[fbsync] Change default of antialias parameter from None to 'warn' (#7160)

Summary:

Reviewed By: vmoens

Differential Revision: D44416275

fbshipit-source-id: 916691a68545d0b487d9ac20b4a8f42ec42315b6

Co-authored-by: Philip Meier <[email protected]>
Co-authored-by: vfdev <[email protected]>

Author: 3 people

gallery/plot_optical_flow.py

@@ -81,9 +81,10 @@ def plot(imgs, **imshow_kwargs):
 
 #########################
 # The RAFT model accepts RGB images. We first get the frames from
-# :func:`~torchvision.io.read_video` and resize them to ensure their
-# dimensions are divisible by 8. Then we use the transforms bundled into the
-# weights in order to preprocess the input and rescale its values to the
+# :func:`~torchvision.io.read_video` and resize them to ensure their dimensions
+# are divisible by 8. Note that we explicitly use ``antialias=False``, because
+# this is how those models were trained. Then we use the transforms bundled into
+# the weights in order to preprocess the input and rescale its values to the
 # required ``[-1, 1]`` interval.
 
 from torchvision.models.optical_flow import Raft_Large_Weights
@@ -93,8 +94,8 @@ def plot(imgs, **imshow_kwargs):
 
 
 def preprocess(img1_batch, img2_batch):
-    img1_batch = F.resize(img1_batch, size=[520, 960])
-    img2_batch = F.resize(img2_batch, size=[520, 960])
+    img1_batch = F.resize(img1_batch, size=[520, 960], antialias=False)
+    img2_batch = F.resize(img2_batch, size=[520, 960], antialias=False)
     return transforms(img1_batch, img2_batch)
 
 

references/depth/stereo/transforms.py

@@ -455,7 +455,11 @@ def forward(
         INTERP_MODE = self._interpolation_mode_strategy()
 
         for img in images:
-            resized_images += (F.resize(img, self.resize_size, interpolation=INTERP_MODE),)
+            # We hard-code antialias=False to preserve results after we changed
+            # its default from None to True (see
+            # https://github.com/pytorch/vision/pull/7160)
+            # TODO: we could re-train the stereo models with antialias=True?
+            resized_images += (F.resize(img, self.resize_size, interpolation=INTERP_MODE, antialias=False),)
 
         for dsp in disparities:
             if dsp is not None:

references/optical_flow/transforms.py

@@ -196,8 +196,12 @@ def forward(self, img1, img2, flow, valid_flow_mask):
 
         if torch.rand(1).item() < self.resize_prob:
             # rescale the images
-            img1 = F.resize(img1, size=(new_h, new_w))
-            img2 = F.resize(img2, size=(new_h, new_w))
+            # We hard-code antialias=False to preserve results after we changed
+            # its default from None to True (see
+            # https://github.com/pytorch/vision/pull/7160)
+            # TODO: we could re-train the OF models with antialias=True?
+            img1 = F.resize(img1, size=(new_h, new_w), antialias=False)
+            img2 = F.resize(img2, size=(new_h, new_w), antialias=False)
             if valid_flow_mask is None:
                 flow = F.resize(flow, size=(new_h, new_w))
                 flow = flow * torch.tensor([scale_x, scale_y])[:, None, None]

references/video_classification/presets.py

@@ -15,7 +15,11 @@ def __init__(
     ):
         trans = [
             transforms.ConvertImageDtype(torch.float32),
-            transforms.Resize(resize_size),
+            # We hard-code antialias=False to preserve results after we changed
+            # its default from None to True (see
+            # https://github.com/pytorch/vision/pull/7160)
+            # TODO: we could re-train the video models with antialias=True?
+            transforms.Resize(resize_size, antialias=False),
         ]
         if hflip_prob > 0:
             trans.append(transforms.RandomHorizontalFlip(hflip_prob))
@@ -31,7 +35,11 @@ def __init__(self, *, crop_size, resize_size, mean=(0.43216, 0.394666, 0.37645),
         self.transforms = transforms.Compose(
             [
                 transforms.ConvertImageDtype(torch.float32),
-                transforms.Resize(resize_size),
+                # We hard-code antialias=False to preserve results after we changed
+                # its default from None to True (see
+                # https://github.com/pytorch/vision/pull/7160)
+                # TODO: we could re-train the video models with antialias=True?
+                transforms.Resize(resize_size, antialias=False),
                 transforms.Normalize(mean=mean, std=std),
                 transforms.CenterCrop(crop_size),
                 ConvertBCHWtoCBHW(),

test/test_functional_tensor.py

@@ -2,6 +2,7 @@
 import itertools
 import math
 import os
+import warnings
 from functools import partial
 from typing import Sequence
 
@@ -483,8 +484,8 @@ def test_resize(device, dt, size, max_size, interpolation):
         tensor = tensor.to(dt)
         batch_tensors = batch_tensors.to(dt)
 
-    resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, max_size=max_size)
-    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation, max_size=max_size)
+    resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, max_size=max_size, antialias=True)
+    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation, max_size=max_size, antialias=True)
 
     assert resized_tensor.size()[1:] == resized_pil_img.size[::-1]
 
@@ -509,10 +510,12 @@ def test_resize(device, dt, size, max_size, interpolation):
     else:
         script_size = size
 
-    resize_result = script_fn(tensor, size=script_size, interpolation=interpolation, max_size=max_size)
+    resize_result = script_fn(tensor, size=script_size, interpolation=interpolation, max_size=max_size, antialias=True)
     assert_equal(resized_tensor, resize_result)
 
-    _test_fn_on_batch(batch_tensors, F.resize, size=script_size, interpolation=interpolation, max_size=max_size)
+    _test_fn_on_batch(
+        batch_tensors, F.resize, size=script_size, interpolation=interpolation, max_size=max_size, antialias=True
+    )
 
 
 @pytest.mark.parametrize("device", cpu_and_gpu())
@@ -547,7 +550,7 @@ def test_resize_antialias(device, dt, size, interpolation):
         tensor = tensor.to(dt)
 
     resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, antialias=True)
-    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation)
+    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation, antialias=True)
 
     assert resized_tensor.size()[1:] == resized_pil_img.size[::-1]
 
@@ -596,6 +599,23 @@ def test_assert_resize_antialias(interpolation):
         F.resize(tensor, size=(5, 5), interpolation=interpolation, antialias=True)
 
 
+def test_resize_antialias_default_warning():
+
+    img = torch.randint(0, 256, size=(3, 44, 56), dtype=torch.uint8)
+
+    match = "The default value of the antialias"
+    with pytest.warns(UserWarning, match=match):
+        F.resize(img, size=(20, 20))
+    with pytest.warns(UserWarning, match=match):
+        F.resized_crop(img, 0, 0, 10, 10, size=(20, 20))
+
+    # For modes that aren't bicubic or bilinear, don't throw a warning
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        F.resize(img, size=(20, 20), interpolation=NEAREST)
+        F.resized_crop(img, 0, 0, 10, 10, size=(20, 20), interpolation=NEAREST)
+
+
 @pytest.mark.parametrize("device", cpu_and_gpu())
 @pytest.mark.parametrize("dt", [torch.float32, torch.float64, torch.float16])
 @pytest.mark.parametrize("size", [[10, 7], [10, 42], [42, 7]])
@@ -924,7 +944,9 @@ def test_resized_crop(device, mode):
     # 1) resize to the same size, crop to the same size => should be identity
     tensor, _ = _create_data(26, 36, device=device)
 
-    out_tensor = F.resized_crop(tensor, top=0, left=0, height=26, width=36, size=[26, 36], interpolation=mode)
+    out_tensor = F.resized_crop(
+        tensor, top=0, left=0, height=26, width=36, size=[26, 36], interpolation=mode, antialias=True
+    )
     assert_equal(tensor, out_tensor, msg=f"{out_tensor[0, :5, :5]} vs {tensor[0, :5, :5]}")
 
     # 2) resize by half and crop a TL corner
@@ -939,7 +961,14 @@ def test_resized_crop(device, mode):
 
     batch_tensors = _create_data_batch(26, 36, num_samples=4, device=device)
     _test_fn_on_batch(
-        batch_tensors, F.resized_crop, top=1, left=2, height=20, width=30, size=[10, 15], interpolation=NEAREST
+        batch_tensors,
+        F.resized_crop,
+        top=1,
+        left=2,
+        height=20,
+        width=30,
+        size=[10, 15],
+        interpolation=NEAREST,
     )
 
 

test/test_models.py

@@ -1050,5 +1050,25 @@ def test_raft(model_fn, scripted):
     _assert_expected(flow_pred.cpu(), name=model_fn.__name__, atol=1e-2, rtol=1)
 
 
+def test_presets_antialias():
+
+    img = torch.randint(0, 256, size=(1, 3, 224, 224), dtype=torch.uint8)
+
+    match = "The default value of the antialias parameter"
+    with pytest.warns(UserWarning, match=match):
+        models.ResNet18_Weights.DEFAULT.transforms()(img)
+    with pytest.warns(UserWarning, match=match):
+        models.segmentation.DeepLabV3_ResNet50_Weights.DEFAULT.transforms()(img)
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        models.ResNet18_Weights.DEFAULT.transforms(antialias=True)(img)
+        models.segmentation.DeepLabV3_ResNet50_Weights.DEFAULT.transforms(antialias=True)(img)
+
+        models.detection.FasterRCNN_ResNet50_FPN_Weights.DEFAULT.transforms()(img)
+        models.video.R3D_18_Weights.DEFAULT.transforms()(img)
+        models.optical_flow.Raft_Small_Weights.DEFAULT.transforms()(img, img)
+
+
 if __name__ == "__main__":
     pytest.main([__file__])

test/test_prototype_transforms.py

@@ -1,5 +1,6 @@
 import itertools
 import re
+import warnings
 from collections import defaultdict
 
 import numpy as np
@@ -94,7 +95,7 @@ def parametrize_from_transforms(*transforms):
 class TestSmoke:
     @parametrize_from_transforms(
         transforms.RandomErasing(p=1.0),
-        transforms.Resize([16, 16]),
+        transforms.Resize([16, 16], antialias=True),
         transforms.CenterCrop([16, 16]),
         transforms.ConvertDtype(),
         transforms.RandomHorizontalFlip(),
@@ -210,7 +211,7 @@ def test_normalize(self, transform, input):
     @parametrize(
         [
             (
-                transforms.RandomResizedCrop([16, 16]),
+                transforms.RandomResizedCrop([16, 16], antialias=True),
                 itertools.chain(
                     make_images(extra_dims=[(4,)]),
                     make_vanilla_tensor_images(),
@@ -1991,6 +1992,70 @@ def test__transform(self, inpt):
         assert output.dtype == inpt.dtype
 
 
+# TODO: remove this test in 0.17 when the default of antialias changes to True
+def test_antialias_warning():
+    pil_img = PIL.Image.new("RGB", size=(10, 10), color=127)
+    tensor_img = torch.randint(0, 256, size=(3, 10, 10), dtype=torch.uint8)
+    tensor_video = torch.randint(0, 256, size=(2, 3, 10, 10), dtype=torch.uint8)
+
+    match = "The default value of the antialias parameter"
+    with pytest.warns(UserWarning, match=match):
+        transforms.Resize((20, 20))(tensor_img)
+    with pytest.warns(UserWarning, match=match):
+        transforms.RandomResizedCrop((20, 20))(tensor_img)
+    with pytest.warns(UserWarning, match=match):
+        transforms.ScaleJitter((20, 20))(tensor_img)
+    with pytest.warns(UserWarning, match=match):
+        transforms.RandomShortestSize((20, 20))(tensor_img)
+    with pytest.warns(UserWarning, match=match):
+        transforms.RandomResize(10, 20)(tensor_img)
+
+    with pytest.warns(UserWarning, match=match):
+        transforms.functional.resize(tensor_img, (20, 20))
+    with pytest.warns(UserWarning, match=match):
+        transforms.functional.resize_image_tensor(tensor_img, (20, 20))
+
+    with pytest.warns(UserWarning, match=match):
+        transforms.functional.resize(tensor_video, (20, 20))
+    with pytest.warns(UserWarning, match=match):
+        transforms.functional.resize_video(tensor_video, (20, 20))
+
+    with pytest.warns(UserWarning, match=match):
+        datapoints.Image(tensor_img).resize((20, 20))
+    with pytest.warns(UserWarning, match=match):
+        datapoints.Image(tensor_img).resized_crop(0, 0, 10, 10, (20, 20))
+
+    with pytest.warns(UserWarning, match=match):
+        datapoints.Video(tensor_video).resize((20, 20))
+    with pytest.warns(UserWarning, match=match):
+        datapoints.Video(tensor_video).resized_crop(0, 0, 10, 10, (20, 20))
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        transforms.Resize((20, 20))(pil_img)
+        transforms.RandomResizedCrop((20, 20))(pil_img)
+        transforms.ScaleJitter((20, 20))(pil_img)
+        transforms.RandomShortestSize((20, 20))(pil_img)
+        transforms.RandomResize(10, 20)(pil_img)
+        transforms.functional.resize(pil_img, (20, 20))
+
+        transforms.Resize((20, 20), antialias=True)(tensor_img)
+        transforms.RandomResizedCrop((20, 20), antialias=True)(tensor_img)
+        transforms.ScaleJitter((20, 20), antialias=True)(tensor_img)
+        transforms.RandomShortestSize((20, 20), antialias=True)(tensor_img)
+        transforms.RandomResize(10, 20, antialias=True)(tensor_img)
+
+        transforms.functional.resize(tensor_img, (20, 20), antialias=True)
+        transforms.functional.resize_image_tensor(tensor_img, (20, 20), antialias=True)
+        transforms.functional.resize(tensor_video, (20, 20), antialias=True)
+        transforms.functional.resize_video(tensor_video, (20, 20), antialias=True)
+
+        datapoints.Image(tensor_img).resize((20, 20), antialias=True)
+        datapoints.Image(tensor_img).resized_crop(0, 0, 10, 10, (20, 20), antialias=True)
+        datapoints.Video(tensor_video).resize((20, 20), antialias=True)
+        datapoints.Video(tensor_video).resized_crop(0, 0, 10, 10, (20, 20), antialias=True)
+
+
 @pytest.mark.parametrize("image_type", (PIL.Image, torch.Tensor, datapoints.Image))
 @pytest.mark.parametrize("label_type", (torch.Tensor, int))
 @pytest.mark.parametrize("dataset_return_type", (dict, tuple))

test/test_transforms.py

@@ -2,6 +2,7 @@
 import os
 import random
 import re
+import warnings
 from functools import partial
 
 import numpy as np
@@ -319,7 +320,7 @@ def test_randomresized_params():
         scale_range = (scale_min, scale_min + round(random.random(), 2))
         aspect_min = max(round(random.random(), 2), epsilon)
         aspect_ratio_range = (aspect_min, aspect_min + round(random.random(), 2))
-        randresizecrop = transforms.RandomResizedCrop(size, scale_range, aspect_ratio_range)
+        randresizecrop = transforms.RandomResizedCrop(size, scale_range, aspect_ratio_range, antialias=True)
         i, j, h, w = randresizecrop.get_params(img, scale_range, aspect_ratio_range)
         aspect_ratio_obtained = w / h
         assert (
@@ -366,7 +367,7 @@ def test_randomresized_params():
 def test_resize(height, width, osize, max_size):
     img = Image.new("RGB", size=(width, height), color=127)
 
-    t = transforms.Resize(osize, max_size=max_size)
+    t = transforms.Resize(osize, max_size=max_size, antialias=True)
     result = t(img)
 
     msg = f"{height}, {width} - {osize} - {max_size}"
@@ -424,7 +425,7 @@ def test_resize_sequence_output(height, width, osize):
     img = Image.new("RGB", size=(width, height), color=127)
     oheight, owidth = osize
 
-    t = transforms.Resize(osize)
+    t = transforms.Resize(osize, antialias=True)
     result = t(img)
 
     assert (owidth, oheight) == result.size
@@ -439,6 +440,16 @@ def test_resize_antialias_error():
         t(img)
 
 
+def test_resize_antialias_default_warning():
+
+    img = Image.new("RGB", size=(10, 10), color=127)
+    # We make sure we don't warn for PIL images since the default behaviour doesn't change
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        transforms.Resize((20, 20))(img)
+        transforms.RandomResizedCrop((20, 20))(img)
+
+
 @pytest.mark.parametrize("height, width", ((32, 64), (64, 32)))
 def test_resize_size_equals_small_edge_size(height, width):
     # Non-regression test for https://github.com/pytorch/vision/issues/5405
@@ -447,7 +458,7 @@ def test_resize_size_equals_small_edge_size(height, width):
     img = Image.new("RGB", size=(width, height), color=127)
 
     small_edge = min(height, width)
-    t = transforms.Resize(small_edge, max_size=max_size)
+    t = transforms.Resize(small_edge, max_size=max_size, antialias=True)
     result = t(img)
     assert max(result.size) == max_size
 
@@ -1424,11 +1435,11 @@ def test_random_choice(proba_passthrough, seed):
 def test_random_order():
     random_state = random.getstate()
     random.seed(42)
-    random_order_transform = transforms.RandomOrder([transforms.Resize(20), transforms.CenterCrop(10)])
+    random_order_transform = transforms.RandomOrder([transforms.Resize(20, antialias=True), transforms.CenterCrop(10)])
     img = transforms.ToPILImage()(torch.rand(3, 25, 25))
     num_samples = 250
     num_normal_order = 0
-    resize_crop_out = transforms.CenterCrop(10)(transforms.Resize(20)(img))
+    resize_crop_out = transforms.CenterCrop(10)(transforms.Resize(20, antialias=True)(img))
     for _ in range(num_samples):
         out = random_order_transform(img)
         if out == resize_crop_out: