Merge branch 'pytorch:main' into main

Author: dhruvbird

.clang-format

@@ -60,9 +60,6 @@ MacroBlockBegin: ''
 MacroBlockEnd:   ''
 MaxEmptyLinesToKeep: 1
 NamespaceIndentation: None
-ObjCBlockIndentWidth: 2
-ObjCSpaceAfterProperty: false
-ObjCSpaceBeforeProtocolList: false
 PenaltyBreakBeforeFirstCallParameter: 1
 PenaltyBreakComment: 300
 PenaltyBreakFirstLessLess: 120
@@ -85,4 +82,11 @@ SpacesInSquareBrackets: false
 Standard:        Cpp11
 TabWidth:        8
 UseTab:          Never
+---
+Language: ObjC
+ColumnLimit: 120
+AlignAfterOpenBracket: Align
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: false
 ...

.git-blame-ignore-revs

@@ -7,3 +7,5 @@
 d367a01a18a3ae6bee13d8be3b63fd6a581ea46f
 # Upgrade usort to 1.0.2 and black to 22.3.0 (#5106) 
 6ca9c76adb6daf2695d603ad623a9cf1c4f4806f
+# Fix unnecessary exploded black formatting (#7709)
+a335d916db0694770e8152f41e19195de3134523

.github/scripts/setup-env.sh

@@ -54,7 +54,11 @@ echo '::endgroup::'
 
 if [[ "${OS_TYPE}" == windows && "${GPU_ARCH_TYPE}" == cuda ]]; then
   echo '::group::Install VisualStudio CUDA extensions on Windows'
-  TARGET_DIR="/c/Program Files (x86)/Microsoft Visual Studio/2019/BuildTools/MSBuild/Microsoft/VC/v160/BuildCustomizations"
+  if [[ "${VC_YEAR:-}" == "2022" ]]; then
+    TARGET_DIR="/c/Program Files (x86)/Microsoft Visual Studio/2022/BuildTools/MSBuild/Microsoft/VC/v170/BuildCustomizations"
+  else
+    TARGET_DIR="/c/Program Files (x86)/Microsoft Visual Studio/2019/BuildTools/MSBuild/Microsoft/VC/v160/BuildCustomizations"
+  fi
   mkdir -p "${TARGET_DIR}"
   cp -r "${CUDA_HOME}/MSBuildExtensions/"* "${TARGET_DIR}"
   echo '::endgroup::'

.github/workflows/build-cmake.yml

@@ -74,9 +74,9 @@ jobs:
       script: |
         set -euo pipefail
 
-        source packaging/windows/internal/vc_install_helper.sh
-
         export PYTHON_VERSION=3.8
+        export VC_YEAR=2022
+        export VSDEVCMD_ARGS=""
         export GPU_ARCH_TYPE=${{ matrix.gpu-arch-type }}
         export GPU_ARCH_VERSION=${{ matrix.gpu-arch-version }}
 

.github/workflows/tests.yml

@@ -98,6 +98,8 @@ jobs:
         set -euxo pipefail
 
         export PYTHON_VERSION=${{ matrix.python-version }}
+        export VC_YEAR=2019
+        export VSDEVCMD_ARGS=""
         export GPU_ARCH_TYPE=${{ matrix.gpu-arch-type }}
         export GPU_ARCH_VERSION=${{ matrix.gpu-arch-version }}
         

.pre-commit-config.yaml

@@ -6,6 +6,8 @@ repos:
       - id: check-toml
       - id: check-yaml
         exclude: packaging/.*
+        args:
+          - --allow-multiple-documents
       - id: mixed-line-ending
         args: [--fix=lf]
       - id: end-of-file-fixer

docs/source/conf.py

@@ -140,7 +140,7 @@
     "logo_only": True,
     "pytorch_project": "docs",
     "navigation_with_keys": True,
-    "analytics_id": "UA-117752657-2",
+    "analytics_id": "GTM-T8XT4PS",
 }
 
 html_logo = "_static/img/pytorch-logo-dark.svg"

packaging/wheel/relocate.py

@@ -2,7 +2,6 @@
 
 import glob
 import hashlib
-import io
 
 # Standard library imports
 import os
@@ -65,21 +64,12 @@
 PYTHON_VERSION = sys.version_info
 
 
-def read_chunks(file, size=io.DEFAULT_BUFFER_SIZE):
-    """Yield pieces of data from a file-like object until EOF."""
-    while True:
-        chunk = file.read(size)
-        if not chunk:
-            break
-        yield chunk
-
-
 def rehash(path, blocksize=1 << 20):
     """Return (hash, length) for path using hashlib.sha256()"""
     h = hashlib.sha256()
     length = 0
     with open(path, "rb") as f:
-        for block in read_chunks(f, size=blocksize):
+        while block := f.read(blocksize):
             length += len(block)
             h.update(block)
     digest = "sha256=" + urlsafe_b64encode(h.digest()).decode("latin1").rstrip("=")

packaging/windows/internal/vc_env_helper.bat

@@ -1,7 +1,11 @@
 @echo on
 
-set VC_VERSION_LOWER=16
-set VC_VERSION_UPPER=17
+set VC_VERSION_LOWER=17
+set VC_VERSION_UPPER=18
+if "%VC_YEAR%" == "2019" (
+    set VC_VERSION_LOWER=16
+    set VC_VERSION_UPPER=17
+)
 if "%VC_YEAR%" == "2017" (
     set VC_VERSION_LOWER=15
     set VC_VERSION_UPPER=16

packaging/windows/internal/vc_install_helper.sh

@@ -1,9 +1,23 @@
 import torch
-from torchvision.transforms import autoaugment, transforms
 from torchvision.transforms.functional import InterpolationMode
 
 
+def get_module(use_v2):
+    # We need a protected import to avoid the V2 warning in case just V1 is used
+    if use_v2:
+        import torchvision.transforms.v2
+
+        return torchvision.transforms.v2
+    else:
+        import torchvision.transforms
+
+        return torchvision.transforms
+
+
 class ClassificationPresetTrain:
+    # Note: this transform assumes that the input to forward() are always PIL
+    # images, regardless of the backend parameter. We may change that in the
+    # future though, if we change the output type from the dataset.
     def __init__(
         self,
         *,
@@ -17,41 +31,44 @@ def __init__(
         augmix_severity=3,
         random_erase_prob=0.0,
         backend="pil",
+        use_v2=False,
     ):
-        trans = []
+        T = get_module(use_v2)
+
+        transforms = []
         backend = backend.lower()
         if backend == "tensor":
-            trans.append(transforms.PILToTensor())
+            transforms.append(T.PILToTensor())
         elif backend != "pil":
             raise ValueError(f"backend can be 'tensor' or 'pil', but got {backend}")
 
-        trans.append(transforms.RandomResizedCrop(crop_size, interpolation=interpolation, antialias=True))
+        transforms.append(T.RandomResizedCrop(crop_size, interpolation=interpolation, antialias=True))
         if hflip_prob > 0:
-            trans.append(transforms.RandomHorizontalFlip(hflip_prob))
+            transforms.append(T.RandomHorizontalFlip(hflip_prob))
         if auto_augment_policy is not None:
             if auto_augment_policy == "ra":
-                trans.append(autoaugment.RandAugment(interpolation=interpolation, magnitude=ra_magnitude))
+                transforms.append(T.RandAugment(interpolation=interpolation, magnitude=ra_magnitude))
             elif auto_augment_policy == "ta_wide":
-                trans.append(autoaugment.TrivialAugmentWide(interpolation=interpolation))
+                transforms.append(T.TrivialAugmentWide(interpolation=interpolation))
             elif auto_augment_policy == "augmix":
-                trans.append(autoaugment.AugMix(interpolation=interpolation, severity=augmix_severity))
+                transforms.append(T.AugMix(interpolation=interpolation, severity=augmix_severity))
             else:
-                aa_policy = autoaugment.AutoAugmentPolicy(auto_augment_policy)
-                trans.append(autoaugment.AutoAugment(policy=aa_policy, interpolation=interpolation))
+                aa_policy = T.AutoAugmentPolicy(auto_augment_policy)
+                transforms.append(T.AutoAugment(policy=aa_policy, interpolation=interpolation))
 
         if backend == "pil":
-            trans.append(transforms.PILToTensor())
+            transforms.append(T.PILToTensor())
 
-        trans.extend(
+        transforms.extend(
             [
-                transforms.ConvertImageDtype(torch.float),
-                transforms.Normalize(mean=mean, std=std),
+                T.ConvertImageDtype(torch.float),
+                T.Normalize(mean=mean, std=std),
             ]
         )
         if random_erase_prob > 0:
-            trans.append(transforms.RandomErasing(p=random_erase_prob))
+            transforms.append(T.RandomErasing(p=random_erase_prob))
 
-        self.transforms = transforms.Compose(trans)
+        self.transforms = T.Compose(transforms)
 
     def __call__(self, img):
         return self.transforms(img)
@@ -67,28 +84,30 @@ def __init__(
         std=(0.229, 0.224, 0.225),
         interpolation=InterpolationMode.BILINEAR,
         backend="pil",
+        use_v2=False,
     ):
-        trans = []
+        T = get_module(use_v2)
+        transforms = []
         backend = backend.lower()
         if backend == "tensor":
-            trans.append(transforms.PILToTensor())
+            transforms.append(T.PILToTensor())
         elif backend != "pil":
             raise ValueError(f"backend can be 'tensor' or 'pil', but got {backend}")
 
-        trans += [
-            transforms.Resize(resize_size, interpolation=interpolation, antialias=True),
-            transforms.CenterCrop(crop_size),
+        transforms += [
+            T.Resize(resize_size, interpolation=interpolation, antialias=True),
+            T.CenterCrop(crop_size),
         ]
 
         if backend == "pil":
-            trans.append(transforms.PILToTensor())
+            transforms.append(T.PILToTensor())
 
-        trans += [
-            transforms.ConvertImageDtype(torch.float),
-            transforms.Normalize(mean=mean, std=std),
+        transforms += [
+            T.ConvertImageDtype(torch.float),
+            T.Normalize(mean=mean, std=std),
         ]
 
-        self.transforms = transforms.Compose(trans)
+        self.transforms = T.Compose(transforms)
 
     def __call__(self, img):
         return self.transforms(img)

references/classification/presets.py

@@ -145,6 +145,7 @@ def load_data(traindir, valdir, args):
                 ra_magnitude=ra_magnitude,
                 augmix_severity=augmix_severity,
                 backend=args.backend,
+                use_v2=args.use_v2,
             ),
         )
         if args.cache_dataset:
@@ -172,6 +173,7 @@ def load_data(traindir, valdir, args):
                 resize_size=val_resize_size,
                 interpolation=interpolation,
                 backend=args.backend,
+                use_v2=args.use_v2,
             )
 
         dataset_test = torchvision.datasets.ImageFolder(
@@ -516,6 +518,7 @@ def get_args_parser(add_help=True):
     )
     parser.add_argument("--weights", default=None, type=str, help="the weights enum name to load")
     parser.add_argument("--backend", default="PIL", type=str.lower, help="PIL or tensor - case insensitive")
+    parser.add_argument("--use-v2", action="store_true", help="Use V2 transforms")
     return parser
 
 

references/classification/train.py

@@ -7,6 +7,7 @@
 import transforms as T
 from pycocotools import mask as coco_mask
 from pycocotools.coco import COCO
+from torchvision.datasets import wrap_dataset_for_transforms_v2
 
 
 class FilterAndRemapCocoCategories:
@@ -49,7 +50,6 @@ def __call__(self, image, target):
         w, h = image.size
 
         image_id = target["image_id"]
-        image_id = torch.tensor([image_id])
 
         anno = target["annotations"]
 
@@ -126,10 +126,6 @@ def _has_valid_annotation(anno):
             return True
         return False
 
-    if not isinstance(dataset, torchvision.datasets.CocoDetection):
-        raise TypeError(
-            f"This function expects dataset of type torchvision.datasets.CocoDetection, instead  got {type(dataset)}"
-        )
     ids = []
     for ds_idx, img_id in enumerate(dataset.ids):
         ann_ids = dataset.coco.getAnnIds(imgIds=img_id, iscrowd=None)
@@ -196,12 +192,15 @@ def convert_to_coco_api(ds):
 
 
 def get_coco_api_from_dataset(dataset):
+    # FIXME: This is... awful?
     for _ in range(10):
         if isinstance(dataset, torchvision.datasets.CocoDetection):
             break
         if isinstance(dataset, torch.utils.data.Subset):
             dataset = dataset.dataset
-    if isinstance(dataset, torchvision.datasets.CocoDetection):
+    if isinstance(dataset, torchvision.datasets.CocoDetection) or isinstance(
+        getattr(dataset, "_dataset", None), torchvision.datasets.CocoDetection
+    ):
         return dataset.coco
     return convert_to_coco_api(dataset)
 
@@ -220,25 +219,29 @@ def __getitem__(self, idx):
         return img, target
 
 
-def get_coco(root, image_set, transforms, mode="instances"):
+def get_coco(root, image_set, transforms, mode="instances", use_v2=False):
     anno_file_template = "{}_{}2017.json"
     PATHS = {
         "train": ("train2017", os.path.join("annotations", anno_file_template.format(mode, "train"))),
         "val": ("val2017", os.path.join("annotations", anno_file_template.format(mode, "val"))),
         # "train": ("val2017", os.path.join("annotations", anno_file_template.format(mode, "val")))
     }
 
-    t = [ConvertCocoPolysToMask()]
-
-    if transforms is not None:
-        t.append(transforms)
-    transforms = T.Compose(t)
-
     img_folder, ann_file = PATHS[image_set]
     img_folder = os.path.join(root, img_folder)
     ann_file = os.path.join(root, ann_file)
 
-    dataset = CocoDetection(img_folder, ann_file, transforms=transforms)
+    if use_v2:
+        dataset = torchvision.datasets.CocoDetection(img_folder, ann_file, transforms=transforms)
+        # TODO: need to update target_keys to handle masks for segmentation!
+        dataset = wrap_dataset_for_transforms_v2(dataset, target_keys={"boxes", "labels", "image_id"})
+    else:
+        t = [ConvertCocoPolysToMask()]
+        if transforms is not None:
+            t.append(transforms)
+        transforms = T.Compose(t)
+
+        dataset = CocoDetection(img_folder, ann_file, transforms=transforms)
 
     if image_set == "train":
         dataset = _coco_remove_images_without_annotations(dataset)
@@ -248,5 +251,7 @@ def get_coco(root, image_set, transforms, mode="instances"):
     return dataset
 
 
-def get_coco_kp(root, image_set, transforms):
+def get_coco_kp(root, image_set, transforms, use_v2=False):
+    if use_v2:
+        raise ValueError("KeyPoints aren't supported by transforms V2 yet.")
     return get_coco(root, image_set, transforms, mode="person_keypoints")

references/detection/coco_utils.py

@@ -26,7 +26,7 @@ def train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq, sc
 
     for images, targets in metric_logger.log_every(data_loader, print_freq, header):
         images = list(image.to(device) for image in images)
-        targets = [{k: v.to(device) for k, v in t.items()} for t in targets]
+        targets = [{k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in t.items()} for t in targets]
         with torch.cuda.amp.autocast(enabled=scaler is not None):
             loss_dict = model(images, targets)
             losses = sum(loss for loss in loss_dict.values())
@@ -97,7 +97,7 @@ def evaluate(model, data_loader, device):
         outputs = [{k: v.to(cpu_device) for k, v in t.items()} for t in outputs]
         model_time = time.time() - model_time
 
-        res = {target["image_id"].item(): output for target, output in zip(targets, outputs)}
+        res = {target["image_id"]: output for target, output in zip(targets, outputs)}
         evaluator_time = time.time()
         coco_evaluator.update(res)
         evaluator_time = time.time() - evaluator_time