7 @@ import numpy as np
 import torch
 from PIL import Image
 
-from .utils import check_integrity, download_and_extract_archive, extract_archive, verify_str_arg
+from .utils import _flip_byte_order, check_integrity, download_and_extract_archive, extract_archive, verify_str_arg
 from .vision import VisionDataset
 
 
@@ -519,13 +519,12 @@ def read_sn3_pascalvincent_tensor(path: str, strict: bool = True) -> torch.Tenso
     torch_type = SN3_PASCALVINCENT_TYPEMAP[ty]
     s = [get_int(data[4 * (i + 1) : 4 * (i + 2)]) for i in range(nd)]
 
-    num_bytes_per_value = torch.iinfo(torch_type).bits // 8
-    # The MNIST format uses the big endian byte order. If the system uses little endian byte order by default,
-    # we need to reverse the bytes before we can read them with torch.frombuffer().
-    needs_byte_reversal = sys.byteorder == "little" and num_bytes_per_value > 1
     parsed = torch.frombuffer(bytearray(data), dtype=torch_type, offset=(4 * (nd + 1)))
-    if needs_byte_reversal:
-        parsed = parsed.flip(0)
+
+    # The MNIST format uses the big endian byte order, while `torch.frombuffer` uses whatever the system uses. In case
+    # that is little endian and the dtype has more than one byte, we need to flip them.
+    if sys.byteorder == "little" and parsed.element_size() > 1:
+        parsed = _flip_byte_order(parsed)
 
     assert parsed.shape[0] == np.prod(s) or not strict
     return parsed.view(*s)

torchvision/datasets/utils.py

@@ -520,3 +520,9 @@ def _read_pfm(file_name: str, slice_channels: int = 2) -> np.ndarray:
     data = np.flip(data, axis=1)  # flip on h dimension
     data = data[:slice_channels, :, :]
     return data.astype(np.float32)
+
+
+def _flip_byte_order(t: torch.Tensor) -> torch.Tensor:
+    return (
+        t.contiguous().view(torch.uint8).view(*t.shape, t.element_size()).flip(-1).view(*t.shape[:-1], -1).view(t.dtype)
+    )