Merge branch 'main' into GaussianNoiseUINT8

Author: NicolasHug

.github/scripts/setup-env.sh

@@ -80,22 +80,8 @@ if [[ $GPU_ARCH_TYPE == 'cuda' ]]; then
 fi
 echo '::endgroup::'
 
-echo '::group::Install third party dependencies prior to TorchVision install'
-# Installing with `easy_install`, e.g. `python setup.py install` or `python setup.py develop`, has some quirks when
-# when pulling in third-party dependencies. For example:
-# - On Windows, we often hit an SSL error although `pip` can install just fine.
-# - It happily pulls in pre-releases, which can lead to more problems down the line.
-#   `pip` does not unless explicitly told to do so.
-# Thus, we use `easy_install` to extract the third-party dependencies here and install them upfront with `pip`.
-python setup.py egg_info
-# The requires.txt cannot be used with `pip install -r` directly. The requirements are listed at the top and the
-# optional dependencies come in non-standard syntax after a blank line. Thus, we just extract the header.
-sed -e '/^$/,$d' *.egg-info/requires.txt | tee requirements.txt
-pip install --progress-bar=off -r requirements.txt
-echo '::endgroup::'
-
 echo '::group::Install TorchVision'
-python setup.py develop
+pip install -e . -v --no-build-isolation
 echo '::endgroup::'
 
 echo '::group::Install torchvision-extra-decoders'

CMakeLists.txt

@@ -11,8 +11,6 @@ option(WITH_JPEG "Enable features requiring LibJPEG." ON)
 # untested. Since building from cmake is very low pri anyway, this is OK. If
 # you're a user and you need this, please open an issue (and a PR!).
 option(WITH_WEBP "Enable features requiring LibWEBP." OFF)
-# Same here
-option(WITH_AVIF "Enable features requiring LibAVIF." OFF)
 
 if(WITH_CUDA)
   enable_language(CUDA)
@@ -40,12 +38,7 @@ endif()
 
 if (WITH_WEBP)
     add_definitions(-DWEBP_FOUND)
-    find_package(WEBP REQUIRED)
-endif()
-
-if (WITH_AVIF)
-    add_definitions(-DAVIF_FOUND)
-    find_package(AVIF REQUIRED)
+    find_package(WebP REQUIRED)
 endif()
 
 function(CUDA_CONVERT_FLAGS EXISTING_TARGET)

CONTRIBUTING.md

@@ -58,11 +58,11 @@ details.
 ```bash
 git clone https://github.com/pytorch/vision.git
 cd vision
-python setup.py develop  # use install instead of develop if you don't care about development.
+pip install -e . -v --no-build-isolation  # leave out the -e switch if you don't care about development.
 # or, for OSX
-# MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py develop
+# MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ pip install -e . -v --no-build-isolation
 # for C++ debugging, use DEBUG=1
-# DEBUG=1 python setup.py develop
+# DEBUG=1 pip install -e . -v --no-build-isolation
 ```
 
 By default, GPU support is built if CUDA is found and `torch.cuda.is_available()` is true. It's possible to force

README.md

@@ -20,19 +20,22 @@ versions.
 
 | `torch`            | `torchvision`      | Python              |
 | ------------------ | ------------------ | ------------------- |
-| `main` / `nightly` | `main` / `nightly` | `>=3.9`, `<=3.12`   |
-| `2.5`              | `0.20`             | `>=3.9`, `<=3.12`   |
-| `2.4`              | `0.19`             | `>=3.8`, `<=3.12`   |
-| `2.3`              | `0.18`             | `>=3.8`, `<=3.12`   |
-| `2.2`              | `0.17`             | `>=3.8`, `<=3.11`   |
-| `2.1`              | `0.16`             | `>=3.8`, `<=3.11`   |
-| `2.0`              | `0.15`             | `>=3.8`, `<=3.11`   |
+| `main` / `nightly` | `main` / `nightly` | `>=3.10`, `<=3.13`   |
+| `2.8`              | `0.23`             | `>=3.9`, `<=3.13`   |
+| `2.7`              | `0.22`             | `>=3.9`, `<=3.13`   |
+| `2.6`              | `0.21`             | `>=3.9`, `<=3.12`   |
 
 <details>
     <summary>older versions</summary>
 
 | `torch` | `torchvision`     | Python                    |
 |---------|-------------------|---------------------------|
+| `2.5`              | `0.20`             | `>=3.9`, `<=3.12`   |
+| `2.4`              | `0.19`             | `>=3.8`, `<=3.12`   |
+| `2.3`              | `0.18`             | `>=3.8`, `<=3.12`   |
+| `2.2`              | `0.17`             | `>=3.8`, `<=3.11`   |
+| `2.1`              | `0.16`             | `>=3.8`, `<=3.11`   |
+| `2.0`              | `0.15`             | `>=3.8`, `<=3.11`   |
 | `1.13`  | `0.14`            | `>=3.7.2`, `<=3.10`       |
 | `1.12`  | `0.13`            | `>=3.7`, `<=3.10`         |
 | `1.11`  | `0.12`            | `>=3.7`, `<=3.10`         |
@@ -61,19 +64,6 @@ Torchvision currently supports the following image backends:
 
 Read more in in our [docs](https://pytorch.org/vision/stable/transforms.html).
 
-## [UNSTABLE] Video Backend
-
-Torchvision currently supports the following video backends:
-
-- [pyav](https://github.com/PyAV-Org/PyAV) (default) - Pythonic binding for ffmpeg libraries.
-- video_reader - This needs ffmpeg to be installed and torchvision to be built from source. There shouldn't be any
-  conflicting version of ffmpeg installed. Currently, this is only supported on Linux.
-
-```
-conda install -c conda-forge 'ffmpeg<4.3'
-python setup.py install
-```
-
 # Using the models on C++
 
 Refer to [example/cpp](https://github.com/pytorch/vision/tree/main/examples/cpp).

test/test_ops.py

@@ -1,7 +1,7 @@
 import math
 import os
 from abc import ABC, abstractmethod
-from functools import lru_cache
+from functools import lru_cache, partial
 from itertools import product
 from typing import Callable
 
@@ -242,7 +242,7 @@ def _helper_boxes_shape(self, func):
             boxes = torch.tensor([[0, 0, 3, 3]], dtype=a.dtype)
             func(a, boxes, output_size=(2, 2))
 
-        # test boxes as List[Tensor[N, 4]]
+        # test boxes as list[Tensor[N, 4]]
         with pytest.raises(AssertionError):
             a = torch.linspace(1, 8 * 8, 8 * 8).reshape(1, 1, 8, 8)
             boxes = torch.tensor([[0, 0, 3]], dtype=a.dtype)
@@ -1073,15 +1073,15 @@ def test_forward(self, device, contiguous, batch_sz, dtype=None):
         expected = self.expected_fn(x, weight, offset, mask, bias, stride=stride, padding=padding, dilation=dilation)
 
         torch.testing.assert_close(
-            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres:\n{res}\nexpected:\n{expected}"
+            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres: \n{res}\nexpected: \n{expected}"
         )
 
         # no modulation test
         res = layer(x, offset)
         expected = self.expected_fn(x, weight, offset, None, bias, stride=stride, padding=padding, dilation=dilation)
 
         torch.testing.assert_close(
-            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres:\n{res}\nexpected:\n{expected}"
+            res.to(expected), expected, rtol=tol, atol=tol, msg=f"\nres: \n{res}\nexpected: \n{expected}"
         )
 
     def test_wrong_sizes(self):
@@ -1446,34 +1446,60 @@ def test_bbox_convert_jit(self):
 
 
 class TestBoxArea:
-    def area_check(self, box, expected, atol=1e-4):
-        out = ops.box_area(box)
+    def area_check(self, box, expected, fmt="xyxy", atol=1e-4):
+        out = ops.box_area(box, fmt=fmt)
         torch.testing.assert_close(out, expected, rtol=0.0, check_dtype=False, atol=atol)
 
     @pytest.mark.parametrize("dtype", [torch.int8, torch.int16, torch.int32, torch.int64])
-    def test_int_boxes(self, dtype):
-        box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=dtype)
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_int_boxes(self, dtype, fmt):
+        box_tensor = ops.box_convert(
+            torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=dtype), in_fmt="xyxy", out_fmt=fmt
+        )
         expected = torch.tensor([10000, 0], dtype=torch.int32)
-        self.area_check(box_tensor, expected)
+        self.area_check(box_tensor, expected, fmt)
 
     @pytest.mark.parametrize("dtype", [torch.float32, torch.float64])
-    def test_float_boxes(self, dtype):
-        box_tensor = torch.tensor(FLOAT_BOXES, dtype=dtype)
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_float_boxes(self, dtype, fmt):
+        box_tensor = ops.box_convert(torch.tensor(FLOAT_BOXES, dtype=dtype), in_fmt="xyxy", out_fmt=fmt)
         expected = torch.tensor([604723.0806, 600965.4666, 592761.0085], dtype=dtype)
-        self.area_check(box_tensor, expected)
-
-    def test_float16_box(self):
-        box_tensor = torch.tensor(
-            [[2.825, 1.8625, 3.90, 4.85], [2.825, 4.875, 19.20, 5.10], [2.925, 1.80, 8.90, 4.90]], dtype=torch.float16
+        self.area_check(box_tensor, expected, fmt)
+
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_float16_box(self, fmt):
+        box_tensor = ops.box_convert(
+            torch.tensor(
+                [[2.825, 1.8625, 3.90, 4.85], [2.825, 4.875, 19.20, 5.10], [2.925, 1.80, 8.90, 4.90]],
+                dtype=torch.float16,
+            ),
+            in_fmt="xyxy",
+            out_fmt=fmt,
         )
 
         expected = torch.tensor([3.2170, 3.7108, 18.5071], dtype=torch.float16)
-        self.area_check(box_tensor, expected, atol=0.01)
+        self.area_check(box_tensor, expected, fmt, atol=0.01)
+
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_box_area_jit(self, fmt):
+        box_tensor = ops.box_convert(
+            torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=torch.float), in_fmt="xyxy", out_fmt=fmt
+        )
+        expected = ops.box_area(box_tensor, fmt)
 
-    def test_box_area_jit(self):
-        box_tensor = torch.tensor([[0, 0, 100, 100], [0, 0, 0, 0]], dtype=torch.float)
-        expected = ops.box_area(box_tensor)
-        scripted_fn = torch.jit.script(ops.box_area)
+        class BoxArea(torch.nn.Module):
+            # We are using this intermediate class
+            # since torchscript does not support
+            # neither partial nor lambda functions for this test.
+            def __init__(self, fmt):
+                super().__init__()
+                self.area = ops.box_area
+                self.fmt = fmt
+
+            def forward(self, boxes):
+                return self.area(boxes, self.fmt)
+
+        scripted_fn = torch.jit.script(BoxArea(fmt))
         scripted_area = scripted_fn(box_tensor)
         torch.testing.assert_close(scripted_area, expected)
 
@@ -1487,25 +1513,28 @@ def test_box_area_jit(self):
 ]
 
 
-def gen_box(size, dtype=torch.float):
+def gen_box(size, dtype=torch.float, fmt="xyxy") -> Tensor:
     xy1 = torch.rand((size, 2), dtype=dtype)
     xy2 = xy1 + torch.rand((size, 2), dtype=dtype)
-    return torch.cat([xy1, xy2], axis=-1)
+    return ops.box_convert(torch.cat([xy1, xy2], axis=-1), in_fmt="xyxy", out_fmt=fmt)
 
 
 class TestIouBase:
     @staticmethod
-    def _run_test(target_fn: Callable, actual_box1, actual_box2, dtypes, atol, expected):
+    def _run_test(target_fn: Callable, actual_box1, actual_box2, dtypes, atol, expected, fmt="xyxy"):
         for dtype in dtypes:
-            actual_box1 = torch.tensor(actual_box1, dtype=dtype)
-            actual_box2 = torch.tensor(actual_box2, dtype=dtype)
+            _actual_box1 = ops.box_convert(torch.tensor(actual_box1, dtype=dtype), in_fmt="xyxy", out_fmt=fmt)
+            _actual_box2 = ops.box_convert(torch.tensor(actual_box2, dtype=dtype), in_fmt="xyxy", out_fmt=fmt)
             expected_box = torch.tensor(expected)
-            out = target_fn(actual_box1, actual_box2)
+            out = target_fn(
+                _actual_box1,
+                _actual_box2,
+            )
             torch.testing.assert_close(out, expected_box, rtol=0.0, check_dtype=False, atol=atol)
 
     @staticmethod
-    def _run_jit_test(target_fn: Callable, actual_box: list):
-        box_tensor = torch.tensor(actual_box, dtype=torch.float)
+    def _run_jit_test(target_fn: Callable, actual_box: list, fmt="xyxy"):
+        box_tensor = ops.box_convert(torch.tensor(actual_box, dtype=torch.float), in_fmt="xyxy", out_fmt=fmt)
         expected = target_fn(box_tensor, box_tensor)
         scripted_fn = torch.jit.script(target_fn)
         scripted_out = scripted_fn(box_tensor, box_tensor)
@@ -1522,13 +1551,21 @@ def _cartesian_product(boxes1, boxes2, target_fn: Callable):
         return result
 
     @staticmethod
-    def _run_cartesian_test(target_fn: Callable):
-        boxes1 = gen_box(5)
-        boxes2 = gen_box(7)
+    def _run_cartesian_test(target_fn: Callable, fmt: str = "xyxy"):
+        boxes1 = gen_box(5, fmt=fmt)
+        boxes2 = gen_box(7, fmt=fmt)
         a = TestIouBase._cartesian_product(boxes1, boxes2, target_fn)
         b = target_fn(boxes1, boxes2)
         torch.testing.assert_close(a, b)
 
+    @staticmethod
+    def _run_batch_test(target_fn: Callable, fmt: str = "xyxy"):
+        boxes1 = torch.stack([gen_box(5, fmt=fmt) for _ in range(3)], dim=0)
+        boxes2 = torch.stack([gen_box(5, fmt=fmt) for _ in range(3)], dim=0)
+        native: Tensor = target_fn(boxes1, boxes2)
+        iterative: Tensor = torch.stack([target_fn(*pairs) for pairs in zip(boxes1, boxes2)], dim=0)
+        torch.testing.assert_close(native, iterative)
+
 
 class TestBoxIou(TestIouBase):
     int_expected = [[1.0, 0.25, 0.0], [0.25, 1.0, 0.0], [0.0, 0.0, 1.0], [0.0625, 0.25, 0.0]]
@@ -1542,14 +1579,33 @@ class TestBoxIou(TestIouBase):
             pytest.param(FLOAT_BOXES, FLOAT_BOXES, [torch.float32, torch.float64], 1e-3, float_expected),
         ],
     )
-    def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected):
-        self._run_test(ops.box_iou, actual_box1, actual_box2, dtypes, atol, expected)
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_iou(self, actual_box1, actual_box2, dtypes, atol, expected, fmt):
+        self._run_test(partial(ops.box_iou, fmt=fmt), actual_box1, actual_box2, dtypes, atol, expected, fmt)
 
-    def test_iou_jit(self):
-        self._run_jit_test(ops.box_iou, INT_BOXES)
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_iou_jit(self, fmt):
+        class IoUJit(torch.nn.Module):
+            # We are using this intermediate class
+            # since torchscript does not support
+            # neither partial nor lambda functions for this test.
+            def __init__(self, fmt):
+                super().__init__()
+                self.iou = ops.box_iou
+                self.fmt = fmt
 
-    def test_iou_cartesian(self):
-        self._run_cartesian_test(ops.box_iou)
+            def forward(self, boxes1, boxes2):
+                return self.iou(boxes1, boxes2, fmt=self.fmt)
+
+        self._run_jit_test(IoUJit(fmt=fmt), INT_BOXES, fmt)
+
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_iou_cartesian(self, fmt):
+        self._run_cartesian_test(partial(ops.box_iou, fmt=fmt))
+
+    @pytest.mark.parametrize("fmt", ["xyxy", "xywh", "cxcywh"])
+    def test_iou_batch(self, fmt):
+        self._run_batch_test(partial(ops.box_iou, fmt=fmt))
 
 
 class TestGeneralizedBoxIou(TestIouBase):
@@ -1573,6 +1629,9 @@ def test_iou_jit(self):
     def test_iou_cartesian(self):
         self._run_cartesian_test(ops.generalized_box_iou)
 
+    def test_iou_batch(self):
+        self._run_batch_test(ops.generalized_box_iou)
+
 
 class TestDistanceBoxIoU(TestIouBase):
     int_expected = [
@@ -1600,6 +1659,9 @@ def test_iou_jit(self):
     def test_iou_cartesian(self):
         self._run_cartesian_test(ops.distance_box_iou)
 
+    def test_iou_batch(self):
+        self._run_batch_test(ops.distance_box_iou)
+
 
 class TestCompleteBoxIou(TestIouBase):
     int_expected = [
@@ -1627,6 +1689,9 @@ def test_iou_jit(self):
     def test_iou_cartesian(self):
         self._run_cartesian_test(ops.complete_box_iou)
 
+    def test_iou_batch(self):
+        self._run_batch_test(ops.complete_box_iou)
+
 
 def get_boxes(dtype, device):
     box1 = torch.tensor([-1, -1, 1, 1], dtype=dtype, device=device)