Merge branch 'main' into references/optimizations

Author: datumbox

.circleci/config.yml

@@ -254,6 +254,27 @@ jobs:
             pip install --user pytest
             python test/test_onnx.py
 
+  prototype_test:
+    docker:
+      - image: circleci/python:3.7
+    steps:
+      - run:
+          name: Install torch
+          command: pip install --user --progress-bar=off --pre torch -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
+      - run:
+          name: Install prototype dependencies
+          command: pip install --user --progress-bar=off git+https://github.com/pytorch/data.git
+      - checkout
+      - run:
+          name: Install torchvision
+          command: pip install --user --progress-bar off --no-build-isolation .
+      - run:
+          name: Install test utilities
+          command: pip install --user --progress-bar=off pytest
+      - run:
+          name: Run tests
+          command: pytest test/test_prototype_*.py
+
   binary_linux_wheel:
     <<: *binary_common
     docker:
@@ -977,6 +998,7 @@ workflows:
       - clang_format
       - torchhub_test
       - torch_onnx_test
+      - prototype_test
       {{ ios_workflows() }}
       {{ android_workflows() }}
 
@@ -998,6 +1020,7 @@ workflows:
       - clang_format
       - torchhub_test
       - torch_onnx_test
+      - prototype_test
       {{ ios_workflows(nightly=True) }}
       {{ android_workflows(nightly=True) }}
       {{ build_workflows(prefix="nightly_", filter_branch="nightly", upload=True) }}

.circleci/config.yml.in

@@ -33,7 +33,7 @@ def build_workflows(prefix="", filter_branch=None, upload=False, indentation=6,
             python_versions = PYTHON_VERSIONS
             cu_versions_dict = {
                 "linux": ["cpu", "cu102", "cu111", "cu113", "rocm4.2", "rocm4.3.1"],
-                "win": ["cpu", "cu102", "cu111", "cu113"],
+                "win": ["cpu", "cu111", "cu113"],
                 "macos": ["cpu"],
             }
             cu_versions = cu_versions_dict[os_type]

.circleci/regenerate.py

@@ -26,10 +26,6 @@ fi
 printf "Installing PyTorch with %s\n" "${cudatoolkit}"
 conda install -y -c "pytorch-${UPLOAD_CHANNEL}" "pytorch-${UPLOAD_CHANNEL}"::pytorch "${cudatoolkit}" pytest
 
-printf "Installing torchdata from source"
-pip install git+https://github.com/pytorch/data.git
-
-
 if [ $PYTHON_VERSION == "3.6" ]; then
     printf "Installing minimal PILLOW version\n"
     # Install the minimal PILLOW version. Otherwise, let setup.py install the latest

.circleci/unittest/linux/scripts/install.sh

@@ -7,4 +7,4 @@ conda activate ./env
 
 export PYTORCH_TEST_WITH_SLOW='1'
 python -m torch.utils.collect_env
-pytest --cov=torchvision --junitxml=test-results/junit.xml -v --durations 20 test --ignore=test/test_datasets_download.py
+pytest --cov=torchvision --junitxml=test-results/junit.xml -v --durations 20

.circleci/unittest/linux/scripts/run_test.sh

@@ -29,10 +29,6 @@ printf "Installing PyTorch with %s\n" "${cudatoolkit}"
 # conda-forge channel is required for cudatoolkit 11.1 on Windows, see https://github.com/pytorch/vision/issues/4458
 conda install -y -c "pytorch-${UPLOAD_CHANNEL}" -c conda-forge "pytorch-${UPLOAD_CHANNEL}"::pytorch "${cudatoolkit}" pytest
 
-printf "Installing torchdata from source"
-pip install git+https://github.com/pytorch/data.git
-
-
 if [ $PYTHON_VERSION == "3.6" ]; then
     printf "Installing minimal PILLOW version\n"
     # Install the minimal PILLOW version. Otherwise, let setup.py install the latest

.circleci/unittest/windows/scripts/install.sh

@@ -10,4 +10,4 @@ source "$this_dir/set_cuda_envs.sh"
 
 export PYTORCH_TEST_WITH_SLOW='1'
 python -m torch.utils.collect_env
-pytest --cov=torchvision --junitxml=test-results/junit.xml -v --durations 20 test --ignore=test/test_datasets_download.py
+pytest --cov=torchvision --junitxml=test-results/junit.xml -v --durations 20

.circleci/unittest/windows/scripts/run_test.sh

@@ -21,6 +21,7 @@
     "bc-breaking",
     "deprecation",
     "other",
+    "prototype",
 }
 
 SECONDARY_LABELS = {

.github/process_commit.py

@@ -25,6 +25,8 @@ supported Python versions.
 +--------------------------+--------------------------+---------------------------------+
 | ``1.9.0``                | ``0.10.0``               | ``>=3.6``, ``<=3.9``            |
 +--------------------------+--------------------------+---------------------------------+
+| ``1.8.2``                | ``0.9.2``                | ``>=3.6``, ``<=3.9``            |
++--------------------------+--------------------------+---------------------------------+
 | ``1.8.1``                | ``0.9.1``                | ``>=3.6``, ``<=3.9``            |
 +--------------------------+--------------------------+---------------------------------+
 | ``1.8.0``                | ``0.9.0``                | ``>=3.6``, ``<=3.9``            |
@@ -110,7 +112,9 @@ Video Backend
 =============
 Torchvision currently supports the following video backends:
 
-* [pyav](https://github.com/PyAV-Org/PyAV) (default) - Pythonic binding for ffmpeg libraries. 
+* `pyav`_ (default) - Pythonic binding for ffmpeg libraries.
+
+.. _pyav : https://github.com/PyAV-Org/PyAV
 
 * 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.
 

README.rst

@@ -21,10 +21,6 @@ ignore_errors=True
 
 ignore_errors = True
 
-[mypy-torchvision.models.detection.backbone_utils]
-
-ignore_errors = True
-
 [mypy-torchvision.models.detection.transform]
 
 ignore_errors = True

mypy.ini

@@ -36,7 +36,13 @@ else
   MKL_CONSTRAINT=''
 fi
 
-conda install -yq \pytorch=$PYTORCH_VERSION $CONDA_CUDATOOLKIT_CONSTRAINT $CONDA_CPUONLY_FEATURE $MKL_CONSTRAINT -c "pytorch-${UPLOAD_CHANNEL}"
+if [[ $CONDA_BUILD_VARIANT == "cpu" ]]; then
+  PYTORCH_MUTEX_CONSTRAINT='pytorch-mutex=1.0=cpu'
+else
+  PYTORCH_MUTEX_CONSTRAINT=''
+fi
+
+conda install -yq \pytorch=$PYTORCH_VERSION $CONDA_CUDATOOLKIT_CONSTRAINT $PYTORCH_MUTEX_CONSTRAINT $MKL_CONSTRAINT -c "pytorch-${UPLOAD_CHANNEL}"
 TORCH_PATH=$(dirname $(python -c "import torch; print(torch.__file__)"))
 
 if [[ "$(uname)" == Darwin || "$OSTYPE" == "msys" ]]; then

packaging/build_cmake.sh

@@ -287,9 +287,9 @@ setup_conda_pytorch_constraint() {
 
 # Translate CUDA_VERSION into CUDA_CUDATOOLKIT_CONSTRAINT
 setup_conda_cudatoolkit_constraint() {
-  export CONDA_CPUONLY_FEATURE=""
+  export CONDA_BUILD_VARIANT="cuda"
   if [[ "$(uname)" == Darwin ]]; then
-    export CONDA_CUDATOOLKIT_CONSTRAINT=""
+    export CONDA_BUILD_VARIANT="cpu"
   else
     case "$CU_VERSION" in
       cu113)
@@ -318,7 +318,7 @@ setup_conda_cudatoolkit_constraint() {
         ;;
       cpu)
         export CONDA_CUDATOOLKIT_CONSTRAINT=""
-        export CONDA_CPUONLY_FEATURE="- cpuonly"
+        export CONDA_BUILD_VARIANT="cpu"
         ;;
       *)
         echo "Unrecognized CU_VERSION=$CU_VERSION"
@@ -329,10 +329,10 @@ setup_conda_cudatoolkit_constraint() {
 }
 
 setup_conda_cudatoolkit_plain_constraint() {
-  export CONDA_CPUONLY_FEATURE=""
+  export CONDA_BUILD_VARIANT="cuda"
   export CMAKE_USE_CUDA=1
   if [[ "$(uname)" == Darwin ]]; then
-    export CONDA_CUDATOOLKIT_CONSTRAINT=""
+    export CONDA_BUILD_VARIANT="cpu"
     export CMAKE_USE_CUDA=0
   else
     case "$CU_VERSION" in
@@ -359,7 +359,7 @@ setup_conda_cudatoolkit_plain_constraint() {
         ;;
       cpu)
         export CONDA_CUDATOOLKIT_CONSTRAINT=""
-        export CONDA_CPUONLY_FEATURE="cpuonly"
+        export CONDA_BUILD_VARIANT="cpu"
         export CMAKE_USE_CUDA=0
         ;;
       *)

packaging/pkg_helpers.bash

@@ -1,3 +1,4 @@
+{% set build_variant = environ.get('CONDA_BUILD_VARIANT', 'cpu') %}
 package:
   name: torchvision
   version: "{{ environ.get('BUILD_VERSION') }}"
@@ -16,9 +17,9 @@ requirements:
   host:
     - python
     - setuptools
+    - pytorch-mutex 1.0 {{ build_variant }}  # [not osx ]
     {{ environ.get('CONDA_PYTORCH_BUILD_CONSTRAINT') }}
-    {{ environ.get('CONDA_CUDATOOLKIT_CONSTRAINT') }}
-    {{ environ.get('CONDA_CPUONLY_FEATURE') }}
+    {{ environ.get('CONDA_CUDATOOLKIT_CONSTRAINT', '') }}
 
   run:
     - python
@@ -27,8 +28,17 @@ requirements:
     - ffmpeg >=4.2  # [not win]
     - jpeg
     - pillow >=5.3.0
+    - pytorch-mutex 1.0 {{ build_variant }}  # [not osx ]
     {{ environ.get('CONDA_PYTORCH_CONSTRAINT') }}
-    {{ environ.get('CONDA_CUDATOOLKIT_CONSTRAINT') }}
+    {{ environ.get('CONDA_CUDATOOLKIT_CONSTRAINT', '') }}
+
+  {% if build_variant == 'cpu' %}
+  run_constrained:
+    - cpuonly
+  {% elif not osx %}
+  run_constrained:
+     - cpuonly <0
+  {% endif %}
 
 build:
   string: py{{py}}_{{ environ['CU_VERSION'] }}
@@ -38,8 +48,6 @@ build:
     - FORCE_CUDA
     - BUILD_VERSION
     - TORCH_CUDA_ARCH_LIST
-  features:
-    {{ environ.get('CONDA_CPUONLY_FEATURE') }}
 
 test:
   imports:

packaging/torchvision/meta.yaml

@@ -6,5 +6,7 @@ addopts =
     --tb=native
     # enable all warnings
     -Wd
+    --ignore=test/test_datasets_download.py
+    --ignore-glob=test/test_prototype_*.py
 testpaths =
     test

pytest.ini

@@ -58,7 +58,7 @@ def evaluate(model, criterion, data_loader, device, print_freq=100, log_suffix="
     header = f"Test: {log_suffix}"
 
     num_processed_samples = 0
-    with torch.no_grad():
+    with torch.inference_mode():
         for image, target in metric_logger.log_every(data_loader, print_freq, header):
             image = image.to(device, non_blocking=True)
             target = target.to(device, non_blocking=True)

references/classification/train.py

@@ -112,7 +112,7 @@ def main(args):
         print("Starting training for epoch", epoch)
         train_one_epoch(model, criterion, optimizer, data_loader, device, epoch, args.print_freq)
         lr_scheduler.step()
-        with torch.no_grad():
+        with torch.inference_mode():
             if epoch >= args.num_observer_update_epochs:
                 print("Disabling observer for subseq epochs, epoch = ", epoch)
                 model.apply(torch.quantization.disable_observer)

references/classification/train_quantization.py

@@ -181,7 +181,7 @@ def update_parameters(self, model):
 
 def accuracy(output, target, topk=(1,)):
     """Computes the accuracy over the k top predictions for the specified values of k"""
-    with torch.no_grad():
+    with torch.inference_mode():
         maxk = max(topk)
         batch_size = target.size(0)
         if target.ndim == 2:

references/classification/utils.py

@@ -68,7 +68,7 @@ def _get_iou_types(model):
     return iou_types
 
 
-@torch.no_grad()
+@torch.inference_mode()
 def evaluate(model, data_loader, device):
     n_threads = torch.get_num_threads()
     # FIXME remove this and make paste_masks_in_image run on the GPU

references/detection/engine.py

@@ -95,7 +95,7 @@ def reduce_dict(input_dict, average=True):
     world_size = get_world_size()
     if world_size < 2:
         return input_dict
-    with torch.no_grad():
+    with torch.inference_mode():
         names = []
         values = []
         # sort the keys so that they are consistent across processes

references/detection/utils.py

@@ -49,7 +49,7 @@ def evaluate(model, data_loader, device, num_classes):
     confmat = utils.ConfusionMatrix(num_classes)
     metric_logger = utils.MetricLogger(delimiter="  ")
     header = "Test:"
-    with torch.no_grad():
+    with torch.inference_mode():
         for image, target in metric_logger.log_every(data_loader, 100, header):
             image, target = image.to(device), target.to(device)
             output = model(image)

references/segmentation/train.py

@@ -76,7 +76,7 @@ def update(self, a, b):
         n = self.num_classes
         if self.mat is None:
             self.mat = torch.zeros((n, n), dtype=torch.int64, device=a.device)
-        with torch.no_grad():
+        with torch.inference_mode():
             k = (a >= 0) & (a < n)
             inds = n * a[k].to(torch.int64) + b[k]
             self.mat += torch.bincount(inds, minlength=n ** 2).reshape(n, n)

references/segmentation/utils.py

@@ -51,7 +51,7 @@ def find_best_threshold(dists, targets, device):
     return best_thresh, accuracy
 
 
-@torch.no_grad()
+@torch.inference_mode()
 def evaluate(model, loader, device):
     model.eval()
     embeds, labels = [], []

references/similarity/train.py

@@ -52,7 +52,7 @@ def evaluate(model, criterion, data_loader, device):
     model.eval()
     metric_logger = utils.MetricLogger(delimiter="  ")
     header = "Test:"
-    with torch.no_grad():
+    with torch.inference_mode():
         for video, target in metric_logger.log_every(data_loader, 100, header):
             video = video.to(device, non_blocking=True)
             target = target.to(device, non_blocking=True)

references/video_classification/train.py

@@ -159,7 +159,7 @@ def log_every(self, iterable, print_freq, header=None):
 
 def accuracy(output, target, topk=(1,)):
     """Computes the accuracy over the k top predictions for the specified values of k"""
-    with torch.no_grad():
+    with torch.inference_mode():
         maxk = max(topk)
         batch_size = target.size(0)
 

references/video_classification/utils.py

@@ -466,7 +466,7 @@ def run(self):
         license="BSD",
         # Package info
         packages=find_packages(exclude=("test",)),
-        package_data={package_name: ["*.dll", "*.dylib", "*.so", "*.categories"]},
+        package_data={package_name: ["*.dll", "*.dylib", "*.so", "prototype/datasets/_builtin/*.categories"]},
         zip_safe=False,
         install_requires=requirements,
         extras_require={

setup.py

@@ -23,7 +23,7 @@
 
 
 def _create_video_frames(num_frames, height, width):
-    y, x = torch.meshgrid(torch.linspace(-2, 2, height), torch.linspace(-2, 2, width))
+    y, x = torch.meshgrid(torch.linspace(-2, 2, height), torch.linspace(-2, 2, width), indexing="ij")
     data = []
     for i in range(num_frames):
         xc = float(i) / num_frames