Merge branch 'main' of github.com:pytorch/vision into tv-tensor-enhance

.github/workflows/prototype-tests-linux-gpu.yml

@@ -12,6 +12,7 @@ jobs:
           - "3.9"
           - "3.10"
           - "3.11"
+          - "3.12"
         runner: ["linux.12xlarge"]
         gpu-arch-type: ["cpu"]
         include:

.github/workflows/tests-schedule.yml

@@ -29,6 +29,9 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v2
 
+      - name: TODO REMOVE THIS! Install non pre-release version of mpmath.
+        run: pip install "mpmath<1.4"
+
       - name: Install torch nightly build
         run: pip install --pre torch -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
 

.github/workflows/tests.yml

@@ -18,6 +18,7 @@ jobs:
           - "3.9"
           - "3.10"
           - "3.11"
+          - "3.12"
         runner: ["linux.12xlarge"]
         gpu-arch-type: ["cpu"]
         include:
@@ -51,6 +52,7 @@ jobs:
           - "3.9"
           - "3.10"
           - "3.11"
+          - "3.12"
         runner: ["macos-12"]
         include:
           - python-version: "3.8"
@@ -81,6 +83,7 @@ jobs:
           - "3.9"
           - "3.10"
           - "3.11"
+          - "3.12"
         runner: ["windows.4xlarge"]
         gpu-arch-type: ["cpu"]
         include:

docs/source/models/ssdlite.rst

@@ -17,7 +17,7 @@ The following model builders can be used to instantiate a SSD Lite model, with o
 without pre-trained weights. All the model builders internally rely on the
 ``torchvision.models.detection.ssd.SSD`` base class. Please refer to the `source
 code
-<https://github.com/pytorch/vision/blob/main/torchvision/models/detection/ssd.py>`_ for
+<https://github.com/pytorch/vision/blob/main/torchvision/models/detection/ssdlite.py>`_ for
 more details about this class.
 
 .. autosummary::

docs/source/transforms.rst

@@ -347,6 +347,7 @@ Color
     v2.RandomChannelPermutation
     v2.RandomPhotometricDistort
     v2.Grayscale
+    v2.RGB
     v2.RandomGrayscale
     v2.GaussianBlur
     v2.RandomInvert
@@ -364,6 +365,7 @@ Functionals
 
     v2.functional.permute_channels
     v2.functional.rgb_to_grayscale
+    v2.functional.grayscale_to_rgb
     v2.functional.to_grayscale
     v2.functional.gaussian_blur
     v2.functional.invert
@@ -405,6 +407,7 @@ Miscellaneous
     v2.SanitizeBoundingBoxes
     v2.ClampBoundingBoxes
     v2.UniformTemporalSubsample
+    v2.JPEG
 
 Functionals
 
@@ -414,8 +417,10 @@ Functionals
 
     v2.functional.normalize
     v2.functional.erase
+    v2.functional.sanitize_bounding_boxes
     v2.functional.clamp_bounding_boxes
     v2.functional.uniform_temporal_subsample
+    v2.functional.jpeg
 
 .. _conversion_transforms:
 
@@ -583,7 +588,7 @@ Conversion
     while performing the conversion, while some may not do any scaling. By
     scaling, we mean e.g. that a ``uint8`` -> ``float32`` would map the [0,
     255] range into [0, 1] (and vice-versa). See :ref:`range_and_dtype`.
-    
+
 .. autosummary::
     :toctree: generated/
     :template: class.rst

gallery/transforms/plot_transforms_e2e.py

@@ -151,8 +151,8 @@
     # We need a custom collation function here, since the object detection
     # models expect a sequence of images and target dictionaries. The default
     # collation function tries to torch.stack() the individual elements,
-    # which fails in general for object detection, because the number of bouding
-    # boxes varies between the images of a same batch.
+    # which fails in general for object detection, because the number of bounding
+    # boxes varies between the images of the same batch.
     collate_fn=lambda batch: tuple(zip(*batch)),
 )
 

gallery/transforms/plot_transforms_illustrations.py

@@ -237,6 +237,17 @@
 equalized_imgs = [equalizer(orig_img) for _ in range(4)]
 plot([orig_img] + equalized_imgs)
 
+# %%
+# JPEG
+# ~~~~~~~~~~~~~~
+# The :class:`~torchvision.transforms.v2.JPEG` transform
+# (see also :func:`~torchvision.transforms.v2.functional.jpeg`)
+# applies JPEG compression to the given image with random
+# degree of compression.
+jpeg = v2.JPEG((5, 50))
+jpeg_imgs = [jpeg(orig_img) for _ in range(4)]
+plot([orig_img] + jpeg_imgs)
+
 # %%
 # Augmentation Transforms
 # -----------------------

maintainer_guide.md

@@ -1,4 +1,4 @@
-## Torchvision maintainers guide
+# Torchvision maintainers guide
 
 This document aims at documenting user-facing policies / principles used when
 developing and maintaining torchvision. Other maintainer info (e.g. release

scripts/release_notes/classify_prs.py

@@ -1,40 +1,22 @@
 # In[1]:
-
-# imports and set configuration
 import pandas as pd
-from retrieve_prs_data import run
-
-exclude_prototype = True
-data_filename = "10.0_to_11.0-rc2.json"
-previous_release = "v10.0"
-current_release = "v11.0-rc2"
 
 # In[2]:
-
-
+data_filename = "data.json"
 df = pd.read_json(data_filename).T
 df.tail()
 
-
 # In[3]:
-
-
 all_labels = {lbl for labels in df["labels"] for lbl in labels}
 all_labels
 
-
 # In[4]:
-
-
 # Add one column per label
 for label in all_labels:
     df[label] = df["labels"].apply(lambda labels_list: label in labels_list)
 df.head()
 
-
 # In[5]:
-
-
 # Add a clean "module" column. It contains tuples since PRs can have more than one module.
 # Maybe we should include "topics" in that column as well?
 
@@ -51,37 +33,26 @@
 df["module"] = df.module.apply(tuple)
 df.head()
 
-
 # In[6]:
-
-
 mod_df = df.set_index("module").sort_index()
 mod_df.tail()
 
-
 # In[7]:
-
-
 # All improvement PRs
 mod_df[mod_df["enhancement"]].head()
 
-
 # In[8]:
-
-
 # improvement f module
 # note: don't filter module name on the index as the index contain tuples with non-exclusive values
 # Use the boolean column instead
 mod_df[mod_df["enhancement"] & mod_df["module: transforms"]]
 
 
 # In[9]:
-
-
-def format_prs(mod_df):
+def format_prs(mod_df, exclude_prototype=True):
     out = []
     for idx, row in mod_df.iterrows():
-        if exclude_prototype and row["prototype"]:
+        if exclude_prototype and "prototype" in row and row["prototype"]:
             continue
         modules = idx
         # Put "documentation" and "tests" first for sorting to be dece
@@ -98,8 +69,6 @@ def format_prs(mod_df):
 
 
 # In[10]:
-
-
 included_prs = pd.DataFrame()
 
 # If labels are accurate, this shouhld generate most of the release notes already
@@ -112,27 +81,40 @@ def format_prs(mod_df):
     ("Bug Fixes", "bug"),
     ("Code Quality", "code quality"),
 ):
-    print(f"## {section_title}")
-    print()
-    tmp_df = mod_df[mod_df[module_idx]]
-    included_prs = pd.concat([included_prs, tmp_df])
-    print(format_prs(tmp_df))
-    print()
+    if module_idx in mod_df:
+        print(f"## {section_title}")
+        print()
+        tmp_df = mod_df[mod_df[module_idx]]
+        included_prs = pd.concat([included_prs, tmp_df])
+        print(format_prs(tmp_df))
+        print()
 
 
 # In[11]:
-
-
 # Missing PRs are these ones... classify them manually
 missing_prs = pd.concat([mod_df, included_prs]).drop_duplicates(subset="pr_number", keep=False)
 print(format_prs(missing_prs))
 
 # In[12]:
-
 # Generate list of contributors
 print()
 print("## Contributors")
 
-command_to_run = f"{{ git shortlog -s {previous_release}..{current_release} | cut -f2- & git log -s {previous_release}..{current_release} | grep Co-authored | cut -f2- -d: | cut -f1 -d\\< | sed 's/^ *//;s/ *$//' ; }} | sort --ignore-case | uniq | tr '\\n' ';' | sed 's/;/, /g;s/, $//' | fold -s"
-rc, output, err = run(command_to_run)
-print(output)
+previous_release = "c35d3855ccbfa6a36e6ae6337a1f2c721c1f1e78"
+current_release = "5181a854d8b127cf465cd22a67c1b5aaf6ccae05"
+print(
+    f"{{ git shortlog -s {previous_release}..{current_release} | cut -f2- & git log -s {previous_release}..{current_release} | grep Co-authored | cut -f2- -d: | cut -f1 -d\\< | sed 's/^ *//;s/ *//' ; }} | sort --ignore-case | uniq | tr '\\n' ';' | sed 's/;/, /g;s/,//' | fold -s"
+)
+
+# In[13]:
+# Utility to extract PR numbers only from multiple lines, useful to bundle all
+# the docs changes for example:
+import re
+
+s = """
+
+[] Remove unnecessary dependency from macOS/Conda binaries (#8077)
+[rocm] [ROCm] remove HCC references (#8070)
+"""
+
+print(", ".join(re.findall("(#\\d+)", s)))

test/test_datasets.py

@@ -827,6 +827,11 @@ def test_transforms_v2_wrapper_spawn(self):
         with self.create_dataset(transform=v2.Resize(size=expected_size)) as (dataset, _):
             datasets_utils.check_transforms_v2_wrapper_spawn(dataset, expected_size=expected_size)
 
+    def test_slice_error(self):
+        with self.create_dataset() as (dataset, _):
+            with pytest.raises(ValueError, match="Index must be of type integer"):
+                dataset[:2]
+
 
 class CocoCaptionsTestCase(CocoDetectionTestCase):
     DATASET_CLASS = datasets.CocoCaptions
@@ -1620,6 +1625,10 @@ def inject_fake_data(self, tmpdir, config):
             num_examples_total += num_examples
             classes.append(cls)
 
+        if config.pop("make_empty_class", False):
+            os.makedirs(pathlib.Path(tmpdir) / "empty_class")
+            classes.append("empty_class")
+
         return dict(num_examples=num_examples_total, classes=classes)
 
     def _file_name_fn(self, cls, ext, idx):
@@ -1644,6 +1653,23 @@ def test_classes(self, config):
             assert len(dataset.classes) == len(info["classes"])
             assert all([a == b for a, b in zip(dataset.classes, info["classes"])])
 
+    def test_allow_empty(self):
+        config = {
+            "extensions": self._EXTENSIONS,
+            "make_empty_class": True,
+        }
+
+        config["allow_empty"] = True
+        with self.create_dataset(config) as (dataset, info):
+            assert "empty_class" in dataset.classes
+            assert len(dataset.classes) == len(info["classes"])
+            assert all([a == b for a, b in zip(dataset.classes, info["classes"])])
+
+        config["allow_empty"] = False
+        with pytest.raises(FileNotFoundError, match="Found no valid file"):
+            with self.create_dataset(config) as (dataset, info):
+                pass
+
 
 class ImageFolderTestCase(datasets_utils.ImageDatasetTestCase):
     DATASET_CLASS = datasets.ImageFolder

test/test_datasets_download.py

@@ -1,12 +1,12 @@
 import contextlib
 import itertools
+import shutil
 import tempfile
 import time
 import traceback
 import unittest.mock
 import warnings
 from datetime import datetime
-from distutils import dir_util
 from os import path
 from urllib.error import HTTPError, URLError
 from urllib.parse import urlparse
@@ -180,7 +180,7 @@ def collect_urls(dataset_cls, *args, **kwargs):
 @pytest.fixture(scope="module", autouse=True)
 def root():
     yield ROOT
-    dir_util.remove_tree(ROOT)
+    shutil.rmtree(ROOT)
 
 
 def places365():
@@ -327,12 +327,6 @@ def kitti():
     )
 
 
-def stanford_cars():
-    return itertools.chain.from_iterable(
-        [collect_urls(datasets.StanfordCars, ROOT, split=split, download=True) for split in ["train", "test"]]
-    )
-
-
 def url_parametrization(*dataset_urls_and_ids_fns):
     return pytest.mark.parametrize(
         "url",
@@ -378,9 +372,9 @@ def test_url_is_accessible(url):
     retry(lambda: assert_url_is_accessible(url))
 
 
-@url_parametrization(
-    stanford_cars,  # https://github.com/pytorch/vision/issues/7545
-)
+# TODO: if e.g. caltech101 starts failing, remove the pytest.mark.parametrize below and use
+# @url_parametrization(caltech101)
+@pytest.mark.parametrize("url", ("http://url_that_doesnt_exist.com",))  # here until we actually have a failing dataset
 @pytest.mark.xfail
 def test_url_is_not_accessible(url):
     """