Add files using upload-large-folder tool

Leffa/preprocess/humanparsing/mhp_extension/detectron2/.circleci/config.yml

@@ -0,0 +1,179 @@
+# Python CircleCI 2.0 configuration file
+#
+# Check https://circleci.com/docs/2.0/language-python/ for more details
+#
+version: 2
+
+# -------------------------------------------------------------------------------------
+# Environments to run the jobs in
+# -------------------------------------------------------------------------------------
+cpu: &cpu
+  docker:
+    - image: circleci/python:3.6.8-stretch
+  resource_class: medium
+
+gpu: &gpu
+  machine:
+    image: ubuntu-1604:201903-01
+    docker_layer_caching: true
+  resource_class: gpu.small
+
+# -------------------------------------------------------------------------------------
+# Re-usable commands
+# -------------------------------------------------------------------------------------
+install_python: &install_python
+  - run:
+      name: Install Python
+      working_directory: ~/
+      command: |
+        pyenv install 3.6.1
+        pyenv global 3.6.1
+
+setup_venv: &setup_venv
+  - run:
+      name: Setup Virtual Env
+      working_directory: ~/
+      command: |
+        python -m venv ~/venv
+        echo ". ~/venv/bin/activate" >> $BASH_ENV
+        . ~/venv/bin/activate
+        python --version
+        which python
+        which pip
+        pip install --upgrade pip
+
+install_dep: &install_dep
+  - run:
+      name: Install Dependencies
+      command: |
+        pip install --progress-bar off -U 'git+https://github.com/facebookresearch/fvcore'
+        pip install --progress-bar off cython opencv-python
+        pip install --progress-bar off 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
+        pip install --progress-bar off torch torchvision
+
+install_detectron2: &install_detectron2
+  - run:
+      name: Install Detectron2
+      command: |
+        gcc --version
+        pip install -U --progress-bar off -e .[dev]
+        python -m detectron2.utils.collect_env
+
+install_nvidia_driver: &install_nvidia_driver
+  - run:
+      name: Install nvidia driver
+      working_directory: ~/
+      command: |
+        wget -q 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
+        sudo /bin/bash ./NVIDIA-Linux-x86_64-430.40.run -s --no-drm
+        nvidia-smi
+
+run_unittests: &run_unittests
+  - run:
+      name: Run Unit Tests
+      command: |
+        python -m unittest discover -v -s tests
+
+# -------------------------------------------------------------------------------------
+# Jobs to run
+# -------------------------------------------------------------------------------------
+jobs:
+  cpu_tests:
+    <<: *cpu
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+      - <<: *setup_venv
+
+      # Cache the venv directory that contains dependencies
+      - restore_cache:
+          keys:
+            - cache-key-{{ .Branch }}-ID-20200425
+
+      - <<: *install_dep
+
+      - save_cache:
+          paths:
+            - ~/venv
+          key: cache-key-{{ .Branch }}-ID-20200425
+
+      - <<: *install_detectron2
+
+      - run:
+          name: isort
+          command: |
+            isort -c -sp .
+      - run:
+          name: black
+          command: |
+            black --check -l 100 .
+      - run:
+          name: flake8
+          command: |
+            flake8 .
+
+      - <<: *run_unittests
+
+  gpu_tests:
+    <<: *gpu
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+      - <<: *install_nvidia_driver
+
+      - run:
+          name: Install nvidia-docker
+          working_directory: ~/
+          command: |
+            curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
+            distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
+            curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
+            sudo tee /etc/apt/sources.list.d/nvidia-docker.list
+            sudo apt-get update && sudo apt-get install -y nvidia-docker2
+            # reload the docker daemon configuration
+            sudo pkill -SIGHUP dockerd
+
+      - run:
+          name: Launch docker
+          working_directory: ~/detectron2/docker
+          command: |
+            nvidia-docker build -t detectron2:v0 -f Dockerfile-circleci .
+            nvidia-docker run -itd --name d2 detectron2:v0
+            docker exec -it d2 nvidia-smi
+
+      - run:
+          name: Build Detectron2
+          command: |
+            docker exec -it d2 pip install 'git+https://github.com/facebookresearch/fvcore'
+            docker cp ~/detectron2 d2:/detectron2
+            # This will build d2 for the target GPU arch only
+            docker exec -it d2 pip install -e /detectron2
+            docker exec -it d2 python3 -m detectron2.utils.collect_env
+            docker exec -it d2 python3 -c 'import torch; assert(torch.cuda.is_available())'
+
+      - run:
+          name: Run Unit Tests
+          command: |
+            docker exec -e CIRCLECI=true -it d2 python3 -m unittest discover -v -s /detectron2/tests
+
+workflows:
+  version: 2
+  regular_test:
+    jobs:
+      - cpu_tests
+      - gpu_tests
+
+  #nightly_test:
+    #jobs:
+      #- gpu_tests
+    #triggers:
+      #- schedule:
+          #cron: "0 0 * * *"
+          #filters:
+            #branches:
+              #only:
+                #- master

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/config/__init__.py

@@ -0,0 +1,13 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .compat import downgrade_config, upgrade_config
+from .config import CfgNode, get_cfg, global_cfg, set_global_cfg, configurable
+
+__all__ = [
+    "CfgNode",
+    "get_cfg",
+    "global_cfg",
+    "set_global_cfg",
+    "downgrade_config",
+    "upgrade_config",
+    "configurable",
+]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/config/compat.py

@@ -0,0 +1,229 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Backward compatibility of configs.
+
+Instructions to bump version:
++ It's not needed to bump version if new keys are added.
+  It's only needed when backward-incompatible changes happen
+  (i.e., some existing keys disappear, or the meaning of a key changes)
++ To bump version, do the following:
+    1. Increment _C.VERSION in defaults.py
+    2. Add a converter in this file.
+
+      Each ConverterVX has a function "upgrade" which in-place upgrades config from X-1 to X,
+      and a function "downgrade" which in-place downgrades config from X to X-1
+
+      In each function, VERSION is left unchanged.
+
+      Each converter assumes that its input has the relevant keys
+      (i.e., the input is not a partial config).
+    3. Run the tests (test_config.py) to make sure the upgrade & downgrade
+       functions are consistent.
+"""
+
+import logging
+from typing import List, Optional, Tuple
+
+from .config import CfgNode as CN
+from .defaults import _C
+
+__all__ = ["upgrade_config", "downgrade_config"]
+
+
+def upgrade_config(cfg: CN, to_version: Optional[int] = None) -> CN:
+    """
+    Upgrade a config from its current version to a newer version.
+
+    Args:
+        cfg (CfgNode):
+        to_version (int): defaults to the latest version.
+    """
+    cfg = cfg.clone()
+    if to_version is None:
+        to_version = _C.VERSION
+
+    assert cfg.VERSION <= to_version, "Cannot upgrade from v{} to v{}!".format(
+        cfg.VERSION, to_version
+    )
+    for k in range(cfg.VERSION, to_version):
+        converter = globals()["ConverterV" + str(k + 1)]
+        converter.upgrade(cfg)
+        cfg.VERSION = k + 1
+    return cfg
+
+
+def downgrade_config(cfg: CN, to_version: int) -> CN:
+    """
+    Downgrade a config from its current version to an older version.
+
+    Args:
+        cfg (CfgNode):
+        to_version (int):
+
+    Note:
+        A general downgrade of arbitrary configs is not always possible due to the
+        different functionalities in different versions.
+        The purpose of downgrade is only to recover the defaults in old versions,
+        allowing it to load an old partial yaml config.
+        Therefore, the implementation only needs to fill in the default values
+        in the old version when a general downgrade is not possible.
+    """
+    cfg = cfg.clone()
+    assert cfg.VERSION >= to_version, "Cannot downgrade from v{} to v{}!".format(
+        cfg.VERSION, to_version
+    )
+    for k in range(cfg.VERSION, to_version, -1):
+        converter = globals()["ConverterV" + str(k)]
+        converter.downgrade(cfg)
+        cfg.VERSION = k - 1
+    return cfg
+
+
+def guess_version(cfg: CN, filename: str) -> int:
+    """
+    Guess the version of a partial config where the VERSION field is not specified.
+    Returns the version, or the latest if cannot make a guess.
+
+    This makes it easier for users to migrate.
+    """
+    logger = logging.getLogger(__name__)
+
+    def _has(name: str) -> bool:
+        cur = cfg
+        for n in name.split("."):
+            if n not in cur:
+                return False
+            cur = cur[n]
+        return True
+
+    # Most users' partial configs have "MODEL.WEIGHT", so guess on it
+    ret = None
+    if _has("MODEL.WEIGHT") or _has("TEST.AUG_ON"):
+        ret = 1
+
+    if ret is not None:
+        logger.warning("Config '{}' has no VERSION. Assuming it to be v{}.".format(filename, ret))
+    else:
+        ret = _C.VERSION
+        logger.warning(
+            "Config '{}' has no VERSION. Assuming it to be compatible with latest v{}.".format(
+                filename, ret
+            )
+        )
+    return ret
+
+
+def _rename(cfg: CN, old: str, new: str) -> None:
+    old_keys = old.split(".")
+    new_keys = new.split(".")
+
+    def _set(key_seq: List[str], val: str) -> None:
+        cur = cfg
+        for k in key_seq[:-1]:
+            if k not in cur:
+                cur[k] = CN()
+            cur = cur[k]
+        cur[key_seq[-1]] = val
+
+    def _get(key_seq: List[str]) -> CN:
+        cur = cfg
+        for k in key_seq:
+            cur = cur[k]
+        return cur
+
+    def _del(key_seq: List[str]) -> None:
+        cur = cfg
+        for k in key_seq[:-1]:
+            cur = cur[k]
+        del cur[key_seq[-1]]
+        if len(cur) == 0 and len(key_seq) > 1:
+            _del(key_seq[:-1])
+
+    _set(new_keys, _get(old_keys))
+    _del(old_keys)
+
+
+class _RenameConverter:
+    """
+    A converter that handles simple rename.
+    """
+
+    RENAME: List[Tuple[str, str]] = []  # list of tuples of (old name, new name)
+
+    @classmethod
+    def upgrade(cls, cfg: CN) -> None:
+        for old, new in cls.RENAME:
+            _rename(cfg, old, new)
+
+    @classmethod
+    def downgrade(cls, cfg: CN) -> None:
+        for old, new in cls.RENAME[::-1]:
+            _rename(cfg, new, old)
+
+
+class ConverterV1(_RenameConverter):
+    RENAME = [("MODEL.RPN_HEAD.NAME", "MODEL.RPN.HEAD_NAME")]
+
+
+class ConverterV2(_RenameConverter):
+    """
+    A large bulk of rename, before public release.
+    """
+
+    RENAME = [
+        ("MODEL.WEIGHT", "MODEL.WEIGHTS"),
+        ("MODEL.PANOPTIC_FPN.SEMANTIC_LOSS_SCALE", "MODEL.SEM_SEG_HEAD.LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.RPN_LOSS_SCALE", "MODEL.RPN.LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.INSTANCE_LOSS_SCALE", "MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.COMBINE_ON", "MODEL.PANOPTIC_FPN.COMBINE.ENABLED"),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_OVERLAP_THRESHOLD",
+            "MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH",
+        ),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_STUFF_AREA_LIMIT",
+            "MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT",
+        ),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_INSTANCES_CONFIDENCE_THRESHOLD",
+            "MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH",
+        ),
+        ("MODEL.ROI_HEADS.SCORE_THRESH", "MODEL.ROI_HEADS.SCORE_THRESH_TEST"),
+        ("MODEL.ROI_HEADS.NMS", "MODEL.ROI_HEADS.NMS_THRESH_TEST"),
+        ("MODEL.RETINANET.INFERENCE_SCORE_THRESHOLD", "MODEL.RETINANET.SCORE_THRESH_TEST"),
+        ("MODEL.RETINANET.INFERENCE_TOPK_CANDIDATES", "MODEL.RETINANET.TOPK_CANDIDATES_TEST"),
+        ("MODEL.RETINANET.INFERENCE_NMS_THRESHOLD", "MODEL.RETINANET.NMS_THRESH_TEST"),
+        ("TEST.DETECTIONS_PER_IMG", "TEST.DETECTIONS_PER_IMAGE"),
+        ("TEST.AUG_ON", "TEST.AUG.ENABLED"),
+        ("TEST.AUG_MIN_SIZES", "TEST.AUG.MIN_SIZES"),
+        ("TEST.AUG_MAX_SIZE", "TEST.AUG.MAX_SIZE"),
+        ("TEST.AUG_FLIP", "TEST.AUG.FLIP"),
+    ]
+
+    @classmethod
+    def upgrade(cls, cfg: CN) -> None:
+        super().upgrade(cfg)
+
+        if cfg.MODEL.META_ARCHITECTURE == "RetinaNet":
+            _rename(
+                cfg, "MODEL.RETINANET.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS"
+            )
+            _rename(cfg, "MODEL.RETINANET.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
+            del cfg["MODEL"]["RPN"]["ANCHOR_SIZES"]
+            del cfg["MODEL"]["RPN"]["ANCHOR_ASPECT_RATIOS"]
+        else:
+            _rename(cfg, "MODEL.RPN.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS")
+            _rename(cfg, "MODEL.RPN.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
+            del cfg["MODEL"]["RETINANET"]["ANCHOR_SIZES"]
+            del cfg["MODEL"]["RETINANET"]["ANCHOR_ASPECT_RATIOS"]
+        del cfg["MODEL"]["RETINANET"]["ANCHOR_STRIDES"]
+
+    @classmethod
+    def downgrade(cls, cfg: CN) -> None:
+        super().downgrade(cfg)
+
+        _rename(cfg, "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS", "MODEL.RPN.ANCHOR_ASPECT_RATIOS")
+        _rename(cfg, "MODEL.ANCHOR_GENERATOR.SIZES", "MODEL.RPN.ANCHOR_SIZES")
+        cfg.MODEL.RETINANET.ANCHOR_ASPECT_RATIOS = cfg.MODEL.RPN.ANCHOR_ASPECT_RATIOS
+        cfg.MODEL.RETINANET.ANCHOR_SIZES = cfg.MODEL.RPN.ANCHOR_SIZES
+        cfg.MODEL.RETINANET.ANCHOR_STRIDES = []  # this is not used anywhere in any version

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/config/config.py

@@ -0,0 +1,202 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import functools
+import inspect
+import logging
+from fvcore.common.config import CfgNode as _CfgNode
+from fvcore.common.file_io import PathManager
+
+
+class CfgNode(_CfgNode):
+    """
+    The same as `fvcore.common.config.CfgNode`, but different in:
+
+    1. Use unsafe yaml loading by default.
+       Note that this may lead to arbitrary code execution: you must not
+       load a config file from untrusted sources before manually inspecting
+       the content of the file.
+    2. Support config versioning.
+       When attempting to merge an old config, it will convert the old config automatically.
+    """
+
+    # Note that the default value of allow_unsafe is changed to True
+    def merge_from_file(self, cfg_filename: str, allow_unsafe: bool = True) -> None:
+        assert PathManager.isfile(cfg_filename), f"Config file '{cfg_filename}' does not exist!"
+        loaded_cfg = _CfgNode.load_yaml_with_base(cfg_filename, allow_unsafe=allow_unsafe)
+        loaded_cfg = type(self)(loaded_cfg)
+
+        # defaults.py needs to import CfgNode
+        from .defaults import _C
+
+        latest_ver = _C.VERSION
+        assert (
+            latest_ver == self.VERSION
+        ), "CfgNode.merge_from_file is only allowed on a config object of latest version!"
+
+        logger = logging.getLogger(__name__)
+
+        loaded_ver = loaded_cfg.get("VERSION", None)
+        if loaded_ver is None:
+            from .compat import guess_version
+
+            loaded_ver = guess_version(loaded_cfg, cfg_filename)
+        assert loaded_ver <= self.VERSION, "Cannot merge a v{} config into a v{} config.".format(
+            loaded_ver, self.VERSION
+        )
+
+        if loaded_ver == self.VERSION:
+            self.merge_from_other_cfg(loaded_cfg)
+        else:
+            # compat.py needs to import CfgNode
+            from .compat import upgrade_config, downgrade_config
+
+            logger.warning(
+                "Loading an old v{} config file '{}' by automatically upgrading to v{}. "
+                "See docs/CHANGELOG.md for instructions to update your files.".format(
+                    loaded_ver, cfg_filename, self.VERSION
+                )
+            )
+            # To convert, first obtain a full config at an old version
+            old_self = downgrade_config(self, to_version=loaded_ver)
+            old_self.merge_from_other_cfg(loaded_cfg)
+            new_config = upgrade_config(old_self)
+            self.clear()
+            self.update(new_config)
+
+    def dump(self, *args, **kwargs):
+        """
+        Returns:
+            str: a yaml string representation of the config
+        """
+        # to make it show up in docs
+        return super().dump(*args, **kwargs)
+
+
+global_cfg = CfgNode()
+
+
+def get_cfg() -> CfgNode:
+    """
+    Get a copy of the default config.
+
+    Returns:
+        a detectron2 CfgNode instance.
+    """
+    from .defaults import _C
+
+    return _C.clone()
+
+
+def set_global_cfg(cfg: CfgNode) -> None:
+    """
+    Let the global config point to the given cfg.
+
+    Assume that the given "cfg" has the key "KEY", after calling
+    `set_global_cfg(cfg)`, the key can be accessed by:
+
+    .. code-block:: python
+
+        from detectron2.config import global_cfg
+        print(global_cfg.KEY)
+
+    By using a hacky global config, you can access these configs anywhere,
+    without having to pass the config object or the values deep into the code.
+    This is a hacky feature introduced for quick prototyping / research exploration.
+    """
+    global global_cfg
+    global_cfg.clear()
+    global_cfg.update(cfg)
+
+
+def configurable(init_func):
+    """
+    Decorate a class's __init__ method so that it can be called with a CfgNode
+    object using the class's from_config classmethod.
+
+    Examples:
+
+    .. code-block:: python
+
+        class A:
+            @configurable
+            def __init__(self, a, b=2, c=3):
+                pass
+
+            @classmethod
+            def from_config(cls, cfg):
+                # Returns kwargs to be passed to __init__
+                return {"a": cfg.A, "b": cfg.B}
+
+        a1 = A(a=1, b=2)  # regular construction
+        a2 = A(cfg)       # construct with a cfg
+        a3 = A(cfg, b=3, c=4)  # construct with extra overwrite
+    """
+    assert init_func.__name__ == "__init__", "@configurable should only be used for __init__!"
+    if init_func.__module__.startswith("detectron2."):
+        assert (
+            init_func.__doc__ is not None and "experimental" in init_func.__doc__
+        ), f"configurable {init_func} should be marked experimental"
+
+    @functools.wraps(init_func)
+    def wrapped(self, *args, **kwargs):
+        try:
+            from_config_func = type(self).from_config
+        except AttributeError:
+            raise AttributeError("Class with @configurable must have a 'from_config' classmethod.")
+        if not inspect.ismethod(from_config_func):
+            raise TypeError("Class with @configurable must have a 'from_config' classmethod.")
+
+        if _called_with_cfg(*args, **kwargs):
+            explicit_args = _get_args_from_config(from_config_func, *args, **kwargs)
+            init_func(self, **explicit_args)
+        else:
+            init_func(self, *args, **kwargs)
+
+    return wrapped
+
+
+def _get_args_from_config(from_config_func, *args, **kwargs):
+    """
+    Use `from_config` to obtain explicit arguments.
+
+    Returns:
+        dict: arguments to be used for cls.__init__
+    """
+    signature = inspect.signature(from_config_func)
+    if list(signature.parameters.keys())[0] != "cfg":
+        raise TypeError(
+            f"{from_config_func.__self__}.from_config must take 'cfg' as the first argument!"
+        )
+    support_var_arg = any(
+        param.kind in [param.VAR_POSITIONAL, param.VAR_KEYWORD]
+        for param in signature.parameters.values()
+    )
+    if support_var_arg:  # forward all arguments to from_config, if from_config accepts them
+        ret = from_config_func(*args, **kwargs)
+    else:
+        # forward supported arguments to from_config
+        supported_arg_names = set(signature.parameters.keys())
+        extra_kwargs = {}
+        for name in list(kwargs.keys()):
+            if name not in supported_arg_names:
+                extra_kwargs[name] = kwargs.pop(name)
+        ret = from_config_func(*args, **kwargs)
+        # forward the other arguments to __init__
+        ret.update(extra_kwargs)
+    return ret
+
+
+def _called_with_cfg(*args, **kwargs):
+    """
+    Returns:
+        bool: whether the arguments contain CfgNode and should be considered
+            forwarded to from_config.
+    """
+    if len(args) and isinstance(args[0], _CfgNode):
+        return True
+    if isinstance(kwargs.pop("cfg", None), _CfgNode):
+        return True
+    # `from_config`'s first argument is forced to be "cfg".
+    # So the above check covers all cases.
+    return False

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/config/defaults.py

@@ -0,0 +1,598 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .config import CfgNode as CN
+
+# -----------------------------------------------------------------------------
+# Convention about Training / Test specific parameters
+# -----------------------------------------------------------------------------
+# Whenever an argument can be either used for training or for testing, the
+# corresponding name will be post-fixed by a _TRAIN for a training parameter,
+# or _TEST for a test-specific parameter.
+# For example, the number of images during training will be
+# IMAGES_PER_BATCH_TRAIN, while the number of images for testing will be
+# IMAGES_PER_BATCH_TEST
+
+# -----------------------------------------------------------------------------
+# Config definition
+# -----------------------------------------------------------------------------
+
+_C = CN()
+
+# The version number, to upgrade from old configs to new ones if any
+# changes happen. It's recommended to keep a VERSION in your config file.
+_C.VERSION = 2
+
+_C.MODEL = CN()
+_C.MODEL.LOAD_PROPOSALS = False
+_C.MODEL.MASK_ON = False
+_C.MODEL.KEYPOINT_ON = False
+_C.MODEL.DEVICE = "cuda"
+_C.MODEL.META_ARCHITECTURE = "GeneralizedRCNN"
+
+# Path (possibly with schema like catalog:// or detectron2://) to a checkpoint file
+# to be loaded to the model. You can find available models in the model zoo.
+_C.MODEL.WEIGHTS = ""
+
+# Values to be used for image normalization (BGR order, since INPUT.FORMAT defaults to BGR).
+# To train on images of different number of channels, just set different mean & std.
+# Default values are the mean pixel value from ImageNet: [103.53, 116.28, 123.675]
+_C.MODEL.PIXEL_MEAN = [103.530, 116.280, 123.675]
+# When using pre-trained models in Detectron1 or any MSRA models,
+# std has been absorbed into its conv1 weights, so the std needs to be set 1.
+# Otherwise, you can use [57.375, 57.120, 58.395] (ImageNet std)
+_C.MODEL.PIXEL_STD = [1.0, 1.0, 1.0]
+
+
+# -----------------------------------------------------------------------------
+# INPUT
+# -----------------------------------------------------------------------------
+_C.INPUT = CN()
+# Size of the smallest side of the image during training
+_C.INPUT.MIN_SIZE_TRAIN = (800,)
+# Sample size of smallest side by choice or random selection from range give by
+# INPUT.MIN_SIZE_TRAIN
+_C.INPUT.MIN_SIZE_TRAIN_SAMPLING = "choice"
+# Maximum size of the side of the image during training
+_C.INPUT.MAX_SIZE_TRAIN = 1333
+# Size of the smallest side of the image during testing. Set to zero to disable resize in testing.
+_C.INPUT.MIN_SIZE_TEST = 800
+# Maximum size of the side of the image during testing
+_C.INPUT.MAX_SIZE_TEST = 1333
+
+# `True` if cropping is used for data augmentation during training
+_C.INPUT.CROP = CN({"ENABLED": False})
+# Cropping type:
+# - "relative" crop (H * CROP.SIZE[0], W * CROP.SIZE[1]) part of an input of size (H, W)
+# - "relative_range" uniformly sample relative crop size from between [CROP.SIZE[0], [CROP.SIZE[1]].
+#   and  [1, 1] and use it as in "relative" scenario.
+# - "absolute" crop part of an input with absolute size: (CROP.SIZE[0], CROP.SIZE[1]).
+_C.INPUT.CROP.TYPE = "relative_range"
+# Size of crop in range (0, 1] if CROP.TYPE is "relative" or "relative_range" and in number of
+# pixels if CROP.TYPE is "absolute"
+_C.INPUT.CROP.SIZE = [0.9, 0.9]
+
+
+# Whether the model needs RGB, YUV, HSV etc.
+# Should be one of the modes defined here, as we use PIL to read the image:
+# https://pillow.readthedocs.io/en/stable/handbook/concepts.html#concept-modes
+# with BGR being the one exception. One can set image format to BGR, we will
+# internally use RGB for conversion and flip the channels over
+_C.INPUT.FORMAT = "BGR"
+# The ground truth mask format that the model will use.
+# Mask R-CNN supports either "polygon" or "bitmask" as ground truth.
+_C.INPUT.MASK_FORMAT = "polygon"  # alternative: "bitmask"
+
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+_C.DATASETS = CN()
+# List of the dataset names for training. Must be registered in DatasetCatalog
+_C.DATASETS.TRAIN = ()
+# List of the pre-computed proposal files for training, which must be consistent
+# with data listed in DATASETS.TRAIN.
+_C.DATASETS.PROPOSAL_FILES_TRAIN = ()
+# Number of top scoring precomputed proposals to keep for training
+_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN = 2000
+# List of the dataset names for testing. Must be registered in DatasetCatalog
+_C.DATASETS.TEST = ()
+# List of the pre-computed proposal files for test, which must be consistent
+# with data listed in DATASETS.TEST.
+_C.DATASETS.PROPOSAL_FILES_TEST = ()
+# Number of top scoring precomputed proposals to keep for test
+_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST = 1000
+
+# -----------------------------------------------------------------------------
+# DataLoader
+# -----------------------------------------------------------------------------
+_C.DATALOADER = CN()
+# Number of data loading threads
+_C.DATALOADER.NUM_WORKERS = 4
+# If True, each batch should contain only images for which the aspect ratio
+# is compatible. This groups portrait images together, and landscape images
+# are not batched with portrait images.
+_C.DATALOADER.ASPECT_RATIO_GROUPING = True
+# Options: TrainingSampler, RepeatFactorTrainingSampler
+_C.DATALOADER.SAMPLER_TRAIN = "TrainingSampler"
+# Repeat threshold for RepeatFactorTrainingSampler
+_C.DATALOADER.REPEAT_THRESHOLD = 0.0
+# if True, the dataloader will filter out images that have no associated
+# annotations at train time.
+_C.DATALOADER.FILTER_EMPTY_ANNOTATIONS = True
+
+# ---------------------------------------------------------------------------- #
+# Backbone options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.BACKBONE = CN()
+
+_C.MODEL.BACKBONE.NAME = "build_resnet_backbone"
+# Freeze the first several stages so they are not trained.
+# There are 5 stages in ResNet. The first is a convolution, and the following
+# stages are each group of residual blocks.
+_C.MODEL.BACKBONE.FREEZE_AT = 2
+
+
+# ---------------------------------------------------------------------------- #
+# FPN options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.FPN = CN()
+# Names of the input feature maps to be used by FPN
+# They must have contiguous power of 2 strides
+# e.g., ["res2", "res3", "res4", "res5"]
+_C.MODEL.FPN.IN_FEATURES = []
+_C.MODEL.FPN.OUT_CHANNELS = 256
+
+# Options: "" (no norm), "GN"
+_C.MODEL.FPN.NORM = ""
+
+# Types for fusing the FPN top-down and lateral features. Can be either "sum" or "avg"
+_C.MODEL.FPN.FUSE_TYPE = "sum"
+
+
+# ---------------------------------------------------------------------------- #
+# Proposal generator options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.PROPOSAL_GENERATOR = CN()
+# Current proposal generators include "RPN", "RRPN" and "PrecomputedProposals"
+_C.MODEL.PROPOSAL_GENERATOR.NAME = "RPN"
+# Proposal height and width both need to be greater than MIN_SIZE
+# (a the scale used during training or inference)
+_C.MODEL.PROPOSAL_GENERATOR.MIN_SIZE = 0
+
+
+# ---------------------------------------------------------------------------- #
+# Anchor generator options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ANCHOR_GENERATOR = CN()
+# The generator can be any name in the ANCHOR_GENERATOR registry
+_C.MODEL.ANCHOR_GENERATOR.NAME = "DefaultAnchorGenerator"
+# Anchor sizes (i.e. sqrt of area) in absolute pixels w.r.t. the network input.
+# Format: list[list[float]]. SIZES[i] specifies the list of sizes
+# to use for IN_FEATURES[i]; len(SIZES) == len(IN_FEATURES) must be true,
+# or len(SIZES) == 1 is true and size list SIZES[0] is used for all
+# IN_FEATURES.
+_C.MODEL.ANCHOR_GENERATOR.SIZES = [[32, 64, 128, 256, 512]]
+# Anchor aspect ratios. For each area given in `SIZES`, anchors with different aspect
+# ratios are generated by an anchor generator.
+# Format: list[list[float]]. ASPECT_RATIOS[i] specifies the list of aspect ratios (H/W)
+# to use for IN_FEATURES[i]; len(ASPECT_RATIOS) == len(IN_FEATURES) must be true,
+# or len(ASPECT_RATIOS) == 1 is true and aspect ratio list ASPECT_RATIOS[0] is used
+# for all IN_FEATURES.
+_C.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS = [[0.5, 1.0, 2.0]]
+# Anchor angles.
+# list[list[float]], the angle in degrees, for each input feature map.
+# ANGLES[i] specifies the list of angles for IN_FEATURES[i].
+_C.MODEL.ANCHOR_GENERATOR.ANGLES = [[-90, 0, 90]]
+# Relative offset between the center of the first anchor and the top-left corner of the image
+# Value has to be in [0, 1). Recommend to use 0.5, which means half stride.
+# The value is not expected to affect model accuracy.
+_C.MODEL.ANCHOR_GENERATOR.OFFSET = 0.0
+
+# ---------------------------------------------------------------------------- #
+# RPN options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RPN = CN()
+_C.MODEL.RPN.HEAD_NAME = "StandardRPNHead"  # used by RPN_HEAD_REGISTRY
+
+# Names of the input feature maps to be used by RPN
+# e.g., ["p2", "p3", "p4", "p5", "p6"] for FPN
+_C.MODEL.RPN.IN_FEATURES = ["res4"]
+# Remove RPN anchors that go outside the image by BOUNDARY_THRESH pixels
+# Set to -1 or a large value, e.g. 100000, to disable pruning anchors
+_C.MODEL.RPN.BOUNDARY_THRESH = -1
+# IOU overlap ratios [BG_IOU_THRESHOLD, FG_IOU_THRESHOLD]
+# Minimum overlap required between an anchor and ground-truth box for the
+# (anchor, gt box) pair to be a positive example (IoU >= FG_IOU_THRESHOLD
+# ==> positive RPN example: 1)
+# Maximum overlap allowed between an anchor and ground-truth box for the
+# (anchor, gt box) pair to be a negative examples (IoU < BG_IOU_THRESHOLD
+# ==> negative RPN example: 0)
+# Anchors with overlap in between (BG_IOU_THRESHOLD <= IoU < FG_IOU_THRESHOLD)
+# are ignored (-1)
+_C.MODEL.RPN.IOU_THRESHOLDS = [0.3, 0.7]
+_C.MODEL.RPN.IOU_LABELS = [0, -1, 1]
+# Total number of RPN examples per image
+_C.MODEL.RPN.BATCH_SIZE_PER_IMAGE = 256
+# Target fraction of foreground (positive) examples per RPN minibatch
+_C.MODEL.RPN.POSITIVE_FRACTION = 0.5
+# Weights on (dx, dy, dw, dh) for normalizing RPN anchor regression targets
+_C.MODEL.RPN.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
+_C.MODEL.RPN.SMOOTH_L1_BETA = 0.0
+_C.MODEL.RPN.LOSS_WEIGHT = 1.0
+# Number of top scoring RPN proposals to keep before applying NMS
+# When FPN is used, this is *per FPN level* (not total)
+_C.MODEL.RPN.PRE_NMS_TOPK_TRAIN = 12000
+_C.MODEL.RPN.PRE_NMS_TOPK_TEST = 6000
+# Number of top scoring RPN proposals to keep after applying NMS
+# When FPN is used, this limit is applied per level and then again to the union
+# of proposals from all levels
+# NOTE: When FPN is used, the meaning of this config is different from Detectron1.
+# It means per-batch topk in Detectron1, but per-image topk here.
+# See "modeling/rpn/rpn_outputs.py" for details.
+_C.MODEL.RPN.POST_NMS_TOPK_TRAIN = 2000
+_C.MODEL.RPN.POST_NMS_TOPK_TEST = 1000
+# NMS threshold used on RPN proposals
+_C.MODEL.RPN.NMS_THRESH = 0.7
+
+# ---------------------------------------------------------------------------- #
+# ROI HEADS options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_HEADS = CN()
+_C.MODEL.ROI_HEADS.NAME = "Res5ROIHeads"
+# Number of foreground classes
+_C.MODEL.ROI_HEADS.NUM_CLASSES = 80
+# Names of the input feature maps to be used by ROI heads
+# Currently all heads (box, mask, ...) use the same input feature map list
+# e.g., ["p2", "p3", "p4", "p5"] is commonly used for FPN
+_C.MODEL.ROI_HEADS.IN_FEATURES = ["res4"]
+# IOU overlap ratios [IOU_THRESHOLD]
+# Overlap threshold for an RoI to be considered background (if < IOU_THRESHOLD)
+# Overlap threshold for an RoI to be considered foreground (if >= IOU_THRESHOLD)
+_C.MODEL.ROI_HEADS.IOU_THRESHOLDS = [0.5]
+_C.MODEL.ROI_HEADS.IOU_LABELS = [0, 1]
+# RoI minibatch size *per image* (number of regions of interest [ROIs])
+# Total number of RoIs per training minibatch =
+#   ROI_HEADS.BATCH_SIZE_PER_IMAGE * SOLVER.IMS_PER_BATCH
+# E.g., a common configuration is: 512 * 16 = 8192
+_C.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512
+# Target fraction of RoI minibatch that is labeled foreground (i.e. class > 0)
+_C.MODEL.ROI_HEADS.POSITIVE_FRACTION = 0.25
+
+# Only used on test mode
+
+# Minimum score threshold (assuming scores in a [0, 1] range); a value chosen to
+# balance obtaining high recall with not having too many low precision
+# detections that will slow down inference post processing steps (like NMS)
+# A default threshold of 0.0 increases AP by ~0.2-0.3 but significantly slows down
+# inference.
+_C.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.05
+# Overlap threshold used for non-maximum suppression (suppress boxes with
+# IoU >= this threshold)
+_C.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.5
+# If True, augment proposals with ground-truth boxes before sampling proposals to
+# train ROI heads.
+_C.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT = True
+
+# ---------------------------------------------------------------------------- #
+# Box Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_BOX_HEAD = CN()
+# C4 don't use head name option
+# Options for non-C4 models: FastRCNNConvFCHead,
+_C.MODEL.ROI_BOX_HEAD.NAME = ""
+# Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets
+# These are empirically chosen to approximately lead to unit variance targets
+_C.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS = (10.0, 10.0, 5.0, 5.0)
+# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
+_C.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA = 0.0
+_C.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_BOX_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+_C.MODEL.ROI_BOX_HEAD.NUM_FC = 0
+# Hidden layer dimension for FC layers in the RoI box head
+_C.MODEL.ROI_BOX_HEAD.FC_DIM = 1024
+_C.MODEL.ROI_BOX_HEAD.NUM_CONV = 0
+# Channel dimension for Conv layers in the RoI box head
+_C.MODEL.ROI_BOX_HEAD.CONV_DIM = 256
+# Normalization method for the convolution layers.
+# Options: "" (no norm), "GN", "SyncBN".
+_C.MODEL.ROI_BOX_HEAD.NORM = ""
+# Whether to use class agnostic for bbox regression
+_C.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG = False
+# If true, RoI heads use bounding boxes predicted by the box head rather than proposal boxes.
+_C.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES = False
+
+# ---------------------------------------------------------------------------- #
+# Cascaded Box Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_BOX_CASCADE_HEAD = CN()
+# The number of cascade stages is implicitly defined by the length of the following two configs.
+_C.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS = (
+    (10.0, 10.0, 5.0, 5.0),
+    (20.0, 20.0, 10.0, 10.0),
+    (30.0, 30.0, 15.0, 15.0),
+)
+_C.MODEL.ROI_BOX_CASCADE_HEAD.IOUS = (0.5, 0.6, 0.7)
+
+
+# ---------------------------------------------------------------------------- #
+# Mask Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_MASK_HEAD = CN()
+_C.MODEL.ROI_MASK_HEAD.NAME = "MaskRCNNConvUpsampleHead"
+_C.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO = 0
+_C.MODEL.ROI_MASK_HEAD.NUM_CONV = 0  # The number of convs in the mask head
+_C.MODEL.ROI_MASK_HEAD.CONV_DIM = 256
+# Normalization method for the convolution layers.
+# Options: "" (no norm), "GN", "SyncBN".
+_C.MODEL.ROI_MASK_HEAD.NORM = ""
+# Whether to use class agnostic for mask prediction
+_C.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK = False
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_MASK_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+
+# ---------------------------------------------------------------------------- #
+# Keypoint Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_KEYPOINT_HEAD = CN()
+_C.MODEL.ROI_KEYPOINT_HEAD.NAME = "KRCNNConvDeconvUpsampleHead"
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO = 0
+_C.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS = tuple(512 for _ in range(8))
+_C.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = 17  # 17 is the number of keypoints in COCO.
+
+# Images with too few (or no) keypoints are excluded from training.
+_C.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE = 1
+# Normalize by the total number of visible keypoints in the minibatch if True.
+# Otherwise, normalize by the total number of keypoints that could ever exist
+# in the minibatch.
+# The keypoint softmax loss is only calculated on visible keypoints.
+# Since the number of visible keypoints can vary significantly between
+# minibatches, this has the effect of up-weighting the importance of
+# minibatches with few visible keypoints. (Imagine the extreme case of
+# only one visible keypoint versus N: in the case of N, each one
+# contributes 1/N to the gradient compared to the single keypoint
+# determining the gradient direction). Instead, we can normalize the
+# loss by the total number of keypoints, if it were the case that all
+# keypoints were visible in a full minibatch. (Returning to the example,
+# this means that the one visible keypoint contributes as much as each
+# of the N keypoints.)
+_C.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS = True
+# Multi-task loss weight to use for keypoints
+# Recommended values:
+#   - use 1.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is True
+#   - use 4.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is False
+_C.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT = 1.0
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+# ---------------------------------------------------------------------------- #
+# Semantic Segmentation Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.SEM_SEG_HEAD = CN()
+_C.MODEL.SEM_SEG_HEAD.NAME = "SemSegFPNHead"
+_C.MODEL.SEM_SEG_HEAD.IN_FEATURES = ["p2", "p3", "p4", "p5"]
+# Label in the semantic segmentation ground truth that is ignored, i.e., no loss is calculated for
+# the correposnding pixel.
+_C.MODEL.SEM_SEG_HEAD.IGNORE_VALUE = 255
+# Number of classes in the semantic segmentation head
+_C.MODEL.SEM_SEG_HEAD.NUM_CLASSES = 54
+# Number of channels in the 3x3 convs inside semantic-FPN heads.
+_C.MODEL.SEM_SEG_HEAD.CONVS_DIM = 128
+# Outputs from semantic-FPN heads are up-scaled to the COMMON_STRIDE stride.
+_C.MODEL.SEM_SEG_HEAD.COMMON_STRIDE = 4
+# Normalization method for the convolution layers. Options: "" (no norm), "GN".
+_C.MODEL.SEM_SEG_HEAD.NORM = "GN"
+_C.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT = 1.0
+
+_C.MODEL.PANOPTIC_FPN = CN()
+# Scaling of all losses from instance detection / segmentation head.
+_C.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT = 1.0
+
+# options when combining instance & semantic segmentation outputs
+_C.MODEL.PANOPTIC_FPN.COMBINE = CN({"ENABLED": True})
+_C.MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH = 0.5
+_C.MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT = 4096
+_C.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = 0.5
+
+
+# ---------------------------------------------------------------------------- #
+# RetinaNet Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RETINANET = CN()
+
+# This is the number of foreground classes.
+_C.MODEL.RETINANET.NUM_CLASSES = 80
+
+_C.MODEL.RETINANET.IN_FEATURES = ["p3", "p4", "p5", "p6", "p7"]
+
+# Convolutions to use in the cls and bbox tower
+# NOTE: this doesn't include the last conv for logits
+_C.MODEL.RETINANET.NUM_CONVS = 4
+
+# IoU overlap ratio [bg, fg] for labeling anchors.
+# Anchors with < bg are labeled negative (0)
+# Anchors  with >= bg and < fg are ignored (-1)
+# Anchors with >= fg are labeled positive (1)
+_C.MODEL.RETINANET.IOU_THRESHOLDS = [0.4, 0.5]
+_C.MODEL.RETINANET.IOU_LABELS = [0, -1, 1]
+
+# Prior prob for rare case (i.e. foreground) at the beginning of training.
+# This is used to set the bias for the logits layer of the classifier subnet.
+# This improves training stability in the case of heavy class imbalance.
+_C.MODEL.RETINANET.PRIOR_PROB = 0.01
+
+# Inference cls score threshold, only anchors with score > INFERENCE_TH are
+# considered for inference (to improve speed)
+_C.MODEL.RETINANET.SCORE_THRESH_TEST = 0.05
+_C.MODEL.RETINANET.TOPK_CANDIDATES_TEST = 1000
+_C.MODEL.RETINANET.NMS_THRESH_TEST = 0.5
+
+# Weights on (dx, dy, dw, dh) for normalizing Retinanet anchor regression targets
+_C.MODEL.RETINANET.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+
+# Loss parameters
+_C.MODEL.RETINANET.FOCAL_LOSS_GAMMA = 2.0
+_C.MODEL.RETINANET.FOCAL_LOSS_ALPHA = 0.25
+_C.MODEL.RETINANET.SMOOTH_L1_LOSS_BETA = 0.1
+
+
+# ---------------------------------------------------------------------------- #
+# ResNe[X]t options (ResNets = {ResNet, ResNeXt}
+# Note that parts of a resnet may be used for both the backbone and the head
+# These options apply to both
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RESNETS = CN()
+
+_C.MODEL.RESNETS.DEPTH = 50
+_C.MODEL.RESNETS.OUT_FEATURES = ["res4"]  # res4 for C4 backbone, res2..5 for FPN backbone
+
+# Number of groups to use; 1 ==> ResNet; > 1 ==> ResNeXt
+_C.MODEL.RESNETS.NUM_GROUPS = 1
+
+# Options: FrozenBN, GN, "SyncBN", "BN"
+_C.MODEL.RESNETS.NORM = "FrozenBN"
+
+# Baseline width of each group.
+# Scaling this parameters will scale the width of all bottleneck layers.
+_C.MODEL.RESNETS.WIDTH_PER_GROUP = 64
+
+# Place the stride 2 conv on the 1x1 filter
+# Use True only for the original MSRA ResNet; use False for C2 and Torch models
+_C.MODEL.RESNETS.STRIDE_IN_1X1 = True
+
+# Apply dilation in stage "res5"
+_C.MODEL.RESNETS.RES5_DILATION = 1
+
+# Output width of res2. Scaling this parameters will scale the width of all 1x1 convs in ResNet
+# For R18 and R34, this needs to be set to 64
+_C.MODEL.RESNETS.RES2_OUT_CHANNELS = 256
+_C.MODEL.RESNETS.STEM_OUT_CHANNELS = 64
+
+# Apply Deformable Convolution in stages
+# Specify if apply deform_conv on Res2, Res3, Res4, Res5
+_C.MODEL.RESNETS.DEFORM_ON_PER_STAGE = [False, False, False, False]
+# Use True to use modulated deform_conv (DeformableV2, https://arxiv.org/abs/1811.11168);
+# Use False for DeformableV1.
+_C.MODEL.RESNETS.DEFORM_MODULATED = False
+# Number of groups in deformable conv.
+_C.MODEL.RESNETS.DEFORM_NUM_GROUPS = 1
+
+
+# ---------------------------------------------------------------------------- #
+# Solver
+# ---------------------------------------------------------------------------- #
+_C.SOLVER = CN()
+
+# See detectron2/solver/build.py for LR scheduler options
+_C.SOLVER.LR_SCHEDULER_NAME = "WarmupMultiStepLR"
+
+_C.SOLVER.MAX_ITER = 40000
+
+_C.SOLVER.BASE_LR = 0.001
+
+_C.SOLVER.MOMENTUM = 0.9
+
+_C.SOLVER.NESTEROV = False
+
+_C.SOLVER.WEIGHT_DECAY = 0.0001
+# The weight decay that's applied to parameters of normalization layers
+# (typically the affine transformation)
+_C.SOLVER.WEIGHT_DECAY_NORM = 0.0
+
+_C.SOLVER.GAMMA = 0.1
+# The iteration number to decrease learning rate by GAMMA.
+_C.SOLVER.STEPS = (30000,)
+
+_C.SOLVER.WARMUP_FACTOR = 1.0 / 1000
+_C.SOLVER.WARMUP_ITERS = 1000
+_C.SOLVER.WARMUP_METHOD = "linear"
+
+# Save a checkpoint after every this number of iterations
+_C.SOLVER.CHECKPOINT_PERIOD = 5000
+
+# Number of images per batch across all machines.
+# If we have 16 GPUs and IMS_PER_BATCH = 32,
+# each GPU will see 2 images per batch.
+_C.SOLVER.IMS_PER_BATCH = 16
+
+# Detectron v1 (and previous detection code) used a 2x higher LR and 0 WD for
+# biases. This is not useful (at least for recent models). You should avoid
+# changing these and they exist only to reproduce Detectron v1 training if
+# desired.
+_C.SOLVER.BIAS_LR_FACTOR = 1.0
+_C.SOLVER.WEIGHT_DECAY_BIAS = _C.SOLVER.WEIGHT_DECAY
+
+# Gradient clipping
+_C.SOLVER.CLIP_GRADIENTS = CN({"ENABLED": False})
+# Type of gradient clipping, currently 2 values are supported:
+# - "value": the absolute values of elements of each gradients are clipped
+# - "norm": the norm of the gradient for each parameter is clipped thus
+#   affecting all elements in the parameter
+_C.SOLVER.CLIP_GRADIENTS.CLIP_TYPE = "value"
+# Maximum absolute value used for clipping gradients
+_C.SOLVER.CLIP_GRADIENTS.CLIP_VALUE = 1.0
+# Floating point number p for L-p norm to be used with the "norm"
+# gradient clipping type; for L-inf, please specify .inf
+_C.SOLVER.CLIP_GRADIENTS.NORM_TYPE = 2.0
+
+# ---------------------------------------------------------------------------- #
+# Specific test options
+# ---------------------------------------------------------------------------- #
+_C.TEST = CN()
+# For end-to-end tests to verify the expected accuracy.
+# Each item is [task, metric, value, tolerance]
+# e.g.: [['bbox', 'AP', 38.5, 0.2]]
+_C.TEST.EXPECTED_RESULTS = []
+# The period (in terms of steps) to evaluate the model during training.
+# Set to 0 to disable.
+_C.TEST.EVAL_PERIOD = 0
+# The sigmas used to calculate keypoint OKS. See http://cocodataset.org/#keypoints-eval
+# When empty it will use the defaults in COCO.
+# Otherwise it should have the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS.
+_C.TEST.KEYPOINT_OKS_SIGMAS = []
+# Maximum number of detections to return per image during inference (100 is
+# based on the limit established for the COCO dataset).
+_C.TEST.DETECTIONS_PER_IMAGE = 100
+
+_C.TEST.AUG = CN({"ENABLED": False})
+_C.TEST.AUG.MIN_SIZES = (400, 500, 600, 700, 800, 900, 1000, 1100, 1200)
+_C.TEST.AUG.MAX_SIZE = 4000
+_C.TEST.AUG.FLIP = True
+
+_C.TEST.PRECISE_BN = CN({"ENABLED": False})
+_C.TEST.PRECISE_BN.NUM_ITER = 200
+
+# ---------------------------------------------------------------------------- #
+# Misc options
+# ---------------------------------------------------------------------------- #
+# Directory where output files are written
+_C.OUTPUT_DIR = "./output"
+# Set seed to negative to fully randomize everything.
+# Set seed to positive to use a fixed seed. Note that a fixed seed increases
+# reproducibility but does not guarantee fully deterministic behavior.
+# Disabling all parallelism further increases reproducibility.
+_C.SEED = -1
+# Benchmark different cudnn algorithms.
+# If input images have very different sizes, this option will have large overhead
+# for about 10k iterations. It usually hurts total time, but can benefit for certain models.
+# If input images have the same or similar sizes, benchmark is often helpful.
+_C.CUDNN_BENCHMARK = False
+# The period (in terms of steps) for minibatch visualization at train time.
+# Set to 0 to disable.
+_C.VIS_PERIOD = 0
+
+# global config is for quick hack purposes.
+# You can set them in command line or config files,
+# and access it with:
+#
+# from detectron2.config import global_cfg
+# print(global_cfg.HACK)
+#
+# Do not commit any configs into it.
+_C.GLOBAL = CN()
+_C.GLOBAL.HACK = 1.0

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from . import transforms  # isort:skip
+
+from .build import (
+    build_detection_test_loader,
+    build_detection_train_loader,
+    get_detection_dataset_dicts,
+    load_proposals_into_dataset,
+    print_instances_class_histogram,
+)
+from .catalog import DatasetCatalog, MetadataCatalog
+from .common import DatasetFromList, MapDataset
+from .dataset_mapper import DatasetMapper
+
+# ensure the builtin data are registered
+from . import datasets, samplers  # isort:skip
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/build.py

@@ -0,0 +1,397 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import bisect
+import copy
+import itertools
+import logging
+import numpy as np
+import operator
+import pickle
+import torch.utils.data
+from fvcore.common.file_io import PathManager
+from tabulate import tabulate
+from termcolor import colored
+
+from detectron2.structures import BoxMode
+from detectron2.utils.comm import get_world_size
+from detectron2.utils.env import seed_all_rng
+from detectron2.utils.logger import log_first_n
+
+from . import samplers
+from .catalog import DatasetCatalog, MetadataCatalog
+from .common import AspectRatioGroupedDataset, DatasetFromList, MapDataset
+from .dataset_mapper import DatasetMapper
+from .detection_utils import check_metadata_consistency
+
+"""
+This file contains the default logic to build a dataloader for training or testing.
+"""
+
+__all__ = [
+    "build_detection_train_loader",
+    "build_detection_test_loader",
+    "get_detection_dataset_dicts",
+    "load_proposals_into_dataset",
+    "print_instances_class_histogram",
+]
+
+
+def filter_images_with_only_crowd_annotations(dataset_dicts):
+    """
+    Filter out images with none annotations or only crowd annotations
+    (i.e., images without non-crowd annotations).
+    A common training-time preprocessing on COCO dataset.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+
+    Returns:
+        list[dict]: the same format, but filtered.
+    """
+    num_before = len(dataset_dicts)
+
+    def valid(anns):
+        for ann in anns:
+            if ann.get("iscrowd", 0) == 0:
+                return True
+        return False
+
+    dataset_dicts = [x for x in dataset_dicts if valid(x["annotations"])]
+    num_after = len(dataset_dicts)
+    logger = logging.getLogger(__name__)
+    logger.info(
+        "Removed {} images with no usable annotations. {} images left.".format(
+            num_before - num_after, num_after
+        )
+    )
+    return dataset_dicts
+
+
+def filter_images_with_few_keypoints(dataset_dicts, min_keypoints_per_image):
+    """
+    Filter out images with too few number of keypoints.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+
+    Returns:
+        list[dict]: the same format as dataset_dicts, but filtered.
+    """
+    num_before = len(dataset_dicts)
+
+    def visible_keypoints_in_image(dic):
+        # Each keypoints field has the format [x1, y1, v1, ...], where v is visibility
+        annotations = dic["annotations"]
+        return sum(
+            (np.array(ann["keypoints"][2::3]) > 0).sum()
+            for ann in annotations
+            if "keypoints" in ann
+        )
+
+    dataset_dicts = [
+        x for x in dataset_dicts if visible_keypoints_in_image(x) >= min_keypoints_per_image
+    ]
+    num_after = len(dataset_dicts)
+    logger = logging.getLogger(__name__)
+    logger.info(
+        "Removed {} images with fewer than {} keypoints.".format(
+            num_before - num_after, min_keypoints_per_image
+        )
+    )
+    return dataset_dicts
+
+
+def load_proposals_into_dataset(dataset_dicts, proposal_file):
+    """
+    Load precomputed object proposals into the dataset.
+
+    The proposal file should be a pickled dict with the following keys:
+
+    - "ids": list[int] or list[str], the image ids
+    - "boxes": list[np.ndarray], each is an Nx4 array of boxes corresponding to the image id
+    - "objectness_logits": list[np.ndarray], each is an N sized array of objectness scores
+      corresponding to the boxes.
+    - "bbox_mode": the BoxMode of the boxes array. Defaults to ``BoxMode.XYXY_ABS``.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+        proposal_file (str): file path of pre-computed proposals, in pkl format.
+
+    Returns:
+        list[dict]: the same format as dataset_dicts, but added proposal field.
+    """
+    logger = logging.getLogger(__name__)
+    logger.info("Loading proposals from: {}".format(proposal_file))
+
+    with PathManager.open(proposal_file, "rb") as f:
+        proposals = pickle.load(f, encoding="latin1")
+
+    # Rename the key names in D1 proposal files
+    rename_keys = {"indexes": "ids", "scores": "objectness_logits"}
+    for key in rename_keys:
+        if key in proposals:
+            proposals[rename_keys[key]] = proposals.pop(key)
+
+    # Fetch the indexes of all proposals that are in the dataset
+    # Convert image_id to str since they could be int.
+    img_ids = set({str(record["image_id"]) for record in dataset_dicts})
+    id_to_index = {str(id): i for i, id in enumerate(proposals["ids"]) if str(id) in img_ids}
+
+    # Assuming default bbox_mode of precomputed proposals are 'XYXY_ABS'
+    bbox_mode = BoxMode(proposals["bbox_mode"]) if "bbox_mode" in proposals else BoxMode.XYXY_ABS
+
+    for record in dataset_dicts:
+        # Get the index of the proposal
+        i = id_to_index[str(record["image_id"])]
+
+        boxes = proposals["boxes"][i]
+        objectness_logits = proposals["objectness_logits"][i]
+        # Sort the proposals in descending order of the scores
+        inds = objectness_logits.argsort()[::-1]
+        record["proposal_boxes"] = boxes[inds]
+        record["proposal_objectness_logits"] = objectness_logits[inds]
+        record["proposal_bbox_mode"] = bbox_mode
+
+    return dataset_dicts
+
+
+def _quantize(x, bin_edges):
+    bin_edges = copy.copy(bin_edges)
+    bin_edges = sorted(bin_edges)
+    quantized = list(map(lambda y: bisect.bisect_right(bin_edges, y), x))
+    return quantized
+
+
+def print_instances_class_histogram(dataset_dicts, class_names):
+    """
+    Args:
+        dataset_dicts (list[dict]): list of dataset dicts.
+        class_names (list[str]): list of class names (zero-indexed).
+    """
+    num_classes = len(class_names)
+    hist_bins = np.arange(num_classes + 1)
+    histogram = np.zeros((num_classes,), dtype=np.int)
+    for entry in dataset_dicts:
+        annos = entry["annotations"]
+        classes = [x["category_id"] for x in annos if not x.get("iscrowd", 0)]
+        histogram += np.histogram(classes, bins=hist_bins)[0]
+
+    N_COLS = min(6, len(class_names) * 2)
+
+    def short_name(x):
+        # make long class names shorter. useful for lvis
+        if len(x) > 13:
+            return x[:11] + ".."
+        return x
+
+    data = list(
+        itertools.chain(*[[short_name(class_names[i]), int(v)] for i, v in enumerate(histogram)])
+    )
+    total_num_instances = sum(data[1::2])
+    data.extend([None] * (N_COLS - (len(data) % N_COLS)))
+    if num_classes > 1:
+        data.extend(["total", total_num_instances])
+    data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
+    table = tabulate(
+        data,
+        headers=["category", "#instances"] * (N_COLS // 2),
+        tablefmt="pipe",
+        numalign="left",
+        stralign="center",
+    )
+    log_first_n(
+        logging.INFO,
+        "Distribution of instances among all {} categories:\n".format(num_classes)
+        + colored(table, "cyan"),
+        key="message",
+    )
+
+
+def get_detection_dataset_dicts(
+    dataset_names, filter_empty=True, min_keypoints=0, proposal_files=None
+):
+    """
+    Load and prepare dataset dicts for instance detection/segmentation and semantic segmentation.
+
+    Args:
+        dataset_names (list[str]): a list of dataset names
+        filter_empty (bool): whether to filter out images without instance annotations
+        min_keypoints (int): filter out images with fewer keypoints than
+            `min_keypoints`. Set to 0 to do nothing.
+        proposal_files (list[str]): if given, a list of object proposal files
+            that match each dataset in `dataset_names`.
+    """
+    assert len(dataset_names)
+    dataset_dicts = [DatasetCatalog.get(dataset_name) for dataset_name in dataset_names]
+    for dataset_name, dicts in zip(dataset_names, dataset_dicts):
+        assert len(dicts), "Dataset '{}' is empty!".format(dataset_name)
+
+    if proposal_files is not None:
+        assert len(dataset_names) == len(proposal_files)
+        # load precomputed proposals from proposal files
+        dataset_dicts = [
+            load_proposals_into_dataset(dataset_i_dicts, proposal_file)
+            for dataset_i_dicts, proposal_file in zip(dataset_dicts, proposal_files)
+        ]
+
+    dataset_dicts = list(itertools.chain.from_iterable(dataset_dicts))
+
+    has_instances = "annotations" in dataset_dicts[0]
+    # Keep images without instance-level GT if the dataset has semantic labels.
+    if filter_empty and has_instances and "sem_seg_file_name" not in dataset_dicts[0]:
+        dataset_dicts = filter_images_with_only_crowd_annotations(dataset_dicts)
+
+    if min_keypoints > 0 and has_instances:
+        dataset_dicts = filter_images_with_few_keypoints(dataset_dicts, min_keypoints)
+
+    if has_instances:
+        try:
+            class_names = MetadataCatalog.get(dataset_names[0]).thing_classes
+            check_metadata_consistency("thing_classes", dataset_names)
+            print_instances_class_histogram(dataset_dicts, class_names)
+        except AttributeError:  # class names are not available for this dataset
+            pass
+    return dataset_dicts
+
+
+def build_detection_train_loader(cfg, mapper=None):
+    """
+    A data loader is created by the following steps:
+
+    1. Use the dataset names in config to query :class:`DatasetCatalog`, and obtain a list of dicts.
+    2. Coordinate a random shuffle order shared among all processes (all GPUs)
+    3. Each process spawn another few workers to process the dicts. Each worker will:
+       * Map each metadata dict into another format to be consumed by the model.
+       * Batch them by simply putting dicts into a list.
+
+    The batched ``list[mapped_dict]`` is what this dataloader will yield.
+
+    Args:
+        cfg (CfgNode): the config
+        mapper (callable): a callable which takes a sample (dict) from dataset and
+            returns the format to be consumed by the model.
+            By default it will be `DatasetMapper(cfg, True)`.
+
+    Returns:
+        an infinite iterator of training data
+    """
+    num_workers = get_world_size()
+    images_per_batch = cfg.SOLVER.IMS_PER_BATCH
+    assert (
+        images_per_batch % num_workers == 0
+    ), "SOLVER.IMS_PER_BATCH ({}) must be divisible by the number of workers ({}).".format(
+        images_per_batch, num_workers
+    )
+    assert (
+        images_per_batch >= num_workers
+    ), "SOLVER.IMS_PER_BATCH ({}) must be larger than the number of workers ({}).".format(
+        images_per_batch, num_workers
+    )
+    images_per_worker = images_per_batch // num_workers
+
+    dataset_dicts = get_detection_dataset_dicts(
+        cfg.DATASETS.TRAIN,
+        filter_empty=cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS,
+        min_keypoints=cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE
+        if cfg.MODEL.KEYPOINT_ON
+        else 0,
+        proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None,
+    )
+    dataset = DatasetFromList(dataset_dicts, copy=False)
+
+    if mapper is None:
+        mapper = DatasetMapper(cfg, True)
+    dataset = MapDataset(dataset, mapper)
+
+    sampler_name = cfg.DATALOADER.SAMPLER_TRAIN
+    logger = logging.getLogger(__name__)
+    logger.info("Using training sampler {}".format(sampler_name))
+    if sampler_name == "TrainingSampler":
+        sampler = samplers.TrainingSampler(len(dataset))
+    elif sampler_name == "RepeatFactorTrainingSampler":
+        sampler = samplers.RepeatFactorTrainingSampler(
+            dataset_dicts, cfg.DATALOADER.REPEAT_THRESHOLD
+        )
+    else:
+        raise ValueError("Unknown training sampler: {}".format(sampler_name))
+
+    if cfg.DATALOADER.ASPECT_RATIO_GROUPING:
+        data_loader = torch.utils.data.DataLoader(
+            dataset,
+            sampler=sampler,
+            num_workers=cfg.DATALOADER.NUM_WORKERS,
+            batch_sampler=None,
+            collate_fn=operator.itemgetter(0),  # don't batch, but yield individual elements
+            worker_init_fn=worker_init_reset_seed,
+        )  # yield individual mapped dict
+        data_loader = AspectRatioGroupedDataset(data_loader, images_per_worker)
+    else:
+        batch_sampler = torch.utils.data.sampler.BatchSampler(
+            sampler, images_per_worker, drop_last=True
+        )
+        # drop_last so the batch always have the same size
+        data_loader = torch.utils.data.DataLoader(
+            dataset,
+            num_workers=cfg.DATALOADER.NUM_WORKERS,
+            batch_sampler=batch_sampler,
+            collate_fn=trivial_batch_collator,
+            worker_init_fn=worker_init_reset_seed,
+        )
+
+    return data_loader
+
+
+def build_detection_test_loader(cfg, dataset_name, mapper=None):
+    """
+    Similar to `build_detection_train_loader`.
+    But this function uses the given `dataset_name` argument (instead of the names in cfg),
+    and uses batch size 1.
+
+    Args:
+        cfg: a detectron2 CfgNode
+        dataset_name (str): a name of the dataset that's available in the DatasetCatalog
+        mapper (callable): a callable which takes a sample (dict) from dataset
+           and returns the format to be consumed by the model.
+           By default it will be `DatasetMapper(cfg, False)`.
+
+    Returns:
+        DataLoader: a torch DataLoader, that loads the given detection
+        dataset, with test-time transformation and batching.
+    """
+    dataset_dicts = get_detection_dataset_dicts(
+        [dataset_name],
+        filter_empty=False,
+        proposal_files=[
+            cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(dataset_name)]
+        ]
+        if cfg.MODEL.LOAD_PROPOSALS
+        else None,
+    )
+
+    dataset = DatasetFromList(dataset_dicts)
+    if mapper is None:
+        mapper = DatasetMapper(cfg, False)
+    dataset = MapDataset(dataset, mapper)
+
+    sampler = samplers.InferenceSampler(len(dataset))
+    # Always use 1 image per worker during inference since this is the
+    # standard when reporting inference time in papers.
+    batch_sampler = torch.utils.data.sampler.BatchSampler(sampler, 1, drop_last=False)
+
+    data_loader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=cfg.DATALOADER.NUM_WORKERS,
+        batch_sampler=batch_sampler,
+        collate_fn=trivial_batch_collator,
+    )
+    return data_loader
+
+
+def trivial_batch_collator(batch):
+    """
+    A batch collator that does nothing.
+    """
+    return batch
+
+
+def worker_init_reset_seed(worker_id):
+    seed_all_rng(np.random.randint(2 ** 31) + worker_id)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/catalog.py

@@ -0,0 +1,221 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import logging
+import types
+from typing import List
+
+from detectron2.utils.logger import log_first_n
+
+__all__ = ["DatasetCatalog", "MetadataCatalog"]
+
+
+class DatasetCatalog(object):
+    """
+    A catalog that stores information about the data and how to obtain them.
+
+    It contains a mapping from strings
+    (which are names that identify a dataset, e.g. "coco_2014_train")
+    to a function which parses the dataset and returns the samples in the
+    format of `list[dict]`.
+
+    The returned dicts should be in Detectron2 Dataset format (See DATASETS.md for details)
+    if used with the data loader functionalities in `data/build.py,data/detection_transform.py`.
+
+    The purpose of having this catalog is to make it easy to choose
+    different data, by just using the strings in the config.
+    """
+
+    _REGISTERED = {}
+
+    @staticmethod
+    def register(name, func):
+        """
+        Args:
+            name (str): the name that identifies a dataset, e.g. "coco_2014_train".
+            func (callable): a callable which takes no arguments and returns a list of dicts.
+        """
+        assert callable(func), "You must register a function with `DatasetCatalog.register`!"
+        assert name not in DatasetCatalog._REGISTERED, "Dataset '{}' is already registered!".format(
+            name
+        )
+        DatasetCatalog._REGISTERED[name] = func
+
+    @staticmethod
+    def get(name):
+        """
+        Call the registered function and return its results.
+
+        Args:
+            name (str): the name that identifies a dataset, e.g. "coco_2014_train".
+
+        Returns:
+            list[dict]: dataset annotations.0
+        """
+        try:
+            f = DatasetCatalog._REGISTERED[name]
+        except KeyError:
+            raise KeyError(
+                "Dataset '{}' is not registered! Available data are: {}".format(
+                    name, ", ".join(DatasetCatalog._REGISTERED.keys())
+                )
+            )
+        return f()
+
+    @staticmethod
+    def list() -> List[str]:
+        """
+        List all registered data.
+
+        Returns:
+            list[str]
+        """
+        return list(DatasetCatalog._REGISTERED.keys())
+
+    @staticmethod
+    def clear():
+        """
+        Remove all registered dataset.
+        """
+        DatasetCatalog._REGISTERED.clear()
+
+
+class Metadata(types.SimpleNamespace):
+    """
+    A class that supports simple attribute setter/getter.
+    It is intended for storing metadata of a dataset and make it accessible globally.
+
+    Examples:
+
+    .. code-block:: python
+
+        # somewhere when you load the data:
+        MetadataCatalog.get("mydataset").thing_classes = ["person", "dog"]
+
+        # somewhere when you print statistics or visualize:
+        classes = MetadataCatalog.get("mydataset").thing_classes
+    """
+
+    # the name of the dataset
+    # set default to N/A so that `self.name` in the errors will not trigger getattr again
+    name: str = "N/A"
+
+    _RENAMED = {
+        "class_names": "thing_classes",
+        "dataset_id_to_contiguous_id": "thing_dataset_id_to_contiguous_id",
+        "stuff_class_names": "stuff_classes",
+    }
+
+    def __getattr__(self, key):
+        if key in self._RENAMED:
+            log_first_n(
+                logging.WARNING,
+                "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
+                n=10,
+            )
+            return getattr(self, self._RENAMED[key])
+
+        raise AttributeError(
+            "Attribute '{}' does not exist in the metadata of '{}'. Available keys are {}.".format(
+                key, self.name, str(self.__dict__.keys())
+            )
+        )
+
+    def __setattr__(self, key, val):
+        if key in self._RENAMED:
+            log_first_n(
+                logging.WARNING,
+                "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
+                n=10,
+            )
+            setattr(self, self._RENAMED[key], val)
+
+        # Ensure that metadata of the same name stays consistent
+        try:
+            oldval = getattr(self, key)
+            assert oldval == val, (
+                "Attribute '{}' in the metadata of '{}' cannot be set "
+                "to a different value!\n{} != {}".format(key, self.name, oldval, val)
+            )
+        except AttributeError:
+            super().__setattr__(key, val)
+
+    def as_dict(self):
+        """
+        Returns all the metadata as a dict.
+        Note that modifications to the returned dict will not reflect on the Metadata object.
+        """
+        return copy.copy(self.__dict__)
+
+    def set(self, **kwargs):
+        """
+        Set multiple metadata with kwargs.
+        """
+        for k, v in kwargs.items():
+            setattr(self, k, v)
+        return self
+
+    def get(self, key, default=None):
+        """
+        Access an attribute and return its value if exists.
+        Otherwise return default.
+        """
+        try:
+            return getattr(self, key)
+        except AttributeError:
+            return default
+
+
+class MetadataCatalog:
+    """
+    MetadataCatalog provides access to "Metadata" of a given dataset.
+
+    The metadata associated with a certain name is a singleton: once created,
+    the metadata will stay alive and will be returned by future calls to `get(name)`.
+
+    It's like global variables, so don't abuse it.
+    It's meant for storing knowledge that's constant and shared across the execution
+    of the program, e.g.: the class names in COCO.
+    """
+
+    _NAME_TO_META = {}
+
+    @staticmethod
+    def get(name):
+        """
+        Args:
+            name (str): name of a dataset (e.g. coco_2014_train).
+
+        Returns:
+            Metadata: The :class:`Metadata` instance associated with this name,
+            or create an empty one if none is available.
+        """
+        assert len(name)
+        if name in MetadataCatalog._NAME_TO_META:
+            ret = MetadataCatalog._NAME_TO_META[name]
+            # TODO this is for the BC breaking change in D15247032.
+            # Remove this in the future.
+            if hasattr(ret, "dataset_name"):
+                logger = logging.getLogger()
+                logger.warning(
+                    """
+The 'dataset_name' key in metadata is no longer used for
+sharing metadata among splits after D15247032! Add
+metadata to each split (now called dataset) separately!
+                    """
+                )
+                parent_meta = MetadataCatalog.get(ret.dataset_name).as_dict()
+                ret.set(**parent_meta)
+            return ret
+        else:
+            m = MetadataCatalog._NAME_TO_META[name] = Metadata(name=name)
+            return m
+
+    @staticmethod
+    def list():
+        """
+        List all registered metadata.
+
+        Returns:
+            list[str]: keys (names of data) of all registered metadata
+        """
+        return list(MetadataCatalog._NAME_TO_META.keys())

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/common.py

@@ -0,0 +1,149 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import logging
+import numpy as np
+import pickle
+import random
+import torch.utils.data as data
+
+from detectron2.utils.serialize import PicklableWrapper
+
+__all__ = ["MapDataset", "DatasetFromList", "AspectRatioGroupedDataset"]
+
+
+class MapDataset(data.Dataset):
+    """
+    Map a function over the elements in a dataset.
+
+    Args:
+        dataset: a dataset where map function is applied.
+        map_func: a callable which maps the element in dataset. map_func is
+            responsible for error handling, when error happens, it needs to
+            return None so the MapDataset will randomly use other
+            elements from the dataset.
+    """
+
+    def __init__(self, dataset, map_func):
+        self._dataset = dataset
+        self._map_func = PicklableWrapper(map_func)  # wrap so that a lambda will work
+
+        self._rng = random.Random(42)
+        self._fallback_candidates = set(range(len(dataset)))
+
+    def __len__(self):
+        return len(self._dataset)
+
+    def __getitem__(self, idx):
+        retry_count = 0
+        cur_idx = int(idx)
+
+        while True:
+            data = self._map_func(self._dataset[cur_idx])
+            if data is not None:
+                self._fallback_candidates.add(cur_idx)
+                return data
+
+            # _map_func fails for this idx, use a random new index from the pool
+            retry_count += 1
+            self._fallback_candidates.discard(cur_idx)
+            cur_idx = self._rng.sample(self._fallback_candidates, k=1)[0]
+
+            if retry_count >= 3:
+                logger = logging.getLogger(__name__)
+                logger.warning(
+                    "Failed to apply `_map_func` for idx: {}, retry count: {}".format(
+                        idx, retry_count
+                    )
+                )
+
+
+class DatasetFromList(data.Dataset):
+    """
+    Wrap a list to a torch Dataset. It produces elements of the list as data.
+    """
+
+    def __init__(self, lst: list, copy: bool = True, serialize: bool = True):
+        """
+        Args:
+            lst (list): a list which contains elements to produce.
+            copy (bool): whether to deepcopy the element when producing it,
+                so that the result can be modified in place without affecting the
+                source in the list.
+            serialize (bool): whether to hold memory using serialized objects, when
+                enabled, data loader workers can use shared RAM from master
+                process instead of making a copy.
+        """
+        self._lst = lst
+        self._copy = copy
+        self._serialize = serialize
+
+        def _serialize(data):
+            buffer = pickle.dumps(data, protocol=-1)
+            return np.frombuffer(buffer, dtype=np.uint8)
+
+        if self._serialize:
+            logger = logging.getLogger(__name__)
+            logger.info(
+                "Serializing {} elements to byte tensors and concatenating them all ...".format(
+                    len(self._lst)
+                )
+            )
+            self._lst = [_serialize(x) for x in self._lst]
+            self._addr = np.asarray([len(x) for x in self._lst], dtype=np.int64)
+            self._addr = np.cumsum(self._addr)
+            self._lst = np.concatenate(self._lst)
+            logger.info("Serialized dataset takes {:.2f} MiB".format(len(self._lst) / 1024 ** 2))
+
+    def __len__(self):
+        if self._serialize:
+            return len(self._addr)
+        else:
+            return len(self._lst)
+
+    def __getitem__(self, idx):
+        if self._serialize:
+            start_addr = 0 if idx == 0 else self._addr[idx - 1].item()
+            end_addr = self._addr[idx].item()
+            bytes = memoryview(self._lst[start_addr:end_addr])
+            return pickle.loads(bytes)
+        elif self._copy:
+            return copy.deepcopy(self._lst[idx])
+        else:
+            return self._lst[idx]
+
+
+class AspectRatioGroupedDataset(data.IterableDataset):
+    """
+    Batch data that have similar aspect ratio together.
+    In this implementation, images whose aspect ratio < (or >) 1 will
+    be batched together.
+    This improves training speed because the images then need less padding
+    to form a batch.
+
+    It assumes the underlying dataset produces dicts with "width" and "height" keys.
+    It will then produce a list of original dicts with length = batch_size,
+    all with similar aspect ratios.
+    """
+
+    def __init__(self, dataset, batch_size):
+        """
+        Args:
+            dataset: an iterable. Each element must be a dict with keys
+                "width" and "height", which will be used to batch data.
+            batch_size (int):
+        """
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self._buckets = [[] for _ in range(2)]
+        # Hard-coded two aspect ratio groups: w > h and w < h.
+        # Can add support for more aspect ratio groups, but doesn't seem useful
+
+    def __iter__(self):
+        for d in self.dataset:
+            w, h = d["width"], d["height"]
+            bucket_id = 0 if w > h else 1
+            bucket = self._buckets[bucket_id]
+            bucket.append(d)
+            if len(bucket) == self.batch_size:
+                yield bucket[:]
+                del bucket[:]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/dataset_mapper.py

@@ -0,0 +1,149 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import logging
+import numpy as np
+import torch
+from fvcore.common.file_io import PathManager
+from PIL import Image
+
+from . import detection_utils as utils
+from . import transforms as T
+
+"""
+This file contains the default mapping that's applied to "dataset dicts".
+"""
+
+__all__ = ["DatasetMapper"]
+
+
+class DatasetMapper:
+    """
+    A callable which takes a dataset dict in Detectron2 Dataset format,
+    and map it into a format used by the model.
+
+    This is the default callable to be used to map your dataset dict into training data.
+    You may need to follow it to implement your own one for customized logic,
+    such as a different way to read or transform images.
+    See :doc:`/tutorials/data_loading` for details.
+
+    The callable currently does the following:
+
+    1. Read the image from "file_name"
+    2. Applies cropping/geometric transforms to the image and annotations
+    3. Prepare data and annotations to Tensor and :class:`Instances`
+    """
+
+    def __init__(self, cfg, is_train=True):
+        if cfg.INPUT.CROP.ENABLED and is_train:
+            self.crop_gen = T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE)
+            logging.getLogger(__name__).info("CropGen used in training: " + str(self.crop_gen))
+        else:
+            self.crop_gen = None
+
+        self.tfm_gens = utils.build_transform_gen(cfg, is_train)
+
+        # fmt: off
+        self.img_format     = cfg.INPUT.FORMAT
+        self.mask_on        = cfg.MODEL.MASK_ON
+        self.mask_format    = cfg.INPUT.MASK_FORMAT
+        self.keypoint_on    = cfg.MODEL.KEYPOINT_ON
+        self.load_proposals = cfg.MODEL.LOAD_PROPOSALS
+        # fmt: on
+        if self.keypoint_on and is_train:
+            # Flip only makes sense in training
+            self.keypoint_hflip_indices = utils.create_keypoint_hflip_indices(cfg.DATASETS.TRAIN)
+        else:
+            self.keypoint_hflip_indices = None
+
+        if self.load_proposals:
+            self.min_box_side_len = cfg.MODEL.PROPOSAL_GENERATOR.MIN_SIZE
+            self.proposal_topk = (
+                cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN
+                if is_train
+                else cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST
+            )
+        self.is_train = is_train
+
+    def __call__(self, dataset_dict):
+        """
+        Args:
+            dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
+
+        Returns:
+            dict: a format that builtin models in detectron2 accept
+        """
+        dataset_dict = copy.deepcopy(dataset_dict)  # it will be modified by code below
+        # USER: Write your own image loading if it's not from a file
+        image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
+        utils.check_image_size(dataset_dict, image)
+
+        if "annotations" not in dataset_dict:
+            image, transforms = T.apply_transform_gens(
+                ([self.crop_gen] if self.crop_gen else []) + self.tfm_gens, image
+            )
+        else:
+            # Crop around an instance if there are instances in the image.
+            # USER: Remove if you don't use cropping
+            if self.crop_gen:
+                crop_tfm = utils.gen_crop_transform_with_instance(
+                    self.crop_gen.get_crop_size(image.shape[:2]),
+                    image.shape[:2],
+                    np.random.choice(dataset_dict["annotations"]),
+                )
+                image = crop_tfm.apply_image(image)
+            image, transforms = T.apply_transform_gens(self.tfm_gens, image)
+            if self.crop_gen:
+                transforms = crop_tfm + transforms
+
+        image_shape = image.shape[:2]  # h, w
+
+        # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
+        # but not efficient on large generic data structures due to the use of pickle & mp.Queue.
+        # Therefore it's important to use torch.Tensor.
+        dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
+
+        # USER: Remove if you don't use pre-computed proposals.
+        if self.load_proposals:
+            utils.transform_proposals(
+                dataset_dict, image_shape, transforms, self.min_box_side_len, self.proposal_topk
+            )
+
+        if not self.is_train:
+            # USER: Modify this if you want to keep them for some reason.
+            dataset_dict.pop("annotations", None)
+            dataset_dict.pop("sem_seg_file_name", None)
+            return dataset_dict
+
+        if "annotations" in dataset_dict:
+            # USER: Modify this if you want to keep them for some reason.
+            for anno in dataset_dict["annotations"]:
+                if not self.mask_on:
+                    anno.pop("segmentation", None)
+                if not self.keypoint_on:
+                    anno.pop("keypoints", None)
+
+            # USER: Implement additional transformations if you have other types of data
+            annos = [
+                utils.transform_instance_annotations(
+                    obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices
+                )
+                for obj in dataset_dict.pop("annotations")
+                if obj.get("iscrowd", 0) == 0
+            ]
+            instances = utils.annotations_to_instances(
+                annos, image_shape, mask_format=self.mask_format
+            )
+            # Create a tight bounding box from masks, useful when image is cropped
+            if self.crop_gen and instances.has("gt_masks"):
+                instances.gt_boxes = instances.gt_masks.get_bounding_boxes()
+            dataset_dict["instances"] = utils.filter_empty_instances(instances)
+
+        # USER: Remove if you don't do semantic/panoptic segmentation.
+        if "sem_seg_file_name" in dataset_dict:
+            with PathManager.open(dataset_dict.pop("sem_seg_file_name"), "rb") as f:
+                sem_seg_gt = Image.open(f)
+                sem_seg_gt = np.asarray(sem_seg_gt, dtype="uint8")
+            sem_seg_gt = transforms.apply_segmentation(sem_seg_gt)
+            sem_seg_gt = torch.as_tensor(sem_seg_gt.astype("long"))
+            dataset_dict["sem_seg"] = sem_seg_gt
+        return dataset_dict

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/README.md

@@ -0,0 +1,9 @@
+
+
+### Common Datasets
+
+The dataset implemented here do not need to load the data into the final format.
+It should provide the minimal data structure needed to use the dataset, so it can be very efficient.
+
+For example, for an image dataset, just provide the file names and labels, but don't read the images.
+Let the downstream decide how to read.

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .cityscapes import load_cityscapes_instances
+from .coco import load_coco_json, load_sem_seg
+from .lvis import load_lvis_json, register_lvis_instances, get_lvis_instances_meta
+from .register_coco import register_coco_instances, register_coco_panoptic_separated
+from . import builtin  # ensure the builtin data are registered
+
+
+__all__ = [k for k in globals().keys() if "builtin" not in k and not k.startswith("_")]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/builtin.py

@@ -0,0 +1,220 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+
+"""
+This file registers pre-defined data at hard-coded paths, and their metadata.
+
+We hard-code metadata for common data. This will enable:
+1. Consistency check when loading the data
+2. Use models on these standard data directly and run demos,
+   without having to download the dataset annotations
+
+We hard-code some paths to the dataset that's assumed to
+exist in "./data/".
+
+Users SHOULD NOT use this file to create new dataset / metadata for new dataset.
+To add new dataset, refer to the tutorial "docs/DATASETS.md".
+"""
+
+import os
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+
+from .builtin_meta import _get_builtin_metadata
+from .cityscapes import load_cityscapes_instances, load_cityscapes_semantic
+from .lvis import get_lvis_instances_meta, register_lvis_instances
+from .pascal_voc import register_pascal_voc
+from .register_coco import register_coco_instances, register_coco_panoptic_separated
+
+# ==== Predefined data and splits for COCO ==========
+
+_PREDEFINED_SPLITS_COCO = {}
+_PREDEFINED_SPLITS_COCO["coco"] = {
+    "coco_2014_train": ("coco/train2014", "coco/annotations/instances_train2014.json"),
+    "coco_2014_val": ("coco/val2014", "coco/annotations/instances_val2014.json"),
+    "coco_2014_minival": ("coco/val2014", "coco/annotations/instances_minival2014.json"),
+    "coco_2014_minival_100": ("coco/val2014", "coco/annotations/instances_minival2014_100.json"),
+    "coco_2014_valminusminival": (
+        "coco/val2014",
+        "coco/annotations/instances_valminusminival2014.json",
+    ),
+    "coco_2017_train": ("coco/train2017", "coco/annotations/instances_train2017.json"),
+    "coco_2017_val": ("coco/val2017", "coco/annotations/instances_val2017.json"),
+    "coco_2017_test": ("coco/test2017", "coco/annotations/image_info_test2017.json"),
+    "coco_2017_test-dev": ("coco/test2017", "coco/annotations/image_info_test-dev2017.json"),
+    "coco_2017_val_100": ("coco/val2017", "coco/annotations/instances_val2017_100.json"),
+}
+
+_PREDEFINED_SPLITS_COCO["coco_person"] = {
+    "keypoints_coco_2014_train": (
+        "coco/train2014",
+        "coco/annotations/person_keypoints_train2014.json",
+    ),
+    "keypoints_coco_2014_val": ("coco/val2014", "coco/annotations/person_keypoints_val2014.json"),
+    "keypoints_coco_2014_minival": (
+        "coco/val2014",
+        "coco/annotations/person_keypoints_minival2014.json",
+    ),
+    "keypoints_coco_2014_valminusminival": (
+        "coco/val2014",
+        "coco/annotations/person_keypoints_valminusminival2014.json",
+    ),
+    "keypoints_coco_2014_minival_100": (
+        "coco/val2014",
+        "coco/annotations/person_keypoints_minival2014_100.json",
+    ),
+    "keypoints_coco_2017_train": (
+        "coco/train2017",
+        "coco/annotations/person_keypoints_train2017.json",
+    ),
+    "keypoints_coco_2017_val": ("coco/val2017", "coco/annotations/person_keypoints_val2017.json"),
+    "keypoints_coco_2017_val_100": (
+        "coco/val2017",
+        "coco/annotations/person_keypoints_val2017_100.json",
+    ),
+}
+
+
+_PREDEFINED_SPLITS_COCO_PANOPTIC = {
+    "coco_2017_train_panoptic": (
+        # This is the original panoptic annotation directory
+        "coco/panoptic_train2017",
+        "coco/annotations/panoptic_train2017.json",
+        # This directory contains semantic annotations that are
+        # converted from panoptic annotations.
+        # It is used by PanopticFPN.
+        # You can use the script at detectron2/data/prepare_panoptic_fpn.py
+        # to create these directories.
+        "coco/panoptic_stuff_train2017",
+    ),
+    "coco_2017_val_panoptic": (
+        "coco/panoptic_val2017",
+        "coco/annotations/panoptic_val2017.json",
+        "coco/panoptic_stuff_val2017",
+    ),
+    "coco_2017_val_100_panoptic": (
+        "coco/panoptic_val2017_100",
+        "coco/annotations/panoptic_val2017_100.json",
+        "coco/panoptic_stuff_val2017_100",
+    ),
+}
+
+
+def register_all_coco(root):
+    for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_COCO.items():
+        for key, (image_root, json_file) in splits_per_dataset.items():
+            # Assume pre-defined data live in `./data`.
+            register_coco_instances(
+                key,
+                _get_builtin_metadata(dataset_name),
+                os.path.join(root, json_file) if "://" not in json_file else json_file,
+                os.path.join(root, image_root),
+            )
+
+    for (
+        prefix,
+        (panoptic_root, panoptic_json, semantic_root),
+    ) in _PREDEFINED_SPLITS_COCO_PANOPTIC.items():
+        prefix_instances = prefix[: -len("_panoptic")]
+        instances_meta = MetadataCatalog.get(prefix_instances)
+        image_root, instances_json = instances_meta.image_root, instances_meta.json_file
+        register_coco_panoptic_separated(
+            prefix,
+            _get_builtin_metadata("coco_panoptic_separated"),
+            image_root,
+            os.path.join(root, panoptic_root),
+            os.path.join(root, panoptic_json),
+            os.path.join(root, semantic_root),
+            instances_json,
+        )
+
+
+# ==== Predefined data and splits for LVIS ==========
+
+
+_PREDEFINED_SPLITS_LVIS = {
+    "lvis_v0.5": {
+        "lvis_v0.5_train": ("coco/train2017", "lvis/lvis_v0.5_train.json"),
+        "lvis_v0.5_val": ("coco/val2017", "lvis/lvis_v0.5_val.json"),
+        "lvis_v0.5_val_rand_100": ("coco/val2017", "lvis/lvis_v0.5_val_rand_100.json"),
+        "lvis_v0.5_test": ("coco/test2017", "lvis/lvis_v0.5_image_info_test.json"),
+    },
+    "lvis_v0.5_cocofied": {
+        "lvis_v0.5_train_cocofied": ("coco/train2017", "lvis/lvis_v0.5_train_cocofied.json"),
+        "lvis_v0.5_val_cocofied": ("coco/val2017", "lvis/lvis_v0.5_val_cocofied.json"),
+    },
+}
+
+
+def register_all_lvis(root):
+    for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_LVIS.items():
+        for key, (image_root, json_file) in splits_per_dataset.items():
+            # Assume pre-defined data live in `./data`.
+            register_lvis_instances(
+                key,
+                get_lvis_instances_meta(dataset_name),
+                os.path.join(root, json_file) if "://" not in json_file else json_file,
+                os.path.join(root, image_root),
+            )
+
+
+# ==== Predefined splits for raw cityscapes images ===========
+
+
+_RAW_CITYSCAPES_SPLITS = {
+    "cityscapes_fine_{task}_train": ("cityscapes/leftImg8bit/train", "cityscapes/gtFine/train"),
+    "cityscapes_fine_{task}_val": ("cityscapes/leftImg8bit/val", "cityscapes/gtFine/val"),
+    "cityscapes_fine_{task}_test": ("cityscapes/leftImg8bit/test", "cityscapes/gtFine/test"),
+}
+
+
+def register_all_cityscapes(root):
+    for key, (image_dir, gt_dir) in _RAW_CITYSCAPES_SPLITS.items():
+        meta = _get_builtin_metadata("cityscapes")
+        image_dir = os.path.join(root, image_dir)
+        gt_dir = os.path.join(root, gt_dir)
+
+        inst_key = key.format(task="instance_seg")
+        DatasetCatalog.register(
+            inst_key,
+            lambda x=image_dir, y=gt_dir: load_cityscapes_instances(
+                x, y, from_json=True, to_polygons=True
+            ),
+        )
+        MetadataCatalog.get(inst_key).set(
+            image_dir=image_dir, gt_dir=gt_dir, evaluator_type="cityscapes_instance", **meta
+        )
+
+        sem_key = key.format(task="sem_seg")
+        DatasetCatalog.register(
+            sem_key, lambda x=image_dir, y=gt_dir: load_cityscapes_semantic(x, y)
+        )
+        MetadataCatalog.get(sem_key).set(
+            image_dir=image_dir, gt_dir=gt_dir, evaluator_type="cityscapes_sem_seg", **meta
+        )
+
+
+# ==== Predefined splits for PASCAL VOC ===========
+def register_all_pascal_voc(root):
+    SPLITS = [
+        ("voc_2007_trainval", "VOC2007", "trainval"),
+        ("voc_2007_train", "VOC2007", "train"),
+        ("voc_2007_val", "VOC2007", "val"),
+        ("voc_2007_test", "VOC2007", "test"),
+        ("voc_2012_trainval", "VOC2012", "trainval"),
+        ("voc_2012_train", "VOC2012", "train"),
+        ("voc_2012_val", "VOC2012", "val"),
+    ]
+    for name, dirname, split in SPLITS:
+        year = 2007 if "2007" in name else 2012
+        register_pascal_voc(name, os.path.join(root, dirname), split, year)
+        MetadataCatalog.get(name).evaluator_type = "pascal_voc"
+
+
+# Register them all under "./data"
+_root = os.getenv("DETECTRON2_DATASETS", "data")
+register_all_coco(_root)
+register_all_lvis(_root)
+register_all_cityscapes(_root)
+register_all_pascal_voc(_root)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/builtin_meta.py

@@ -0,0 +1,267 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+
+# All coco categories, together with their nice-looking visualization colors
+# It's from https://github.com/cocodataset/panopticapi/blob/master/panoptic_coco_categories.json
+COCO_CATEGORIES = [
+    {"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "person"},
+    {"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "bicycle"},
+    {"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "car"},
+    {"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "motorcycle"},
+    {"color": [106, 0, 228], "isthing": 1, "id": 5, "name": "airplane"},
+    {"color": [0, 60, 100], "isthing": 1, "id": 6, "name": "bus"},
+    {"color": [0, 80, 100], "isthing": 1, "id": 7, "name": "train"},
+    {"color": [0, 0, 70], "isthing": 1, "id": 8, "name": "truck"},
+    {"color": [0, 0, 192], "isthing": 1, "id": 9, "name": "boat"},
+    {"color": [250, 170, 30], "isthing": 1, "id": 10, "name": "traffic light"},
+    {"color": [100, 170, 30], "isthing": 1, "id": 11, "name": "fire hydrant"},
+    {"color": [220, 220, 0], "isthing": 1, "id": 13, "name": "stop sign"},
+    {"color": [175, 116, 175], "isthing": 1, "id": 14, "name": "parking meter"},
+    {"color": [250, 0, 30], "isthing": 1, "id": 15, "name": "bench"},
+    {"color": [165, 42, 42], "isthing": 1, "id": 16, "name": "bird"},
+    {"color": [255, 77, 255], "isthing": 1, "id": 17, "name": "cat"},
+    {"color": [0, 226, 252], "isthing": 1, "id": 18, "name": "dog"},
+    {"color": [182, 182, 255], "isthing": 1, "id": 19, "name": "horse"},
+    {"color": [0, 82, 0], "isthing": 1, "id": 20, "name": "sheep"},
+    {"color": [120, 166, 157], "isthing": 1, "id": 21, "name": "cow"},
+    {"color": [110, 76, 0], "isthing": 1, "id": 22, "name": "elephant"},
+    {"color": [174, 57, 255], "isthing": 1, "id": 23, "name": "bear"},
+    {"color": [199, 100, 0], "isthing": 1, "id": 24, "name": "zebra"},
+    {"color": [72, 0, 118], "isthing": 1, "id": 25, "name": "giraffe"},
+    {"color": [255, 179, 240], "isthing": 1, "id": 27, "name": "backpack"},
+    {"color": [0, 125, 92], "isthing": 1, "id": 28, "name": "umbrella"},
+    {"color": [209, 0, 151], "isthing": 1, "id": 31, "name": "handbag"},
+    {"color": [188, 208, 182], "isthing": 1, "id": 32, "name": "tie"},
+    {"color": [0, 220, 176], "isthing": 1, "id": 33, "name": "suitcase"},
+    {"color": [255, 99, 164], "isthing": 1, "id": 34, "name": "frisbee"},
+    {"color": [92, 0, 73], "isthing": 1, "id": 35, "name": "skis"},
+    {"color": [133, 129, 255], "isthing": 1, "id": 36, "name": "snowboard"},
+    {"color": [78, 180, 255], "isthing": 1, "id": 37, "name": "sports ball"},
+    {"color": [0, 228, 0], "isthing": 1, "id": 38, "name": "kite"},
+    {"color": [174, 255, 243], "isthing": 1, "id": 39, "name": "baseball bat"},
+    {"color": [45, 89, 255], "isthing": 1, "id": 40, "name": "baseball glove"},
+    {"color": [134, 134, 103], "isthing": 1, "id": 41, "name": "skateboard"},
+    {"color": [145, 148, 174], "isthing": 1, "id": 42, "name": "surfboard"},
+    {"color": [255, 208, 186], "isthing": 1, "id": 43, "name": "tennis racket"},
+    {"color": [197, 226, 255], "isthing": 1, "id": 44, "name": "bottle"},
+    {"color": [171, 134, 1], "isthing": 1, "id": 46, "name": "wine glass"},
+    {"color": [109, 63, 54], "isthing": 1, "id": 47, "name": "cup"},
+    {"color": [207, 138, 255], "isthing": 1, "id": 48, "name": "fork"},
+    {"color": [151, 0, 95], "isthing": 1, "id": 49, "name": "knife"},
+    {"color": [9, 80, 61], "isthing": 1, "id": 50, "name": "spoon"},
+    {"color": [84, 105, 51], "isthing": 1, "id": 51, "name": "bowl"},
+    {"color": [74, 65, 105], "isthing": 1, "id": 52, "name": "banana"},
+    {"color": [166, 196, 102], "isthing": 1, "id": 53, "name": "apple"},
+    {"color": [208, 195, 210], "isthing": 1, "id": 54, "name": "sandwich"},
+    {"color": [255, 109, 65], "isthing": 1, "id": 55, "name": "orange"},
+    {"color": [0, 143, 149], "isthing": 1, "id": 56, "name": "broccoli"},
+    {"color": [179, 0, 194], "isthing": 1, "id": 57, "name": "carrot"},
+    {"color": [209, 99, 106], "isthing": 1, "id": 58, "name": "hot dog"},
+    {"color": [5, 121, 0], "isthing": 1, "id": 59, "name": "pizza"},
+    {"color": [227, 255, 205], "isthing": 1, "id": 60, "name": "donut"},
+    {"color": [147, 186, 208], "isthing": 1, "id": 61, "name": "cake"},
+    {"color": [153, 69, 1], "isthing": 1, "id": 62, "name": "chair"},
+    {"color": [3, 95, 161], "isthing": 1, "id": 63, "name": "couch"},
+    {"color": [163, 255, 0], "isthing": 1, "id": 64, "name": "potted plant"},
+    {"color": [119, 0, 170], "isthing": 1, "id": 65, "name": "bed"},
+    {"color": [0, 182, 199], "isthing": 1, "id": 67, "name": "dining table"},
+    {"color": [0, 165, 120], "isthing": 1, "id": 70, "name": "toilet"},
+    {"color": [183, 130, 88], "isthing": 1, "id": 72, "name": "tv"},
+    {"color": [95, 32, 0], "isthing": 1, "id": 73, "name": "laptop"},
+    {"color": [130, 114, 135], "isthing": 1, "id": 74, "name": "mouse"},
+    {"color": [110, 129, 133], "isthing": 1, "id": 75, "name": "remote"},
+    {"color": [166, 74, 118], "isthing": 1, "id": 76, "name": "keyboard"},
+    {"color": [219, 142, 185], "isthing": 1, "id": 77, "name": "cell phone"},
+    {"color": [79, 210, 114], "isthing": 1, "id": 78, "name": "microwave"},
+    {"color": [178, 90, 62], "isthing": 1, "id": 79, "name": "oven"},
+    {"color": [65, 70, 15], "isthing": 1, "id": 80, "name": "toaster"},
+    {"color": [127, 167, 115], "isthing": 1, "id": 81, "name": "sink"},
+    {"color": [59, 105, 106], "isthing": 1, "id": 82, "name": "refrigerator"},
+    {"color": [142, 108, 45], "isthing": 1, "id": 84, "name": "book"},
+    {"color": [196, 172, 0], "isthing": 1, "id": 85, "name": "clock"},
+    {"color": [95, 54, 80], "isthing": 1, "id": 86, "name": "vase"},
+    {"color": [128, 76, 255], "isthing": 1, "id": 87, "name": "scissors"},
+    {"color": [201, 57, 1], "isthing": 1, "id": 88, "name": "teddy bear"},
+    {"color": [246, 0, 122], "isthing": 1, "id": 89, "name": "hair drier"},
+    {"color": [191, 162, 208], "isthing": 1, "id": 90, "name": "toothbrush"},
+    {"color": [255, 255, 128], "isthing": 0, "id": 92, "name": "banner"},
+    {"color": [147, 211, 203], "isthing": 0, "id": 93, "name": "blanket"},
+    {"color": [150, 100, 100], "isthing": 0, "id": 95, "name": "bridge"},
+    {"color": [168, 171, 172], "isthing": 0, "id": 100, "name": "cardboard"},
+    {"color": [146, 112, 198], "isthing": 0, "id": 107, "name": "counter"},
+    {"color": [210, 170, 100], "isthing": 0, "id": 109, "name": "curtain"},
+    {"color": [92, 136, 89], "isthing": 0, "id": 112, "name": "door-stuff"},
+    {"color": [218, 88, 184], "isthing": 0, "id": 118, "name": "floor-wood"},
+    {"color": [241, 129, 0], "isthing": 0, "id": 119, "name": "flower"},
+    {"color": [217, 17, 255], "isthing": 0, "id": 122, "name": "fruit"},
+    {"color": [124, 74, 181], "isthing": 0, "id": 125, "name": "gravel"},
+    {"color": [70, 70, 70], "isthing": 0, "id": 128, "name": "house"},
+    {"color": [255, 228, 255], "isthing": 0, "id": 130, "name": "light"},
+    {"color": [154, 208, 0], "isthing": 0, "id": 133, "name": "mirror-stuff"},
+    {"color": [193, 0, 92], "isthing": 0, "id": 138, "name": "net"},
+    {"color": [76, 91, 113], "isthing": 0, "id": 141, "name": "pillow"},
+    {"color": [255, 180, 195], "isthing": 0, "id": 144, "name": "platform"},
+    {"color": [106, 154, 176], "isthing": 0, "id": 145, "name": "playingfield"},
+    {"color": [230, 150, 140], "isthing": 0, "id": 147, "name": "railroad"},
+    {"color": [60, 143, 255], "isthing": 0, "id": 148, "name": "river"},
+    {"color": [128, 64, 128], "isthing": 0, "id": 149, "name": "road"},
+    {"color": [92, 82, 55], "isthing": 0, "id": 151, "name": "roof"},
+    {"color": [254, 212, 124], "isthing": 0, "id": 154, "name": "sand"},
+    {"color": [73, 77, 174], "isthing": 0, "id": 155, "name": "sea"},
+    {"color": [255, 160, 98], "isthing": 0, "id": 156, "name": "shelf"},
+    {"color": [255, 255, 255], "isthing": 0, "id": 159, "name": "snow"},
+    {"color": [104, 84, 109], "isthing": 0, "id": 161, "name": "stairs"},
+    {"color": [169, 164, 131], "isthing": 0, "id": 166, "name": "tent"},
+    {"color": [225, 199, 255], "isthing": 0, "id": 168, "name": "towel"},
+    {"color": [137, 54, 74], "isthing": 0, "id": 171, "name": "wall-brick"},
+    {"color": [135, 158, 223], "isthing": 0, "id": 175, "name": "wall-stone"},
+    {"color": [7, 246, 231], "isthing": 0, "id": 176, "name": "wall-tile"},
+    {"color": [107, 255, 200], "isthing": 0, "id": 177, "name": "wall-wood"},
+    {"color": [58, 41, 149], "isthing": 0, "id": 178, "name": "water-other"},
+    {"color": [183, 121, 142], "isthing": 0, "id": 180, "name": "window-blind"},
+    {"color": [255, 73, 97], "isthing": 0, "id": 181, "name": "window-other"},
+    {"color": [107, 142, 35], "isthing": 0, "id": 184, "name": "tree-merged"},
+    {"color": [190, 153, 153], "isthing": 0, "id": 185, "name": "fence-merged"},
+    {"color": [146, 139, 141], "isthing": 0, "id": 186, "name": "ceiling-merged"},
+    {"color": [70, 130, 180], "isthing": 0, "id": 187, "name": "sky-other-merged"},
+    {"color": [134, 199, 156], "isthing": 0, "id": 188, "name": "cabinet-merged"},
+    {"color": [209, 226, 140], "isthing": 0, "id": 189, "name": "table-merged"},
+    {"color": [96, 36, 108], "isthing": 0, "id": 190, "name": "floor-other-merged"},
+    {"color": [96, 96, 96], "isthing": 0, "id": 191, "name": "pavement-merged"},
+    {"color": [64, 170, 64], "isthing": 0, "id": 192, "name": "mountain-merged"},
+    {"color": [152, 251, 152], "isthing": 0, "id": 193, "name": "grass-merged"},
+    {"color": [208, 229, 228], "isthing": 0, "id": 194, "name": "dirt-merged"},
+    {"color": [206, 186, 171], "isthing": 0, "id": 195, "name": "paper-merged"},
+    {"color": [152, 161, 64], "isthing": 0, "id": 196, "name": "food-other-merged"},
+    {"color": [116, 112, 0], "isthing": 0, "id": 197, "name": "building-other-merged"},
+    {"color": [0, 114, 143], "isthing": 0, "id": 198, "name": "rock-merged"},
+    {"color": [102, 102, 156], "isthing": 0, "id": 199, "name": "wall-other-merged"},
+    {"color": [250, 141, 255], "isthing": 0, "id": 200, "name": "rug-merged"},
+]
+
+# fmt: off
+COCO_PERSON_KEYPOINT_NAMES = (
+    "nose",
+    "left_eye", "right_eye",
+    "left_ear", "right_ear",
+    "left_shoulder", "right_shoulder",
+    "left_elbow", "right_elbow",
+    "left_wrist", "right_wrist",
+    "left_hip", "right_hip",
+    "left_knee", "right_knee",
+    "left_ankle", "right_ankle",
+)
+# fmt: on
+
+# Pairs of keypoints that should be exchanged under horizontal flipping
+COCO_PERSON_KEYPOINT_FLIP_MAP = (
+    ("left_eye", "right_eye"),
+    ("left_ear", "right_ear"),
+    ("left_shoulder", "right_shoulder"),
+    ("left_elbow", "right_elbow"),
+    ("left_wrist", "right_wrist"),
+    ("left_hip", "right_hip"),
+    ("left_knee", "right_knee"),
+    ("left_ankle", "right_ankle"),
+)
+
+# rules for pairs of keypoints to draw a line between, and the line color to use.
+KEYPOINT_CONNECTION_RULES = [
+    # face
+    ("left_ear", "left_eye", (102, 204, 255)),
+    ("right_ear", "right_eye", (51, 153, 255)),
+    ("left_eye", "nose", (102, 0, 204)),
+    ("nose", "right_eye", (51, 102, 255)),
+    # upper-body
+    ("left_shoulder", "right_shoulder", (255, 128, 0)),
+    ("left_shoulder", "left_elbow", (153, 255, 204)),
+    ("right_shoulder", "right_elbow", (128, 229, 255)),
+    ("left_elbow", "left_wrist", (153, 255, 153)),
+    ("right_elbow", "right_wrist", (102, 255, 224)),
+    # lower-body
+    ("left_hip", "right_hip", (255, 102, 0)),
+    ("left_hip", "left_knee", (255, 255, 77)),
+    ("right_hip", "right_knee", (153, 255, 204)),
+    ("left_knee", "left_ankle", (191, 255, 128)),
+    ("right_knee", "right_ankle", (255, 195, 77)),
+]
+
+
+def _get_coco_instances_meta():
+    thing_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    assert len(thing_ids) == 80, len(thing_ids)
+    # Mapping from the incontiguous COCO category id to an id in [0, 79]
+    thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}
+    thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    ret = {
+        "thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id,
+        "thing_classes": thing_classes,
+        "thing_colors": thing_colors,
+    }
+    return ret
+
+
+def _get_coco_panoptic_separated_meta():
+    """
+    Returns metadata for "separated" version of the panoptic segmentation dataset.
+    """
+    stuff_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 0]
+    assert len(stuff_ids) == 53, len(stuff_ids)
+
+    # For semantic segmentation, this mapping maps from contiguous stuff id
+    # (in [0, 53], used in models) to ids in the dataset (used for processing results)
+    # The id 0 is mapped to an extra category "thing".
+    stuff_dataset_id_to_contiguous_id = {k: i + 1 for i, k in enumerate(stuff_ids)}
+    # When converting COCO panoptic annotations to semantic annotations
+    # We label the "thing" category to 0
+    stuff_dataset_id_to_contiguous_id[0] = 0
+
+    # 54 names for COCO stuff categories (including "things")
+    stuff_classes = ["things"] + [
+        k["name"].replace("-other", "").replace("-merged", "")
+        for k in COCO_CATEGORIES
+        if k["isthing"] == 0
+    ]
+
+    # NOTE: I randomly picked a color for things
+    stuff_colors = [[82, 18, 128]] + [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 0]
+    ret = {
+        "stuff_dataset_id_to_contiguous_id": stuff_dataset_id_to_contiguous_id,
+        "stuff_classes": stuff_classes,
+        "stuff_colors": stuff_colors,
+    }
+    ret.update(_get_coco_instances_meta())
+    return ret
+
+
+def _get_builtin_metadata(dataset_name):
+    if dataset_name == "coco":
+        return _get_coco_instances_meta()
+    if dataset_name == "coco_panoptic_separated":
+        return _get_coco_panoptic_separated_meta()
+    elif dataset_name == "coco_person":
+        return {
+            "thing_classes": ["person"],
+            "keypoint_names": COCO_PERSON_KEYPOINT_NAMES,
+            "keypoint_flip_map": COCO_PERSON_KEYPOINT_FLIP_MAP,
+            "keypoint_connection_rules": KEYPOINT_CONNECTION_RULES,
+        }
+    elif dataset_name == "cityscapes":
+        # fmt: off
+        CITYSCAPES_THING_CLASSES = [
+            "person", "rider", "car", "truck",
+            "bus", "train", "motorcycle", "bicycle",
+        ]
+        CITYSCAPES_STUFF_CLASSES = [
+            "road", "sidewalk", "building", "wall", "fence", "pole", "traffic light",
+            "traffic sign", "vegetation", "terrain", "sky", "person", "rider", "car",
+            "truck", "bus", "train", "motorcycle", "bicycle", "license plate",
+        ]
+        # fmt: on
+        return {
+            "thing_classes": CITYSCAPES_THING_CLASSES,
+            "stuff_classes": CITYSCAPES_STUFF_CLASSES,
+        }
+    raise KeyError("No built-in metadata for dataset {}".format(dataset_name))

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/cityscapes.py

@@ -0,0 +1,329 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import functools
+import json
+import logging
+import multiprocessing as mp
+import numpy as np
+import os
+from itertools import chain
+import pycocotools.mask as mask_util
+from fvcore.common.file_io import PathManager
+from PIL import Image
+
+from detectron2.structures import BoxMode
+from detectron2.utils.comm import get_world_size
+from detectron2.utils.logger import setup_logger
+
+try:
+    import cv2  # noqa
+except ImportError:
+    # OpenCV is an optional dependency at the moment
+    pass
+
+
+logger = logging.getLogger(__name__)
+
+
+def get_cityscapes_files(image_dir, gt_dir):
+    files = []
+    # scan through the directory
+    cities = PathManager.ls(image_dir)
+    logger.info(f"{len(cities)} cities found in '{image_dir}'.")
+    for city in cities:
+        city_img_dir = os.path.join(image_dir, city)
+        city_gt_dir = os.path.join(gt_dir, city)
+        for basename in PathManager.ls(city_img_dir):
+            image_file = os.path.join(city_img_dir, basename)
+
+            suffix = "leftImg8bit.png"
+            assert basename.endswith(suffix)
+            basename = basename[: -len(suffix)]
+
+            instance_file = os.path.join(city_gt_dir, basename + "gtFine_instanceIds.png")
+            label_file = os.path.join(city_gt_dir, basename + "gtFine_labelIds.png")
+            json_file = os.path.join(city_gt_dir, basename + "gtFine_polygons.json")
+
+            files.append((image_file, instance_file, label_file, json_file))
+    assert len(files), "No images found in {}".format(image_dir)
+    for f in files[0]:
+        assert PathManager.isfile(f), f
+    return files
+
+
+def load_cityscapes_instances(image_dir, gt_dir, from_json=True, to_polygons=True):
+    """
+    Args:
+        image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train".
+        gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train".
+        from_json (bool): whether to read annotations from the raw json file or the png files.
+        to_polygons (bool): whether to represent the segmentation as polygons
+            (COCO's format) instead of masks (cityscapes's format).
+
+    Returns:
+        list[dict]: a list of dicts in Detectron2 standard format. (See
+        `Using Custom Datasets </tutorials/data.html>`_ )
+    """
+    if from_json:
+        assert to_polygons, (
+            "Cityscapes's json annotations are in polygon format. "
+            "Converting to mask format is not supported now."
+        )
+    files = get_cityscapes_files(image_dir, gt_dir)
+
+    logger.info("Preprocessing cityscapes annotations ...")
+    # This is still not fast: all workers will execute duplicate works and will
+    # take up to 10m on a 8GPU server.
+    pool = mp.Pool(processes=max(mp.cpu_count() // get_world_size() // 2, 4))
+
+    ret = pool.map(
+        functools.partial(cityscapes_files_to_dict, from_json=from_json, to_polygons=to_polygons),
+        files,
+    )
+    logger.info("Loaded {} images from {}".format(len(ret), image_dir))
+
+    # Map cityscape ids to contiguous ids
+    from cityscapesscripts.helpers.labels import labels
+
+    labels = [l for l in labels if l.hasInstances and not l.ignoreInEval]
+    dataset_id_to_contiguous_id = {l.id: idx for idx, l in enumerate(labels)}
+    for dict_per_image in ret:
+        for anno in dict_per_image["annotations"]:
+            anno["category_id"] = dataset_id_to_contiguous_id[anno["category_id"]]
+    return ret
+
+
+def load_cityscapes_semantic(image_dir, gt_dir):
+    """
+    Args:
+        image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train".
+        gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train".
+
+    Returns:
+        list[dict]: a list of dict, each has "file_name" and
+            "sem_seg_file_name".
+    """
+    ret = []
+    # gt_dir is small and contain many small files. make sense to fetch to local first
+    gt_dir = PathManager.get_local_path(gt_dir)
+    for image_file, _, label_file, json_file in get_cityscapes_files(image_dir, gt_dir):
+        label_file = label_file.replace("labelIds", "labelTrainIds")
+
+        with PathManager.open(json_file, "r") as f:
+            jsonobj = json.load(f)
+        ret.append(
+            {
+                "file_name": image_file,
+                "sem_seg_file_name": label_file,
+                "height": jsonobj["imgHeight"],
+                "width": jsonobj["imgWidth"],
+            }
+        )
+    assert len(ret), f"No images found in {image_dir}!"
+    assert PathManager.isfile(
+        ret[0]["sem_seg_file_name"]
+    ), "Please generate labelTrainIds.png with cityscapesscripts/preparation/createTrainIdLabelImgs.py"  # noqa
+    return ret
+
+
+def cityscapes_files_to_dict(files, from_json, to_polygons):
+    """
+    Parse cityscapes annotation files to a instance segmentation dataset dict.
+
+    Args:
+        files (tuple): consists of (image_file, instance_id_file, label_id_file, json_file)
+        from_json (bool): whether to read annotations from the raw json file or the png files.
+        to_polygons (bool): whether to represent the segmentation as polygons
+            (COCO's format) instead of masks (cityscapes's format).
+
+    Returns:
+        A dict in Detectron2 Dataset format.
+    """
+    from cityscapesscripts.helpers.labels import id2label, name2label
+
+    image_file, instance_id_file, _, json_file = files
+
+    annos = []
+
+    if from_json:
+        from shapely.geometry import MultiPolygon, Polygon
+
+        with PathManager.open(json_file, "r") as f:
+            jsonobj = json.load(f)
+        ret = {
+            "file_name": image_file,
+            "image_id": os.path.basename(image_file),
+            "height": jsonobj["imgHeight"],
+            "width": jsonobj["imgWidth"],
+        }
+
+        # `polygons_union` contains the union of all valid polygons.
+        polygons_union = Polygon()
+
+        # CityscapesScripts draw the polygons in sequential order
+        # and each polygon *overwrites* existing ones. See
+        # (https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/json2instanceImg.py) # noqa
+        # We use reverse order, and each polygon *avoids* early ones.
+        # This will resolve the ploygon overlaps in the same way as CityscapesScripts.
+        for obj in jsonobj["objects"][::-1]:
+            if "deleted" in obj:  # cityscapes data format specific
+                continue
+            label_name = obj["label"]
+
+            try:
+                label = name2label[label_name]
+            except KeyError:
+                if label_name.endswith("group"):  # crowd area
+                    label = name2label[label_name[: -len("group")]]
+                else:
+                    raise
+            if label.id < 0:  # cityscapes data format
+                continue
+
+            # Cityscapes's raw annotations uses integer coordinates
+            # Therefore +0.5 here
+            poly_coord = np.asarray(obj["polygon"], dtype="f4") + 0.5
+            # CityscapesScript uses PIL.ImageDraw.polygon to rasterize
+            # polygons for evaluation. This function operates in integer space
+            # and draws each pixel whose center falls into the polygon.
+            # Therefore it draws a polygon which is 0.5 "fatter" in expectation.
+            # We therefore dilate the input polygon by 0.5 as our input.
+            poly = Polygon(poly_coord).buffer(0.5, resolution=4)
+
+            if not label.hasInstances or label.ignoreInEval:
+                # even if we won't store the polygon it still contributes to overlaps resolution
+                polygons_union = polygons_union.union(poly)
+                continue
+
+            # Take non-overlapping part of the polygon
+            poly_wo_overlaps = poly.difference(polygons_union)
+            if poly_wo_overlaps.is_empty:
+                continue
+            polygons_union = polygons_union.union(poly)
+
+            anno = {}
+            anno["iscrowd"] = label_name.endswith("group")
+            anno["category_id"] = label.id
+
+            if isinstance(poly_wo_overlaps, Polygon):
+                poly_list = [poly_wo_overlaps]
+            elif isinstance(poly_wo_overlaps, MultiPolygon):
+                poly_list = poly_wo_overlaps.geoms
+            else:
+                raise NotImplementedError("Unknown geometric structure {}".format(poly_wo_overlaps))
+
+            poly_coord = []
+            for poly_el in poly_list:
+                # COCO API can work only with exterior boundaries now, hence we store only them.
+                # TODO: store both exterior and interior boundaries once other parts of the
+                # codebase support holes in polygons.
+                poly_coord.append(list(chain(*poly_el.exterior.coords)))
+            anno["segmentation"] = poly_coord
+            (xmin, ymin, xmax, ymax) = poly_wo_overlaps.bounds
+
+            anno["bbox"] = (xmin, ymin, xmax, ymax)
+            anno["bbox_mode"] = BoxMode.XYXY_ABS
+
+            annos.append(anno)
+    else:
+        # See also the official annotation parsing scripts at
+        # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/instances2dict.py  # noqa
+        with PathManager.open(instance_id_file, "rb") as f:
+            inst_image = np.asarray(Image.open(f), order="F")
+        # ids < 24 are stuff labels (filtering them first is about 5% faster)
+        flattened_ids = np.unique(inst_image[inst_image >= 24])
+
+        ret = {
+            "file_name": image_file,
+            "image_id": os.path.basename(image_file),
+            "height": inst_image.shape[0],
+            "width": inst_image.shape[1],
+        }
+
+        for instance_id in flattened_ids:
+            # For non-crowd annotations, instance_id // 1000 is the label_id
+            # Crowd annotations have <1000 instance ids
+            label_id = instance_id // 1000 if instance_id >= 1000 else instance_id
+            label = id2label[label_id]
+            if not label.hasInstances or label.ignoreInEval:
+                continue
+
+            anno = {}
+            anno["iscrowd"] = instance_id < 1000
+            anno["category_id"] = label.id
+
+            mask = np.asarray(inst_image == instance_id, dtype=np.uint8, order="F")
+
+            inds = np.nonzero(mask)
+            ymin, ymax = inds[0].min(), inds[0].max()
+            xmin, xmax = inds[1].min(), inds[1].max()
+            anno["bbox"] = (xmin, ymin, xmax, ymax)
+            if xmax <= xmin or ymax <= ymin:
+                continue
+            anno["bbox_mode"] = BoxMode.XYXY_ABS
+            if to_polygons:
+                # This conversion comes from D4809743 and D5171122,
+                # when Mask-RCNN was first developed.
+                contours = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)[
+                    -2
+                ]
+                polygons = [c.reshape(-1).tolist() for c in contours if len(c) >= 3]
+                # opencv's can produce invalid polygons
+                if len(polygons) == 0:
+                    continue
+                anno["segmentation"] = polygons
+            else:
+                anno["segmentation"] = mask_util.encode(mask[:, :, None])[0]
+            annos.append(anno)
+    ret["annotations"] = annos
+    return ret
+
+
+if __name__ == "__main__":
+    """
+    Test the cityscapes dataset loader.
+
+    Usage:
+        python -m detectron2.data.data.cityscapes \
+            cityscapes/leftImg8bit/train cityscapes/gtFine/train
+    """
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("image_dir")
+    parser.add_argument("gt_dir")
+    parser.add_argument("--type", choices=["instance", "semantic"], default="instance")
+    args = parser.parse_args()
+    from detectron2.data.catalog import Metadata
+    from detectron2.utils.visualizer import Visualizer
+    from cityscapesscripts.helpers.labels import labels
+
+    logger = setup_logger(name=__name__)
+
+    dirname = "cityscapes-data-vis"
+    os.makedirs(dirname, exist_ok=True)
+
+    if args.type == "instance":
+        dicts = load_cityscapes_instances(
+            args.image_dir, args.gt_dir, from_json=True, to_polygons=True
+        )
+        logger.info("Done loading {} samples.".format(len(dicts)))
+
+        thing_classes = [k.name for k in labels if k.hasInstances and not k.ignoreInEval]
+        meta = Metadata().set(thing_classes=thing_classes)
+
+    else:
+        dicts = load_cityscapes_semantic(args.image_dir, args.gt_dir)
+        logger.info("Done loading {} samples.".format(len(dicts)))
+
+        stuff_names = [k.name for k in labels if k.trainId != 255]
+        stuff_colors = [k.color for k in labels if k.trainId != 255]
+        meta = Metadata().set(stuff_names=stuff_names, stuff_colors=stuff_colors)
+
+    for d in dicts:
+        img = np.array(Image.open(PathManager.open(d["file_name"], "rb")))
+        visualizer = Visualizer(img, metadata=meta)
+        vis = visualizer.draw_dataset_dict(d)
+        # cv2.imshow("a", vis.get_image()[:, :, ::-1])
+        # cv2.waitKey()
+        fpath = os.path.join(dirname, os.path.basename(d["file_name"]))
+        vis.save(fpath)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/coco.py

@@ -0,0 +1,466 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import contextlib
+import datetime
+import io
+import json
+import logging
+import numpy as np
+import os
+import pycocotools.mask as mask_util
+from fvcore.common.file_io import PathManager, file_lock
+from fvcore.common.timer import Timer
+from PIL import Image
+
+from detectron2.structures import Boxes, BoxMode, PolygonMasks
+
+from .. import DatasetCatalog, MetadataCatalog
+
+"""
+This file contains functions to parse COCO-format annotations into dicts in "Detectron2 format".
+"""
+
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["load_coco_json", "load_sem_seg", "convert_to_coco_json"]
+
+
+def load_coco_json(json_file, image_root, dataset_name=None, extra_annotation_keys=None):
+    """
+    Load a json file with COCO's instances annotation format.
+    Currently supports instance detection, instance segmentation,
+    and person keypoints annotations.
+
+    Args:
+        json_file (str): full path to the json file in COCO instances annotation format.
+        image_root (str or path-like): the directory where the images in this json file exists.
+        dataset_name (str): the name of the dataset (e.g., coco_2017_train).
+            If provided, this function will also put "thing_classes" into
+            the metadata associated with this dataset.
+        extra_annotation_keys (list[str]): list of per-annotation keys that should also be
+            loaded into the dataset dict (besides "iscrowd", "bbox", "keypoints",
+            "category_id", "segmentation"). The values for these keys will be returned as-is.
+            For example, the densepose annotations are loaded in this way.
+
+    Returns:
+        list[dict]: a list of dicts in Detectron2 standard dataset dicts format. (See
+        `Using Custom Datasets </tutorials/data.html>`_ )
+
+    Notes:
+        1. This function does not read the image files.
+           The results do not have the "image" field.
+    """
+    from pycocotools.coco import COCO
+
+    timer = Timer()
+    json_file = PathManager.get_local_path(json_file)
+    with contextlib.redirect_stdout(io.StringIO()):
+        coco_api = COCO(json_file)
+    if timer.seconds() > 1:
+        logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds()))
+
+    id_map = None
+    if dataset_name is not None:
+        meta = MetadataCatalog.get(dataset_name)
+        cat_ids = sorted(coco_api.getCatIds())
+        cats = coco_api.loadCats(cat_ids)
+        # The categories in a custom json file may not be sorted.
+        thing_classes = [c["name"] for c in sorted(cats, key=lambda x: x["id"])]
+        meta.thing_classes = thing_classes
+
+        # In COCO, certain category ids are artificially removed,
+        # and by convention they are always ignored.
+        # We deal with COCO's id issue and translate
+        # the category ids to contiguous ids in [0, 80).
+
+        # It works by looking at the "categories" field in the json, therefore
+        # if users' own json also have incontiguous ids, we'll
+        # apply this mapping as well but print a warning.
+        if not (min(cat_ids) == 1 and max(cat_ids) == len(cat_ids)):
+            if "coco" not in dataset_name:
+                logger.warning(
+                    """
+Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you.
+"""
+                )
+        id_map = {v: i for i, v in enumerate(cat_ids)}
+        meta.thing_dataset_id_to_contiguous_id = id_map
+
+    # sort indices for reproducible results
+    img_ids = sorted(coco_api.imgs.keys())
+    # imgs is a list of dicts, each looks something like:
+    # {'license': 4,
+    #  'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
+    #  'file_name': 'COCO_val2014_000000001268.jpg',
+    #  'height': 427,
+    #  'width': 640,
+    #  'date_captured': '2013-11-17 05:57:24',
+    #  'id': 1268}
+    imgs = coco_api.loadImgs(img_ids)
+    # anns is a list[list[dict]], where each dict is an annotation
+    # record for an object. The inner list enumerates the objects in an image
+    # and the outer list enumerates over images. Example of anns[0]:
+    # [{'segmentation': [[192.81,
+    #     247.09,
+    #     ...
+    #     219.03,
+    #     249.06]],
+    #   'area': 1035.749,
+    #   'iscrowd': 0,
+    #   'image_id': 1268,
+    #   'bbox': [192.81, 224.8, 74.73, 33.43],
+    #   'category_id': 16,
+    #   'id': 42986},
+    #  ...]
+    anns = [coco_api.imgToAnns[img_id] for img_id in img_ids]
+
+    if "minival" not in json_file:
+        # The popular valminusminival & minival annotations for COCO2014 contain this bug.
+        # However the ratio of buggy annotations there is tiny and does not affect accuracy.
+        # Therefore we explicitly white-list them.
+        ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image]
+        assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique!".format(
+            json_file
+        )
+
+    imgs_anns = list(zip(imgs, anns))
+
+    logger.info("Loaded {} images in COCO format from {}".format(len(imgs_anns), json_file))
+
+    dataset_dicts = []
+
+    ann_keys = ["iscrowd", "bbox", "keypoints", "category_id"] + (extra_annotation_keys or [])
+
+    num_instances_without_valid_segmentation = 0
+
+    for (img_dict, anno_dict_list) in imgs_anns:
+        record = {}
+        record["file_name"] = os.path.join(image_root, img_dict["file_name"])
+        record["height"] = img_dict["height"]
+        record["width"] = img_dict["width"]
+        image_id = record["image_id"] = img_dict["id"]
+
+        objs = []
+        for anno in anno_dict_list:
+            # Check that the image_id in this annotation is the same as
+            # the image_id we're looking at.
+            # This fails only when the data parsing logic or the annotation file is buggy.
+
+            # The original COCO valminusminival2014 & minival2014 annotation files
+            # actually contains bugs that, together with certain ways of using COCO API,
+            # can trigger this assertion.
+            assert anno["image_id"] == image_id
+
+            assert anno.get("ignore", 0) == 0, '"ignore" in COCO json file is not supported.'
+
+            obj = {key: anno[key] for key in ann_keys if key in anno}
+
+            segm = anno.get("segmentation", None)
+            if segm:  # either list[list[float]] or dict(RLE)
+                if not isinstance(segm, dict):
+                    # filter out invalid polygons (< 3 points)
+                    segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6]
+                    if len(segm) == 0:
+                        num_instances_without_valid_segmentation += 1
+                        continue  # ignore this instance
+                obj["segmentation"] = segm
+
+            keypts = anno.get("keypoints", None)
+            if keypts:  # list[int]
+                for idx, v in enumerate(keypts):
+                    if idx % 3 != 2:
+                        # COCO's segmentation coordinates are floating points in [0, H or W],
+                        # but keypoint coordinates are integers in [0, H-1 or W-1]
+                        # Therefore we assume the coordinates are "pixel indices" and
+                        # add 0.5 to convert to floating point coordinates.
+                        keypts[idx] = v + 0.5
+                obj["keypoints"] = keypts
+
+            obj["bbox_mode"] = BoxMode.XYWH_ABS
+            if id_map:
+                obj["category_id"] = id_map[obj["category_id"]]
+            objs.append(obj)
+        record["annotations"] = objs
+        dataset_dicts.append(record)
+
+    if num_instances_without_valid_segmentation > 0:
+        logger.warning(
+            "Filtered out {} instances without valid segmentation. "
+            "There might be issues in your dataset generation process.".format(
+                num_instances_without_valid_segmentation
+            )
+        )
+    return dataset_dicts
+
+
+def load_sem_seg(gt_root, image_root, gt_ext="png", image_ext="jpg"):
+    """
+    Load semantic segmentation data. All files under "gt_root" with "gt_ext" extension are
+    treated as ground truth annotations and all files under "image_root" with "image_ext" extension
+    as input images. Ground truth and input images are matched using file paths relative to
+    "gt_root" and "image_root" respectively without taking into account file extensions.
+    This works for COCO as well as some other data.
+
+    Args:
+        gt_root (str): full path to ground truth semantic segmentation files. Semantic segmentation
+            annotations are stored as images with integer values in pixels that represent
+            corresponding semantic labels.
+        image_root (str): the directory where the input images are.
+        gt_ext (str): file extension for ground truth annotations.
+        image_ext (str): file extension for input images.
+
+    Returns:
+        list[dict]:
+            a list of dicts in detectron2 standard format without instance-level
+            annotation.
+
+    Notes:
+        1. This function does not read the image and ground truth files.
+           The results do not have the "image" and "sem_seg" fields.
+    """
+
+    # We match input images with ground truth based on their relative filepaths (without file
+    # extensions) starting from 'image_root' and 'gt_root' respectively.
+    def file2id(folder_path, file_path):
+        # extract relative path starting from `folder_path`
+        image_id = os.path.normpath(os.path.relpath(file_path, start=folder_path))
+        # remove file extension
+        image_id = os.path.splitext(image_id)[0]
+        return image_id
+
+    input_files = sorted(
+        (os.path.join(image_root, f) for f in PathManager.ls(image_root) if f.endswith(image_ext)),
+        key=lambda file_path: file2id(image_root, file_path),
+    )
+    gt_files = sorted(
+        (os.path.join(gt_root, f) for f in PathManager.ls(gt_root) if f.endswith(gt_ext)),
+        key=lambda file_path: file2id(gt_root, file_path),
+    )
+
+    assert len(gt_files) > 0, "No annotations found in {}.".format(gt_root)
+
+    # Use the intersection, so that val2017_100 annotations can run smoothly with val2017 images
+    if len(input_files) != len(gt_files):
+        logger.warn(
+            "Directory {} and {} has {} and {} files, respectively.".format(
+                image_root, gt_root, len(input_files), len(gt_files)
+            )
+        )
+        input_basenames = [os.path.basename(f)[: -len(image_ext)] for f in input_files]
+        gt_basenames = [os.path.basename(f)[: -len(gt_ext)] for f in gt_files]
+        intersect = list(set(input_basenames) & set(gt_basenames))
+        # sort, otherwise each worker may obtain a list[dict] in different order
+        intersect = sorted(intersect)
+        logger.warn("Will use their intersection of {} files.".format(len(intersect)))
+        input_files = [os.path.join(image_root, f + image_ext) for f in intersect]
+        gt_files = [os.path.join(gt_root, f + gt_ext) for f in intersect]
+
+    logger.info(
+        "Loaded {} images with semantic segmentation from {}".format(len(input_files), image_root)
+    )
+
+    dataset_dicts = []
+    for (img_path, gt_path) in zip(input_files, gt_files):
+        record = {}
+        record["file_name"] = img_path
+        record["sem_seg_file_name"] = gt_path
+        dataset_dicts.append(record)
+
+    return dataset_dicts
+
+
+def convert_to_coco_dict(dataset_name):
+    """
+    Convert an instance detection/segmentation or keypoint detection dataset
+    in detectron2's standard format into COCO json format.
+
+    Generic dataset description can be found here:
+    https://detectron2.readthedocs.io/tutorials/datasets.html#register-a-dataset
+
+    COCO data format description can be found here:
+    http://cocodataset.org/#format-data
+
+    Args:
+        dataset_name (str):
+            name of the source dataset
+            Must be registered in DatastCatalog and in detectron2's standard format.
+            Must have corresponding metadata "thing_classes"
+    Returns:
+        coco_dict: serializable dict in COCO json format
+    """
+
+    dataset_dicts = DatasetCatalog.get(dataset_name)
+    metadata = MetadataCatalog.get(dataset_name)
+
+    # unmap the category mapping ids for COCO
+    if hasattr(metadata, "thing_dataset_id_to_contiguous_id"):
+        reverse_id_mapping = {v: k for k, v in metadata.thing_dataset_id_to_contiguous_id.items()}
+        reverse_id_mapper = lambda contiguous_id: reverse_id_mapping[contiguous_id]  # noqa
+    else:
+        reverse_id_mapper = lambda contiguous_id: contiguous_id  # noqa
+
+    categories = [
+        {"id": reverse_id_mapper(id), "name": name}
+        for id, name in enumerate(metadata.thing_classes)
+    ]
+
+    logger.info("Converting dataset dicts into COCO format")
+    coco_images = []
+    coco_annotations = []
+
+    for image_id, image_dict in enumerate(dataset_dicts):
+        coco_image = {
+            "id": image_dict.get("image_id", image_id),
+            "width": image_dict["width"],
+            "height": image_dict["height"],
+            "file_name": image_dict["file_name"],
+        }
+        coco_images.append(coco_image)
+
+        anns_per_image = image_dict["annotations"]
+        for annotation in anns_per_image:
+            # create a new dict with only COCO fields
+            coco_annotation = {}
+
+            # COCO requirement: XYWH box format
+            bbox = annotation["bbox"]
+            bbox_mode = annotation["bbox_mode"]
+            bbox = BoxMode.convert(bbox, bbox_mode, BoxMode.XYWH_ABS)
+
+            # COCO requirement: instance area
+            if "segmentation" in annotation:
+                # Computing areas for instances by counting the pixels
+                segmentation = annotation["segmentation"]
+                # TODO: check segmentation type: RLE, BinaryMask or Polygon
+                if isinstance(segmentation, list):
+                    polygons = PolygonMasks([segmentation])
+                    area = polygons.area()[0].item()
+                elif isinstance(segmentation, dict):  # RLE
+                    area = mask_util.area(segmentation).item()
+                else:
+                    raise TypeError(f"Unknown segmentation type {type(segmentation)}!")
+            else:
+                # Computing areas using bounding boxes
+                bbox_xy = BoxMode.convert(bbox, BoxMode.XYWH_ABS, BoxMode.XYXY_ABS)
+                area = Boxes([bbox_xy]).area()[0].item()
+
+            if "keypoints" in annotation:
+                keypoints = annotation["keypoints"]  # list[int]
+                for idx, v in enumerate(keypoints):
+                    if idx % 3 != 2:
+                        # COCO's segmentation coordinates are floating points in [0, H or W],
+                        # but keypoint coordinates are integers in [0, H-1 or W-1]
+                        # For COCO format consistency we substract 0.5
+                        # https://github.com/facebookresearch/detectron2/pull/175#issuecomment-551202163
+                        keypoints[idx] = v - 0.5
+                if "num_keypoints" in annotation:
+                    num_keypoints = annotation["num_keypoints"]
+                else:
+                    num_keypoints = sum(kp > 0 for kp in keypoints[2::3])
+
+            # COCO requirement:
+            #   linking annotations to images
+            #   "id" field must start with 1
+            coco_annotation["id"] = len(coco_annotations) + 1
+            coco_annotation["image_id"] = coco_image["id"]
+            coco_annotation["bbox"] = [round(float(x), 3) for x in bbox]
+            coco_annotation["area"] = float(area)
+            coco_annotation["iscrowd"] = annotation.get("iscrowd", 0)
+            coco_annotation["category_id"] = reverse_id_mapper(annotation["category_id"])
+
+            # Add optional fields
+            if "keypoints" in annotation:
+                coco_annotation["keypoints"] = keypoints
+                coco_annotation["num_keypoints"] = num_keypoints
+
+            if "segmentation" in annotation:
+                coco_annotation["segmentation"] = annotation["segmentation"]
+                if isinstance(coco_annotation["segmentation"], dict):  # RLE
+                    coco_annotation["segmentation"]["counts"] = coco_annotation["segmentation"][
+                        "counts"
+                    ].decode("ascii")
+
+            coco_annotations.append(coco_annotation)
+
+    logger.info(
+        "Conversion finished, "
+        f"#images: {len(coco_images)}, #annotations: {len(coco_annotations)}"
+    )
+
+    info = {
+        "date_created": str(datetime.datetime.now()),
+        "description": "Automatically generated COCO json file for Detectron2.",
+    }
+    coco_dict = {
+        "info": info,
+        "images": coco_images,
+        "annotations": coco_annotations,
+        "categories": categories,
+        "licenses": None,
+    }
+    return coco_dict
+
+
+def convert_to_coco_json(dataset_name, output_file, allow_cached=True):
+    """
+    Converts dataset into COCO format and saves it to a json file.
+    dataset_name must be registered in DatasetCatalog and in detectron2's standard format.
+
+    Args:
+        dataset_name:
+            reference from the config file to the catalogs
+            must be registered in DatasetCatalog and in detectron2's standard format
+        output_file: path of json file that will be saved to
+        allow_cached: if json file is already present then skip conversion
+    """
+
+    # TODO: The dataset or the conversion script *may* change,
+    # a checksum would be useful for validating the cached data
+
+    PathManager.mkdirs(os.path.dirname(output_file))
+    with file_lock(output_file):
+        if PathManager.exists(output_file) and allow_cached:
+            logger.warning(
+                f"Using previously cached COCO format annotations at '{output_file}'. "
+                "You need to clear the cache file if your dataset has been modified."
+            )
+        else:
+            logger.info(f"Converting annotations of dataset '{dataset_name}' to COCO format ...)")
+            coco_dict = convert_to_coco_dict(dataset_name)
+
+            logger.info(f"Caching COCO format annotations at '{output_file}' ...")
+            with PathManager.open(output_file, "w") as f:
+                json.dump(coco_dict, f)
+
+
+if __name__ == "__main__":
+    """
+    Test the COCO json dataset loader.
+
+    Usage:
+        python -m detectron2.data.data.coco \
+            path/to/json path/to/image_root dataset_name
+
+        "dataset_name" can be "coco_2014_minival_100", or other
+        pre-registered ones
+    """
+    from detectron2.utils.logger import setup_logger
+    from detectron2.utils.visualizer import Visualizer
+    import detectron2.data.datasets  # noqa # add pre-defined metadata
+    import sys
+
+    logger = setup_logger(name=__name__)
+    assert sys.argv[3] in DatasetCatalog.list()
+    meta = MetadataCatalog.get(sys.argv[3])
+
+    dicts = load_coco_json(sys.argv[1], sys.argv[2], sys.argv[3])
+    logger.info("Done loading {} samples.".format(len(dicts)))
+
+    dirname = "coco-data-vis"
+    os.makedirs(dirname, exist_ok=True)
+    for d in dicts:
+        img = np.array(Image.open(d["file_name"]))
+        visualizer = Visualizer(img, metadata=meta)
+        vis = visualizer.draw_dataset_dict(d)
+        fpath = os.path.join(dirname, os.path.basename(d["file_name"]))
+        vis.save(fpath)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/lvis.py

@@ -0,0 +1,209 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
+import os
+from fvcore.common.file_io import PathManager
+from fvcore.common.timer import Timer
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+from detectron2.structures import BoxMode
+
+from .builtin_meta import _get_coco_instances_meta
+from .lvis_v0_5_categories import LVIS_CATEGORIES
+
+"""
+This file contains functions to parse LVIS-format annotations into dicts in the
+"Detectron2 format".
+"""
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["load_lvis_json", "register_lvis_instances", "get_lvis_instances_meta"]
+
+
+def register_lvis_instances(name, metadata, json_file, image_root):
+    """
+    Register a dataset in LVIS's json annotation format for instance detection and segmentation.
+
+    Args:
+        name (str): a name that identifies the dataset, e.g. "lvis_v0.5_train".
+        metadata (dict): extra metadata associated with this dataset. It can be an empty dict.
+        json_file (str): path to the json instance annotation file.
+        image_root (str or path-like): directory which contains all the images.
+    """
+    DatasetCatalog.register(name, lambda: load_lvis_json(json_file, image_root, name))
+    MetadataCatalog.get(name).set(
+        json_file=json_file, image_root=image_root, evaluator_type="lvis", **metadata
+    )
+
+
+def load_lvis_json(json_file, image_root, dataset_name=None):
+    """
+    Load a json file in LVIS's annotation format.
+
+    Args:
+        json_file (str): full path to the LVIS json annotation file.
+        image_root (str): the directory where the images in this json file exists.
+        dataset_name (str): the name of the dataset (e.g., "lvis_v0.5_train").
+            If provided, this function will put "thing_classes" into the metadata
+            associated with this dataset.
+
+    Returns:
+        list[dict]: a list of dicts in Detectron2 standard format. (See
+        `Using Custom Datasets </tutorials/data.html>`_ )
+
+    Notes:
+        1. This function does not read the image files.
+           The results do not have the "image" field.
+    """
+    from lvis import LVIS
+
+    json_file = PathManager.get_local_path(json_file)
+
+    timer = Timer()
+    lvis_api = LVIS(json_file)
+    if timer.seconds() > 1:
+        logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds()))
+
+    if dataset_name is not None:
+        meta = get_lvis_instances_meta(dataset_name)
+        MetadataCatalog.get(dataset_name).set(**meta)
+
+    # sort indices for reproducible results
+    img_ids = sorted(lvis_api.imgs.keys())
+    # imgs is a list of dicts, each looks something like:
+    # {'license': 4,
+    #  'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
+    #  'file_name': 'COCO_val2014_000000001268.jpg',
+    #  'height': 427,
+    #  'width': 640,
+    #  'date_captured': '2013-11-17 05:57:24',
+    #  'id': 1268}
+    imgs = lvis_api.load_imgs(img_ids)
+    # anns is a list[list[dict]], where each dict is an annotation
+    # record for an object. The inner list enumerates the objects in an image
+    # and the outer list enumerates over images. Example of anns[0]:
+    # [{'segmentation': [[192.81,
+    #     247.09,
+    #     ...
+    #     219.03,
+    #     249.06]],
+    #   'area': 1035.749,
+    #   'image_id': 1268,
+    #   'bbox': [192.81, 224.8, 74.73, 33.43],
+    #   'category_id': 16,
+    #   'id': 42986},
+    #  ...]
+    anns = [lvis_api.img_ann_map[img_id] for img_id in img_ids]
+
+    # Sanity check that each annotation has a unique id
+    ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image]
+    assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique".format(
+        json_file
+    )
+
+    imgs_anns = list(zip(imgs, anns))
+
+    logger.info("Loaded {} images in the LVIS format from {}".format(len(imgs_anns), json_file))
+
+    dataset_dicts = []
+
+    for (img_dict, anno_dict_list) in imgs_anns:
+        record = {}
+        file_name = img_dict["file_name"]
+        if img_dict["file_name"].startswith("COCO"):
+            # Convert form the COCO 2014 file naming convention of
+            # COCO_[train/val/test]2014_000000000000.jpg to the 2017 naming convention of
+            # 000000000000.jpg (LVIS v1 will fix this naming issue)
+            file_name = file_name[-16:]
+        record["file_name"] = os.path.join(image_root, file_name)
+        record["height"] = img_dict["height"]
+        record["width"] = img_dict["width"]
+        record["not_exhaustive_category_ids"] = img_dict.get("not_exhaustive_category_ids", [])
+        record["neg_category_ids"] = img_dict.get("neg_category_ids", [])
+        image_id = record["image_id"] = img_dict["id"]
+
+        objs = []
+        for anno in anno_dict_list:
+            # Check that the image_id in this annotation is the same as
+            # the image_id we're looking at.
+            # This fails only when the data parsing logic or the annotation file is buggy.
+            assert anno["image_id"] == image_id
+            obj = {"bbox": anno["bbox"], "bbox_mode": BoxMode.XYWH_ABS}
+            obj["category_id"] = anno["category_id"] - 1  # Convert 1-indexed to 0-indexed
+            segm = anno["segmentation"]  # list[list[float]]
+            # filter out invalid polygons (< 3 points)
+            valid_segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6]
+            assert len(segm) == len(
+                valid_segm
+            ), "Annotation contains an invalid polygon with < 3 points"
+            assert len(segm) > 0
+            obj["segmentation"] = segm
+            objs.append(obj)
+        record["annotations"] = objs
+        dataset_dicts.append(record)
+
+    return dataset_dicts
+
+
+def get_lvis_instances_meta(dataset_name):
+    """
+    Load LVIS metadata.
+
+    Args:
+        dataset_name (str): LVIS dataset name without the split name (e.g., "lvis_v0.5").
+
+    Returns:
+        dict: LVIS metadata with keys: thing_classes
+    """
+    if "cocofied" in dataset_name:
+        return _get_coco_instances_meta()
+    if "v0.5" in dataset_name:
+        return _get_lvis_instances_meta_v0_5()
+    # There will be a v1 in the future
+    # elif dataset_name == "lvis_v1":
+    #   return get_lvis_instances_meta_v1()
+    raise ValueError("No built-in metadata for dataset {}".format(dataset_name))
+
+
+def _get_lvis_instances_meta_v0_5():
+    assert len(LVIS_CATEGORIES) == 1230
+    cat_ids = [k["id"] for k in LVIS_CATEGORIES]
+    assert min(cat_ids) == 1 and max(cat_ids) == len(
+        cat_ids
+    ), "Category ids are not in [1, #categories], as expected"
+    # Ensure that the category list is sorted by id
+    lvis_categories = sorted(LVIS_CATEGORIES, key=lambda x: x["id"])
+    thing_classes = [k["synonyms"][0] for k in lvis_categories]
+    meta = {"thing_classes": thing_classes}
+    return meta
+
+
+if __name__ == "__main__":
+    """
+    Test the LVIS json dataset loader.
+
+    Usage:
+        python -m detectron2.data.data.lvis \
+            path/to/json path/to/image_root dataset_name vis_limit
+    """
+    import sys
+    import numpy as np
+    from detectron2.utils.logger import setup_logger
+    from PIL import Image
+    import detectron2.data.datasets  # noqa # add pre-defined metadata
+    from detectron2.utils.visualizer import Visualizer
+
+    logger = setup_logger(name=__name__)
+    meta = MetadataCatalog.get(sys.argv[3])
+
+    dicts = load_lvis_json(sys.argv[1], sys.argv[2], sys.argv[3])
+    logger.info("Done loading {} samples.".format(len(dicts)))
+
+    dirname = "lvis-data-vis"
+    os.makedirs(dirname, exist_ok=True)
+    for d in dicts[: int(sys.argv[4])]:
+        img = np.array(Image.open(d["file_name"]))
+        visualizer = Visualizer(img, metadata=meta)
+        vis = visualizer.draw_dataset_dict(d)
+        fpath = os.path.join(dirname, os.path.basename(d["file_name"]))
+        vis.save(fpath)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/pascal_voc.py

@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import numpy as np
+import os
+import xml.etree.ElementTree as ET
+from fvcore.common.file_io import PathManager
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+from detectron2.structures import BoxMode
+
+__all__ = ["register_pascal_voc"]
+
+
+# fmt: off
+CLASS_NAMES = [
+    "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat",
+    "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person",
+    "pottedplant", "sheep", "sofa", "train", "tvmonitor",
+]
+# fmt: on
+
+
+def load_voc_instances(dirname: str, split: str):
+    """
+    Load Pascal VOC detection annotations to Detectron2 format.
+
+    Args:
+        dirname: Contain "Annotations", "ImageSets", "JPEGImages"
+        split (str): one of "train", "test", "val", "trainval"
+    """
+    with PathManager.open(os.path.join(dirname, "ImageSets", "Main", split + ".txt")) as f:
+        fileids = np.loadtxt(f, dtype=np.str)
+
+    # Needs to read many small annotation files. Makes sense at local
+    annotation_dirname = PathManager.get_local_path(os.path.join(dirname, "Annotations/"))
+    dicts = []
+    for fileid in fileids:
+        anno_file = os.path.join(annotation_dirname, fileid + ".xml")
+        jpeg_file = os.path.join(dirname, "JPEGImages", fileid + ".jpg")
+
+        with PathManager.open(anno_file) as f:
+            tree = ET.parse(f)
+
+        r = {
+            "file_name": jpeg_file,
+            "image_id": fileid,
+            "height": int(tree.findall("./size/height")[0].text),
+            "width": int(tree.findall("./size/width")[0].text),
+        }
+        instances = []
+
+        for obj in tree.findall("object"):
+            cls = obj.find("name").text
+            # We include "difficult" samples in training.
+            # Based on limited experiments, they don't hurt accuracy.
+            # difficult = int(obj.find("difficult").text)
+            # if difficult == 1:
+            # continue
+            bbox = obj.find("bndbox")
+            bbox = [float(bbox.find(x).text) for x in ["xmin", "ymin", "xmax", "ymax"]]
+            # Original annotations are integers in the range [1, W or H]
+            # Assuming they mean 1-based pixel indices (inclusive),
+            # a box with annotation (xmin=1, xmax=W) covers the whole image.
+            # In coordinate space this is represented by (xmin=0, xmax=W)
+            bbox[0] -= 1.0
+            bbox[1] -= 1.0
+            instances.append(
+                {"category_id": CLASS_NAMES.index(cls), "bbox": bbox, "bbox_mode": BoxMode.XYXY_ABS}
+            )
+        r["annotations"] = instances
+        dicts.append(r)
+    return dicts
+
+
+def register_pascal_voc(name, dirname, split, year):
+    DatasetCatalog.register(name, lambda: load_voc_instances(dirname, split))
+    MetadataCatalog.get(name).set(
+        thing_classes=CLASS_NAMES, dirname=dirname, year=year, split=split
+    )

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/datasets/register_coco.py

@@ -0,0 +1,129 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import os
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+
+from .coco import load_coco_json, load_sem_seg
+
+"""
+This file contains functions to register a COCO-format dataset to the DatasetCatalog.
+"""
+
+__all__ = ["register_coco_instances", "register_coco_panoptic_separated"]
+
+
+def register_coco_instances(name, metadata, json_file, image_root):
+    """
+    Register a dataset in COCO's json annotation format for
+    instance detection, instance segmentation and keypoint detection.
+    (i.e., Type 1 and 2 in http://cocodataset.org/#format-data.
+    `instances*.json` and `person_keypoints*.json` in the dataset).
+
+    This is an example of how to register a new dataset.
+    You can do something similar to this function, to register new data.
+
+    Args:
+        name (str): the name that identifies a dataset, e.g. "coco_2014_train".
+        metadata (dict): extra metadata associated with this dataset.  You can
+            leave it as an empty dict.
+        json_file (str): path to the json instance annotation file.
+        image_root (str or path-like): directory which contains all the images.
+    """
+    assert isinstance(name, str), name
+    assert isinstance(json_file, (str, os.PathLike)), json_file
+    assert isinstance(image_root, (str, os.PathLike)), image_root
+    # 1. register a function which returns dicts
+    DatasetCatalog.register(name, lambda: load_coco_json(json_file, image_root, name))
+
+    # 2. Optionally, add metadata about this dataset,
+    # since they might be useful in evaluation, visualization or logging
+    MetadataCatalog.get(name).set(
+        json_file=json_file, image_root=image_root, evaluator_type="coco", **metadata
+    )
+
+
+def register_coco_panoptic_separated(
+    name, metadata, image_root, panoptic_root, panoptic_json, sem_seg_root, instances_json
+):
+    """
+    Register a COCO panoptic segmentation dataset named `name`.
+    The annotations in this registered dataset will contain both instance annotations and
+    semantic annotations, each with its own contiguous ids. Hence it's called "separated".
+
+    It follows the setting used by the PanopticFPN paper:
+
+    1. The instance annotations directly come from polygons in the COCO
+       instances annotation task, rather than from the masks in the COCO panoptic annotations.
+
+       The two format have small differences:
+       Polygons in the instance annotations may have overlaps.
+       The mask annotations are produced by labeling the overlapped polygons
+       with depth ordering.
+
+    2. The semantic annotations are converted from panoptic annotations, where
+       all "things" are assigned a semantic id of 0.
+       All semantic categories will therefore have ids in contiguous
+       range [1, #stuff_categories].
+
+    This function will also register a pure semantic segmentation dataset
+    named ``name + '_stuffonly'``.
+
+    Args:
+        name (str): the name that identifies a dataset,
+            e.g. "coco_2017_train_panoptic"
+        metadata (dict): extra metadata associated with this dataset.
+        image_root (str): directory which contains all the images
+        panoptic_root (str): directory which contains panoptic annotation images
+        panoptic_json (str): path to the json panoptic annotation file
+        sem_seg_root (str): directory which contains all the ground truth segmentation annotations.
+        instances_json (str): path to the json instance annotation file
+    """
+    panoptic_name = name + "_separated"
+    DatasetCatalog.register(
+        panoptic_name,
+        lambda: merge_to_panoptic(
+            load_coco_json(instances_json, image_root, panoptic_name),
+            load_sem_seg(sem_seg_root, image_root),
+        ),
+    )
+    MetadataCatalog.get(panoptic_name).set(
+        panoptic_root=panoptic_root,
+        image_root=image_root,
+        panoptic_json=panoptic_json,
+        sem_seg_root=sem_seg_root,
+        json_file=instances_json,  # TODO rename
+        evaluator_type="coco_panoptic_seg",
+        **metadata
+    )
+
+    semantic_name = name + "_stuffonly"
+    DatasetCatalog.register(semantic_name, lambda: load_sem_seg(sem_seg_root, image_root))
+    MetadataCatalog.get(semantic_name).set(
+        sem_seg_root=sem_seg_root, image_root=image_root, evaluator_type="sem_seg", **metadata
+    )
+
+
+def merge_to_panoptic(detection_dicts, sem_seg_dicts):
+    """
+    Create dataset dicts for panoptic segmentation, by
+    merging two dicts using "file_name" field to match their entries.
+
+    Args:
+        detection_dicts (list[dict]): lists of dicts for object detection or instance segmentation.
+        sem_seg_dicts (list[dict]): lists of dicts for semantic segmentation.
+
+    Returns:
+        list[dict] (one per input image): Each dict contains all (key, value) pairs from dicts in
+            both detection_dicts and sem_seg_dicts that correspond to the same image.
+            The function assumes that the same key in different dicts has the same value.
+    """
+    results = []
+    sem_seg_file_to_entry = {x["file_name"]: x for x in sem_seg_dicts}
+    assert len(sem_seg_file_to_entry) > 0
+
+    for det_dict in detection_dicts:
+        dic = copy.copy(det_dict)
+        dic.update(sem_seg_file_to_entry[dic["file_name"]])
+        results.append(dic)
+    return results

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/detection_utils.py

@@ -0,0 +1,516 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+"""
+Common data processing utilities that are used in a
+typical object detection data pipeline.
+"""
+import logging
+import numpy as np
+import pycocotools.mask as mask_util
+import torch
+from fvcore.common.file_io import PathManager
+from PIL import Image, ImageOps
+
+from detectron2.structures import (
+    BitMasks,
+    Boxes,
+    BoxMode,
+    Instances,
+    Keypoints,
+    PolygonMasks,
+    RotatedBoxes,
+    polygons_to_bitmask,
+)
+
+from . import transforms as T
+from .catalog import MetadataCatalog
+
+
+class SizeMismatchError(ValueError):
+    """
+    When loaded image has difference width/height compared with annotation.
+    """
+
+
+# https://en.wikipedia.org/wiki/YUV#SDTV_with_BT.601
+_M_RGB2YUV = [[0.299, 0.587, 0.114], [-0.14713, -0.28886, 0.436], [0.615, -0.51499, -0.10001]]
+_M_YUV2RGB = [[1.0, 0.0, 1.13983], [1.0, -0.39465, -0.58060], [1.0, 2.03211, 0.0]]
+
+
+def convert_PIL_to_numpy(image, format):
+    """
+    Convert PIL image to numpy array of target format.
+
+    Args:
+        image (PIL.Image): a PIL image
+        format (str): the format of output image
+
+    Returns:
+        (np.ndarray): also see `read_image`
+    """
+    if format is not None:
+        # PIL only supports RGB, so convert to RGB and flip channels over below
+        conversion_format = format
+        if format in ["BGR", "YUV-BT.601"]:
+            conversion_format = "RGB"
+        image = image.convert(conversion_format)
+    image = np.asarray(image)
+    # PIL squeezes out the channel dimension for "L", so make it HWC
+    if format == "L":
+        image = np.expand_dims(image, -1)
+
+    # handle formats not supported by PIL
+    elif format == "BGR":
+        # flip channels if needed
+        image = image[:, :, ::-1]
+    elif format == "YUV-BT.601":
+        image = image / 255.0
+        image = np.dot(image, np.array(_M_RGB2YUV).T)
+
+    return image
+
+
+def convert_image_to_rgb(image, format):
+    """
+    Convert numpy image from given format to RGB.
+
+    Args:
+        image (np.ndarray): a numpy image
+        format (str): the format of input image, also see `read_image`
+
+    Returns:
+        (np.ndarray): HWC RGB image in 0-255 range, can be either float or uint8
+    """
+    if format == "BGR":
+        image = image[:, :, [2, 1, 0]]
+    elif format == "YUV-BT.601":
+        image = np.dot(image, np.array(_M_YUV2RGB).T)
+        image = image * 255.0
+    else:
+        if format == "L":
+            image = image[:, :, 0]
+        image = image.astype(np.uint8)
+        image = np.asarray(Image.fromarray(image, mode=format).convert("RGB"))
+    return image
+
+
+def read_image(file_name, format=None):
+    """
+    Read an image into the given format.
+    Will apply rotation and flipping if the image has such exif information.
+
+    Args:
+        file_name (str): image file path
+        format (str): one of the supported image modes in PIL, or "BGR" or "YUV-BT.601"
+
+    Returns:
+        image (np.ndarray): an HWC image in the given format, which is 0-255, uint8 for
+            supported image modes in PIL or "BGR"; float (0-1 for Y) for YUV-BT.601.
+    """
+    with PathManager.open(file_name, "rb") as f:
+        image = Image.open(f)
+
+        # capture and ignore this bug: https://github.com/python-pillow/Pillow/issues/3973
+        try:
+            image = ImageOps.exif_transpose(image)
+        except Exception:
+            pass
+
+        return convert_PIL_to_numpy(image, format)
+
+
+def check_image_size(dataset_dict, image):
+    """
+    Raise an error if the image does not match the size specified in the dict.
+    """
+    if "width" in dataset_dict or "height" in dataset_dict:
+        image_wh = (image.shape[1], image.shape[0])
+        expected_wh = (dataset_dict["width"], dataset_dict["height"])
+        if not image_wh == expected_wh:
+            raise SizeMismatchError(
+                "Mismatched (W,H){}, got {}, expect {}".format(
+                    " for image " + dataset_dict["file_name"]
+                    if "file_name" in dataset_dict
+                    else "",
+                    image_wh,
+                    expected_wh,
+                )
+            )
+
+    # To ensure bbox always remap to original image size
+    if "width" not in dataset_dict:
+        dataset_dict["width"] = image.shape[1]
+    if "height" not in dataset_dict:
+        dataset_dict["height"] = image.shape[0]
+
+
+def transform_proposals(dataset_dict, image_shape, transforms, min_box_side_len, proposal_topk):
+    """
+    Apply transformations to the proposals in dataset_dict, if any.
+
+    Args:
+        dataset_dict (dict): a dict read from the dataset, possibly
+            contains fields "proposal_boxes", "proposal_objectness_logits", "proposal_bbox_mode"
+        image_shape (tuple): height, width
+        transforms (TransformList):
+        min_box_side_len (int): keep proposals with at least this size
+        proposal_topk (int): only keep top-K scoring proposals
+
+    The input dict is modified in-place, with abovementioned keys removed. A new
+    key "proposals" will be added. Its value is an `Instances`
+    object which contains the transformed proposals in its field
+    "proposal_boxes" and "objectness_logits".
+    """
+    if "proposal_boxes" in dataset_dict:
+        # Transform proposal boxes
+        boxes = transforms.apply_box(
+            BoxMode.convert(
+                dataset_dict.pop("proposal_boxes"),
+                dataset_dict.pop("proposal_bbox_mode"),
+                BoxMode.XYXY_ABS,
+            )
+        )
+        boxes = Boxes(boxes)
+        objectness_logits = torch.as_tensor(
+            dataset_dict.pop("proposal_objectness_logits").astype("float32")
+        )
+
+        boxes.clip(image_shape)
+        keep = boxes.nonempty(threshold=min_box_side_len)
+        boxes = boxes[keep]
+        objectness_logits = objectness_logits[keep]
+
+        proposals = Instances(image_shape)
+        proposals.proposal_boxes = boxes[:proposal_topk]
+        proposals.objectness_logits = objectness_logits[:proposal_topk]
+        dataset_dict["proposals"] = proposals
+
+
+def transform_instance_annotations(
+    annotation, transforms, image_size, *, keypoint_hflip_indices=None
+):
+    """
+    Apply transforms to box, segmentation and keypoints annotations of a single instance.
+
+    It will use `transforms.apply_box` for the box, and
+    `transforms.apply_coords` for segmentation polygons & keypoints.
+    If you need anything more specially designed for each data structure,
+    you'll need to implement your own version of this function or the transforms.
+
+    Args:
+        annotation (dict): dict of instance annotations for a single instance.
+            It will be modified in-place.
+        transforms (TransformList):
+        image_size (tuple): the height, width of the transformed image
+        keypoint_hflip_indices (ndarray[int]): see `create_keypoint_hflip_indices`.
+
+    Returns:
+        dict:
+            the same input dict with fields "bbox", "segmentation", "keypoints"
+            transformed according to `transforms`.
+            The "bbox_mode" field will be set to XYXY_ABS.
+    """
+    bbox = BoxMode.convert(annotation["bbox"], annotation["bbox_mode"], BoxMode.XYXY_ABS)
+    # Note that bbox is 1d (per-instance bounding box)
+    annotation["bbox"] = transforms.apply_box([bbox])[0]
+    annotation["bbox_mode"] = BoxMode.XYXY_ABS
+
+    if "segmentation" in annotation:
+        # each instance contains 1 or more polygons
+        segm = annotation["segmentation"]
+        if isinstance(segm, list):
+            # polygons
+            polygons = [np.asarray(p).reshape(-1, 2) for p in segm]
+            annotation["segmentation"] = [
+                p.reshape(-1) for p in transforms.apply_polygons(polygons)
+            ]
+        elif isinstance(segm, dict):
+            # RLE
+            mask = mask_util.decode(segm)
+            mask = transforms.apply_segmentation(mask)
+            assert tuple(mask.shape[:2]) == image_size
+            annotation["segmentation"] = mask
+        else:
+            raise ValueError(
+                "Cannot transform segmentation of type '{}'!"
+                "Supported types are: polygons as list[list[float] or ndarray],"
+                " COCO-style RLE as a dict.".format(type(segm))
+            )
+
+    if "keypoints" in annotation:
+        keypoints = transform_keypoint_annotations(
+            annotation["keypoints"], transforms, image_size, keypoint_hflip_indices
+        )
+        annotation["keypoints"] = keypoints
+
+    return annotation
+
+
+def transform_keypoint_annotations(keypoints, transforms, image_size, keypoint_hflip_indices=None):
+    """
+    Transform keypoint annotations of an image.
+
+    Args:
+        keypoints (list[float]): Nx3 float in Detectron2 Dataset format.
+        transforms (TransformList):
+        image_size (tuple): the height, width of the transformed image
+        keypoint_hflip_indices (ndarray[int]): see `create_keypoint_hflip_indices`.
+    """
+    # (N*3,) -> (N, 3)
+    keypoints = np.asarray(keypoints, dtype="float64").reshape(-1, 3)
+    keypoints[:, :2] = transforms.apply_coords(keypoints[:, :2])
+
+    # This assumes that HorizFlipTransform is the only one that does flip
+    do_hflip = sum(isinstance(t, T.HFlipTransform) for t in transforms.transforms) % 2 == 1
+
+    # Alternative way: check if probe points was horizontally flipped.
+    # probe = np.asarray([[0.0, 0.0], [image_width, 0.0]])
+    # probe_aug = transforms.apply_coords(probe.copy())
+    # do_hflip = np.sign(probe[1][0] - probe[0][0]) != np.sign(probe_aug[1][0] - probe_aug[0][0])  # noqa
+
+    # If flipped, swap each keypoint with its opposite-handed equivalent
+    if do_hflip:
+        assert keypoint_hflip_indices is not None
+        keypoints = keypoints[keypoint_hflip_indices, :]
+
+    # Maintain COCO convention that if visibility == 0, then x, y = 0
+    # TODO may need to reset visibility for cropped keypoints,
+    # but it does not matter for our existing algorithms
+    keypoints[keypoints[:, 2] == 0] = 0
+    return keypoints
+
+
+def annotations_to_instances(annos, image_size, mask_format="polygon"):
+    """
+    Create an :class:`Instances` object used by the models,
+    from instance annotations in the dataset dict.
+
+    Args:
+        annos (list[dict]): a list of instance annotations in one image, each
+            element for one instance.
+        image_size (tuple): height, width
+
+    Returns:
+        Instances:
+            It will contain fields "gt_boxes", "gt_classes",
+            "gt_masks", "gt_keypoints", if they can be obtained from `annos`.
+            This is the format that builtin models expect.
+    """
+    boxes = [BoxMode.convert(obj["bbox"], obj["bbox_mode"], BoxMode.XYXY_ABS) for obj in annos]
+    target = Instances(image_size)
+    boxes = target.gt_boxes = Boxes(boxes)
+    boxes.clip(image_size)
+
+    classes = [obj["category_id"] for obj in annos]
+    classes = torch.tensor(classes, dtype=torch.int64)
+    target.gt_classes = classes
+
+    if len(annos) and "segmentation" in annos[0]:
+        segms = [obj["segmentation"] for obj in annos]
+        if mask_format == "polygon":
+            masks = PolygonMasks(segms)
+        else:
+            assert mask_format == "bitmask", mask_format
+            masks = []
+            for segm in segms:
+                if isinstance(segm, list):
+                    # polygon
+                    masks.append(polygons_to_bitmask(segm, *image_size))
+                elif isinstance(segm, dict):
+                    # COCO RLE
+                    masks.append(mask_util.decode(segm))
+                elif isinstance(segm, np.ndarray):
+                    assert segm.ndim == 2, "Expect segmentation of 2 dimensions, got {}.".format(
+                        segm.ndim
+                    )
+                    # mask array
+                    masks.append(segm)
+                else:
+                    raise ValueError(
+                        "Cannot convert segmentation of type '{}' to BitMasks!"
+                        "Supported types are: polygons as list[list[float] or ndarray],"
+                        " COCO-style RLE as a dict, or a full-image segmentation mask "
+                        "as a 2D ndarray.".format(type(segm))
+                    )
+            # torch.from_numpy does not support array with negative stride.
+            masks = BitMasks(
+                torch.stack([torch.from_numpy(np.ascontiguousarray(x)) for x in masks])
+            )
+        target.gt_masks = masks
+
+    if len(annos) and "keypoints" in annos[0]:
+        kpts = [obj.get("keypoints", []) for obj in annos]
+        target.gt_keypoints = Keypoints(kpts)
+
+    return target
+
+
+def annotations_to_instances_rotated(annos, image_size):
+    """
+    Create an :class:`Instances` object used by the models,
+    from instance annotations in the dataset dict.
+    Compared to `annotations_to_instances`, this function is for rotated boxes only
+
+    Args:
+        annos (list[dict]): a list of instance annotations in one image, each
+            element for one instance.
+        image_size (tuple): height, width
+
+    Returns:
+        Instances:
+            Containing fields "gt_boxes", "gt_classes",
+            if they can be obtained from `annos`.
+            This is the format that builtin models expect.
+    """
+    boxes = [obj["bbox"] for obj in annos]
+    target = Instances(image_size)
+    boxes = target.gt_boxes = RotatedBoxes(boxes)
+    boxes.clip(image_size)
+
+    classes = [obj["category_id"] for obj in annos]
+    classes = torch.tensor(classes, dtype=torch.int64)
+    target.gt_classes = classes
+
+    return target
+
+
+def filter_empty_instances(instances, by_box=True, by_mask=True, box_threshold=1e-5):
+    """
+    Filter out empty instances in an `Instances` object.
+
+    Args:
+        instances (Instances):
+        by_box (bool): whether to filter out instances with empty boxes
+        by_mask (bool): whether to filter out instances with empty masks
+        box_threshold (float): minimum width and height to be considered non-empty
+
+    Returns:
+        Instances: the filtered instances.
+    """
+    assert by_box or by_mask
+    r = []
+    if by_box:
+        r.append(instances.gt_boxes.nonempty(threshold=box_threshold))
+    if instances.has("gt_masks") and by_mask:
+        r.append(instances.gt_masks.nonempty())
+
+    # TODO: can also filter visible keypoints
+
+    if not r:
+        return instances
+    m = r[0]
+    for x in r[1:]:
+        m = m & x
+    return instances[m]
+
+
+def create_keypoint_hflip_indices(dataset_names):
+    """
+    Args:
+        dataset_names (list[str]): list of dataset names
+    Returns:
+        ndarray[int]: a vector of size=#keypoints, storing the
+        horizontally-flipped keypoint indices.
+    """
+
+    check_metadata_consistency("keypoint_names", dataset_names)
+    check_metadata_consistency("keypoint_flip_map", dataset_names)
+
+    meta = MetadataCatalog.get(dataset_names[0])
+    names = meta.keypoint_names
+    # TODO flip -> hflip
+    flip_map = dict(meta.keypoint_flip_map)
+    flip_map.update({v: k for k, v in flip_map.items()})
+    flipped_names = [i if i not in flip_map else flip_map[i] for i in names]
+    flip_indices = [names.index(i) for i in flipped_names]
+    return np.asarray(flip_indices)
+
+
+def gen_crop_transform_with_instance(crop_size, image_size, instance):
+    """
+    Generate a CropTransform so that the cropping region contains
+    the center of the given instance.
+
+    Args:
+        crop_size (tuple): h, w in pixels
+        image_size (tuple): h, w
+        instance (dict): an annotation dict of one instance, in Detectron2's
+            dataset format.
+    """
+    crop_size = np.asarray(crop_size, dtype=np.int32)
+    bbox = BoxMode.convert(instance["bbox"], instance["bbox_mode"], BoxMode.XYXY_ABS)
+    center_yx = (bbox[1] + bbox[3]) * 0.5, (bbox[0] + bbox[2]) * 0.5
+    assert (
+        image_size[0] >= center_yx[0] and image_size[1] >= center_yx[1]
+    ), "The annotation bounding box is outside of the image!"
+    assert (
+        image_size[0] >= crop_size[0] and image_size[1] >= crop_size[1]
+    ), "Crop size is larger than image size!"
+
+    min_yx = np.maximum(np.floor(center_yx).astype(np.int32) - crop_size, 0)
+    max_yx = np.maximum(np.asarray(image_size, dtype=np.int32) - crop_size, 0)
+    max_yx = np.minimum(max_yx, np.ceil(center_yx).astype(np.int32))
+
+    y0 = np.random.randint(min_yx[0], max_yx[0] + 1)
+    x0 = np.random.randint(min_yx[1], max_yx[1] + 1)
+    return T.CropTransform(x0, y0, crop_size[1], crop_size[0])
+
+
+def check_metadata_consistency(key, dataset_names):
+    """
+    Check that the data have consistent metadata.
+
+    Args:
+        key (str): a metadata key
+        dataset_names (list[str]): a list of dataset names
+
+    Raises:
+        AttributeError: if the key does not exist in the metadata
+        ValueError: if the given data do not have the same metadata values defined by key
+    """
+    if len(dataset_names) == 0:
+        return
+    logger = logging.getLogger(__name__)
+    entries_per_dataset = [getattr(MetadataCatalog.get(d), key) for d in dataset_names]
+    for idx, entry in enumerate(entries_per_dataset):
+        if entry != entries_per_dataset[0]:
+            logger.error(
+                "Metadata '{}' for dataset '{}' is '{}'".format(key, dataset_names[idx], str(entry))
+            )
+            logger.error(
+                "Metadata '{}' for dataset '{}' is '{}'".format(
+                    key, dataset_names[0], str(entries_per_dataset[0])
+                )
+            )
+            raise ValueError("Datasets have different metadata '{}'!".format(key))
+
+
+def build_transform_gen(cfg, is_train):
+    """
+    Create a list of :class:`TransformGen` from config.
+    Now it includes resizing and flipping.
+
+    Returns:
+        list[TransformGen]
+    """
+    if is_train:
+        min_size = cfg.INPUT.MIN_SIZE_TRAIN
+        max_size = cfg.INPUT.MAX_SIZE_TRAIN
+        sample_style = cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING
+    else:
+        min_size = cfg.INPUT.MIN_SIZE_TEST
+        max_size = cfg.INPUT.MAX_SIZE_TEST
+        sample_style = "choice"
+    if sample_style == "range":
+        assert len(min_size) == 2, "more than 2 ({}) min_size(s) are provided for ranges".format(
+            len(min_size)
+        )
+
+    logger = logging.getLogger(__name__)
+    tfm_gens = []
+    tfm_gens.append(T.ResizeShortestEdge(min_size, max_size, sample_style))
+    if is_train:
+        tfm_gens.append(T.RandomFlip())
+        logger.info("TransformGens used in training: " + str(tfm_gens))
+    return tfm_gens

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/samplers/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .distributed_sampler import InferenceSampler, RepeatFactorTrainingSampler, TrainingSampler
+from .grouped_batch_sampler import GroupedBatchSampler
+
+__all__ = [
+    "GroupedBatchSampler",
+    "TrainingSampler",
+    "InferenceSampler",
+    "RepeatFactorTrainingSampler",
+]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/samplers/distributed_sampler.py

@@ -0,0 +1,199 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import itertools
+import math
+from collections import defaultdict
+from typing import Optional
+import torch
+from torch.utils.data.sampler import Sampler
+
+from detectron2.utils import comm
+
+
+class TrainingSampler(Sampler):
+    """
+    In training, we only care about the "infinite stream" of training data.
+    So this sampler produces an infinite stream of indices and
+    all workers cooperate to correctly shuffle the indices and sample different indices.
+
+    The samplers in each worker effectively produces `indices[worker_id::num_workers]`
+    where `indices` is an infinite stream of indices consisting of
+    `shuffle(range(size)) + shuffle(range(size)) + ...` (if shuffle is True)
+    or `range(size) + range(size) + ...` (if shuffle is False)
+    """
+
+    def __init__(self, size: int, shuffle: bool = True, seed: Optional[int] = None):
+        """
+        Args:
+            size (int): the total number of data of the underlying dataset to sample from
+            shuffle (bool): whether to shuffle the indices or not
+            seed (int): the initial seed of the shuffle. Must be the same
+                across all workers. If None, will use a random seed shared
+                among workers (require synchronization among all workers).
+        """
+        self._size = size
+        assert size > 0
+        self._shuffle = shuffle
+        if seed is None:
+            seed = comm.shared_random_seed()
+        self._seed = int(seed)
+
+        self._rank = comm.get_rank()
+        self._world_size = comm.get_world_size()
+
+    def __iter__(self):
+        start = self._rank
+        yield from itertools.islice(self._infinite_indices(), start, None, self._world_size)
+
+    def _infinite_indices(self):
+        g = torch.Generator()
+        g.manual_seed(self._seed)
+        while True:
+            if self._shuffle:
+                yield from torch.randperm(self._size, generator=g)
+            else:
+                yield from torch.arange(self._size)
+
+
+class RepeatFactorTrainingSampler(Sampler):
+    """
+    Similar to TrainingSampler, but suitable for training on class imbalanced data
+    like LVIS. In each epoch, an image may appear multiple times based on its "repeat
+    factor". The repeat factor for an image is a function of the frequency the rarest
+    category labeled in that image. The "frequency of category c" in [0, 1] is defined
+    as the fraction of images in the training set (without repeats) in which category c
+    appears.
+
+    See :paper:`lvis` (>= v2) Appendix B.2.
+    """
+
+    def __init__(self, dataset_dicts, repeat_thresh, shuffle=True, seed=None):
+        """
+        Args:
+            dataset_dicts (list[dict]): annotations in Detectron2 dataset format.
+            repeat_thresh (float): frequency threshold below which data is repeated.
+            shuffle (bool): whether to shuffle the indices or not
+            seed (int): the initial seed of the shuffle. Must be the same
+                across all workers. If None, will use a random seed shared
+                among workers (require synchronization among all workers).
+        """
+        self._shuffle = shuffle
+        if seed is None:
+            seed = comm.shared_random_seed()
+        self._seed = int(seed)
+
+        self._rank = comm.get_rank()
+        self._world_size = comm.get_world_size()
+
+        # Get fractional repeat factors and split into whole number (_int_part)
+        # and fractional (_frac_part) parts.
+        rep_factors = self._get_repeat_factors(dataset_dicts, repeat_thresh)
+        self._int_part = torch.trunc(rep_factors)
+        self._frac_part = rep_factors - self._int_part
+
+    def _get_repeat_factors(self, dataset_dicts, repeat_thresh):
+        """
+        Compute (fractional) per-image repeat factors.
+
+        Args:
+            See __init__.
+
+        Returns:
+            torch.Tensor: the i-th element is the repeat factor for the dataset image
+                at index i.
+        """
+        # 1. For each category c, compute the fraction of images that contain it: f(c)
+        category_freq = defaultdict(int)
+        for dataset_dict in dataset_dicts:  # For each image (without repeats)
+            cat_ids = {ann["category_id"] for ann in dataset_dict["annotations"]}
+            for cat_id in cat_ids:
+                category_freq[cat_id] += 1
+        num_images = len(dataset_dicts)
+        for k, v in category_freq.items():
+            category_freq[k] = v / num_images
+
+        # 2. For each category c, compute the category-level repeat factor:
+        #    r(c) = max(1, sqrt(t / f(c)))
+        category_rep = {
+            cat_id: max(1.0, math.sqrt(repeat_thresh / cat_freq))
+            for cat_id, cat_freq in category_freq.items()
+        }
+
+        # 3. For each image I, compute the image-level repeat factor:
+        #    r(I) = max_{c in I} r(c)
+        rep_factors = []
+        for dataset_dict in dataset_dicts:
+            cat_ids = {ann["category_id"] for ann in dataset_dict["annotations"]}
+            rep_factor = max({category_rep[cat_id] for cat_id in cat_ids})
+            rep_factors.append(rep_factor)
+
+        return torch.tensor(rep_factors, dtype=torch.float32)
+
+    def _get_epoch_indices(self, generator):
+        """
+        Create a list of dataset indices (with repeats) to use for one epoch.
+
+        Args:
+            generator (torch.Generator): pseudo random number generator used for
+                stochastic rounding.
+
+        Returns:
+            torch.Tensor: list of dataset indices to use in one epoch. Each index
+                is repeated based on its calculated repeat factor.
+        """
+        # Since repeat factors are fractional, we use stochastic rounding so
+        # that the target repeat factor is achieved in expectation over the
+        # course of training
+        rands = torch.rand(len(self._frac_part), generator=generator)
+        rep_factors = self._int_part + (rands < self._frac_part).float()
+        # Construct a list of indices in which we repeat images as specified
+        indices = []
+        for dataset_index, rep_factor in enumerate(rep_factors):
+            indices.extend([dataset_index] * int(rep_factor.item()))
+        return torch.tensor(indices, dtype=torch.int64)
+
+    def __iter__(self):
+        start = self._rank
+        yield from itertools.islice(self._infinite_indices(), start, None, self._world_size)
+
+    def _infinite_indices(self):
+        g = torch.Generator()
+        g.manual_seed(self._seed)
+        while True:
+            # Sample indices with repeats determined by stochastic rounding; each
+            # "epoch" may have a slightly different size due to the rounding.
+            indices = self._get_epoch_indices(g)
+            if self._shuffle:
+                randperm = torch.randperm(len(indices), generator=g)
+                yield from indices[randperm]
+            else:
+                yield from indices
+
+
+class InferenceSampler(Sampler):
+    """
+    Produce indices for inference.
+    Inference needs to run on the __exact__ set of samples,
+    therefore when the total number of samples is not divisible by the number of workers,
+    this sampler produces different number of samples on different workers.
+    """
+
+    def __init__(self, size: int):
+        """
+        Args:
+            size (int): the total number of data of the underlying dataset to sample from
+        """
+        self._size = size
+        assert size > 0
+        self._rank = comm.get_rank()
+        self._world_size = comm.get_world_size()
+
+        shard_size = (self._size - 1) // self._world_size + 1
+        begin = shard_size * self._rank
+        end = min(shard_size * (self._rank + 1), self._size)
+        self._local_indices = range(begin, end)
+
+    def __iter__(self):
+        yield from self._local_indices
+
+    def __len__(self):
+        return len(self._local_indices)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/samplers/grouped_batch_sampler.py

@@ -0,0 +1,47 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import numpy as np
+from torch.utils.data.sampler import BatchSampler, Sampler
+
+
+class GroupedBatchSampler(BatchSampler):
+    """
+    Wraps another sampler to yield a mini-batch of indices.
+    It enforces that the batch only contain elements from the same group.
+    It also tries to provide mini-batches which follows an ordering which is
+    as close as possible to the ordering from the original sampler.
+    """
+
+    def __init__(self, sampler, group_ids, batch_size):
+        """
+        Args:
+            sampler (Sampler): Base sampler.
+            group_ids (list[int]): If the sampler produces indices in range [0, N),
+                `group_ids` must be a list of `N` ints which contains the group id of each sample.
+                The group ids must be a set of integers in the range [0, num_groups).
+            batch_size (int): Size of mini-batch.
+        """
+        if not isinstance(sampler, Sampler):
+            raise ValueError(
+                "sampler should be an instance of "
+                "torch.utils.data.Sampler, but got sampler={}".format(sampler)
+            )
+        self.sampler = sampler
+        self.group_ids = np.asarray(group_ids)
+        assert self.group_ids.ndim == 1
+        self.batch_size = batch_size
+        groups = np.unique(self.group_ids).tolist()
+
+        # buffer the indices of each group until batch size is reached
+        self.buffer_per_group = {k: [] for k in groups}
+
+    def __iter__(self):
+        for idx in self.sampler:
+            group_id = self.group_ids[idx]
+            group_buffer = self.buffer_per_group[group_id]
+            group_buffer.append(idx)
+            if len(group_buffer) == self.batch_size:
+                yield group_buffer[:]  # yield a copy of the list
+                del group_buffer[:]
+
+    def __len__(self):
+        raise NotImplementedError("len() of GroupedBatchSampler is not well-defined.")

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/transforms/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .transform import *
+from fvcore.transforms.transform import *
+from .transform_gen import *
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/transforms/transform.py

@@ -0,0 +1,241 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+# File: transform.py
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from fvcore.transforms.transform import HFlipTransform, NoOpTransform, Transform
+from PIL import Image
+
+try:
+    import cv2  # noqa
+except ImportError:
+    # OpenCV is an optional dependency at the moment
+    pass
+
+__all__ = ["ExtentTransform", "ResizeTransform", "RotationTransform"]
+
+
+class ExtentTransform(Transform):
+    """
+    Extracts a subregion from the source image and scales it to the output size.
+
+    The fill color is used to map pixels from the source rect that fall outside
+    the source image.
+
+    See: https://pillow.readthedocs.io/en/latest/PIL.html#PIL.ImageTransform.ExtentTransform
+    """
+
+    def __init__(self, src_rect, output_size, interp=Image.LINEAR, fill=0):
+        """
+        Args:
+            src_rect (x0, y0, x1, y1): src coordinates
+            output_size (h, w): dst image size
+            interp: PIL interpolation methods
+            fill: Fill color used when src_rect extends outside image
+        """
+        super().__init__()
+        self._set_attributes(locals())
+
+    def apply_image(self, img, interp=None):
+        h, w = self.output_size
+        ret = Image.fromarray(img).transform(
+            size=(w, h),
+            method=Image.EXTENT,
+            data=self.src_rect,
+            resample=interp if interp else self.interp,
+            fill=self.fill,
+        )
+        return np.asarray(ret)
+
+    def apply_coords(self, coords):
+        # Transform image center from source coordinates into output coordinates
+        # and then map the new origin to the corner of the output image.
+        h, w = self.output_size
+        x0, y0, x1, y1 = self.src_rect
+        new_coords = coords.astype(np.float32)
+        new_coords[:, 0] -= 0.5 * (x0 + x1)
+        new_coords[:, 1] -= 0.5 * (y0 + y1)
+        new_coords[:, 0] *= w / (x1 - x0)
+        new_coords[:, 1] *= h / (y1 - y0)
+        new_coords[:, 0] += 0.5 * w
+        new_coords[:, 1] += 0.5 * h
+        return new_coords
+
+    def apply_segmentation(self, segmentation):
+        segmentation = self.apply_image(segmentation, interp=Image.NEAREST)
+        return segmentation
+
+
+class ResizeTransform(Transform):
+    """
+    Resize the image to a target size.
+    """
+
+    def __init__(self, h, w, new_h, new_w, interp=None):
+        """
+        Args:
+            h, w (int): original image size
+            new_h, new_w (int): new image size
+            interp: PIL interpolation methods, defaults to bilinear.
+        """
+        # TODO decide on PIL vs opencv
+        super().__init__()
+        if interp is None:
+            interp = Image.BILINEAR
+        self._set_attributes(locals())
+
+    def apply_image(self, img, interp=None):
+        assert img.shape[:2] == (self.h, self.w)
+        assert len(img.shape) <= 4
+
+        if img.dtype == np.uint8:
+            pil_image = Image.fromarray(img)
+            interp_method = interp if interp is not None else self.interp
+            pil_image = pil_image.resize((self.new_w, self.new_h), interp_method)
+            ret = np.asarray(pil_image)
+        else:
+            # PIL only supports uint8
+            img = torch.from_numpy(img)
+            shape = list(img.shape)
+            shape_4d = shape[:2] + [1] * (4 - len(shape)) + shape[2:]
+            img = img.view(shape_4d).permute(2, 3, 0, 1)  # hw(c) -> nchw
+            _PIL_RESIZE_TO_INTERPOLATE_MODE = {Image.BILINEAR: "bilinear", Image.BICUBIC: "bicubic"}
+            mode = _PIL_RESIZE_TO_INTERPOLATE_MODE[self.interp]
+            img = F.interpolate(img, (self.new_h, self.new_w), mode=mode, align_corners=False)
+            shape[:2] = (self.new_h, self.new_w)
+            ret = img.permute(2, 3, 0, 1).view(shape).numpy()  # nchw -> hw(c)
+
+        return ret
+
+    def apply_coords(self, coords):
+        coords[:, 0] = coords[:, 0] * (self.new_w * 1.0 / self.w)
+        coords[:, 1] = coords[:, 1] * (self.new_h * 1.0 / self.h)
+        return coords
+
+    def apply_segmentation(self, segmentation):
+        segmentation = self.apply_image(segmentation, interp=Image.NEAREST)
+        return segmentation
+
+    def inverse(self):
+        return ResizeTransform(self.new_h, self.new_w, self.h, self.w, self.interp)
+
+
+class RotationTransform(Transform):
+    """
+    This method returns a copy of this image, rotated the given
+    number of degrees counter clockwise around its center.
+    """
+
+    def __init__(self, h, w, angle, expand=True, center=None, interp=None):
+        """
+        Args:
+            h, w (int): original image size
+            angle (float): degrees for rotation
+            expand (bool): choose if the image should be resized to fit the whole
+                rotated image (default), or simply cropped
+            center (tuple (width, height)): coordinates of the rotation center
+                if left to None, the center will be fit to the center of each image
+                center has no effect if expand=True because it only affects shifting
+            interp: cv2 interpolation method, default cv2.INTER_LINEAR
+        """
+        super().__init__()
+        image_center = np.array((w / 2, h / 2))
+        if center is None:
+            center = image_center
+        if interp is None:
+            interp = cv2.INTER_LINEAR
+        abs_cos, abs_sin = abs(np.cos(np.deg2rad(angle))), abs(np.sin(np.deg2rad(angle)))
+        if expand:
+            # find the new width and height bounds
+            bound_w, bound_h = np.rint(
+                [h * abs_sin + w * abs_cos, h * abs_cos + w * abs_sin]
+            ).astype(int)
+        else:
+            bound_w, bound_h = w, h
+
+        self._set_attributes(locals())
+        self.rm_coords = self.create_rotation_matrix()
+        # Needed because of this problem https://github.com/opencv/opencv/issues/11784
+        self.rm_image = self.create_rotation_matrix(offset=-0.5)
+
+    def apply_image(self, img, interp=None):
+        """
+        demo should be a numpy array, formatted as Height * Width * Nchannels
+        """
+        if len(img) == 0 or self.angle % 360 == 0:
+            return img
+        assert img.shape[:2] == (self.h, self.w)
+        interp = interp if interp is not None else self.interp
+        return cv2.warpAffine(img, self.rm_image, (self.bound_w, self.bound_h), flags=interp)
+
+    def apply_coords(self, coords):
+        """
+        coords should be a N * 2 array-like, containing N couples of (x, y) points
+        """
+        coords = np.asarray(coords, dtype=float)
+        if len(coords) == 0 or self.angle % 360 == 0:
+            return coords
+        return cv2.transform(coords[:, np.newaxis, :], self.rm_coords)[:, 0, :]
+
+    def apply_segmentation(self, segmentation):
+        segmentation = self.apply_image(segmentation, interp=cv2.INTER_NEAREST)
+        return segmentation
+
+    def create_rotation_matrix(self, offset=0):
+        center = (self.center[0] + offset, self.center[1] + offset)
+        rm = cv2.getRotationMatrix2D(tuple(center), self.angle, 1)
+        if self.expand:
+            # Find the coordinates of the center of rotation in the new image
+            # The only point for which we know the future coordinates is the center of the image
+            rot_im_center = cv2.transform(self.image_center[None, None, :] + offset, rm)[0, 0, :]
+            new_center = np.array([self.bound_w / 2, self.bound_h / 2]) + offset - rot_im_center
+            # shift the rotation center to the new coordinates
+            rm[:, 2] += new_center
+        return rm
+
+
+def HFlip_rotated_box(transform, rotated_boxes):
+    """
+    Apply the horizontal flip transform on rotated boxes.
+
+    Args:
+        rotated_boxes (ndarray): Nx5 floating point array of
+            (x_center, y_center, width, height, angle_degrees) format
+            in absolute coordinates.
+    """
+    # Transform x_center
+    rotated_boxes[:, 0] = transform.width - rotated_boxes[:, 0]
+    # Transform angle
+    rotated_boxes[:, 4] = -rotated_boxes[:, 4]
+    return rotated_boxes
+
+
+def Resize_rotated_box(transform, rotated_boxes):
+    """
+    Apply the resizing transform on rotated boxes. For details of how these (approximation)
+    formulas are derived, please refer to :meth:`RotatedBoxes.scale`.
+
+    Args:
+        rotated_boxes (ndarray): Nx5 floating point array of
+            (x_center, y_center, width, height, angle_degrees) format
+            in absolute coordinates.
+    """
+    scale_factor_x = transform.new_w * 1.0 / transform.w
+    scale_factor_y = transform.new_h * 1.0 / transform.h
+    rotated_boxes[:, 0] *= scale_factor_x
+    rotated_boxes[:, 1] *= scale_factor_y
+    theta = rotated_boxes[:, 4] * np.pi / 180.0
+    c = np.cos(theta)
+    s = np.sin(theta)
+    rotated_boxes[:, 2] *= np.sqrt(np.square(scale_factor_x * c) + np.square(scale_factor_y * s))
+    rotated_boxes[:, 3] *= np.sqrt(np.square(scale_factor_x * s) + np.square(scale_factor_y * c))
+    rotated_boxes[:, 4] = np.arctan2(scale_factor_x * s, scale_factor_y * c) * 180 / np.pi
+
+    return rotated_boxes
+
+
+HFlipTransform.register_type("rotated_box", HFlip_rotated_box)
+NoOpTransform.register_type("rotated_box", lambda t, x: x)
+ResizeTransform.register_type("rotated_box", Resize_rotated_box)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/data/transforms/transform_gen.py

@@ -0,0 +1,534 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+# File: transformer.py
+
+import inspect
+import numpy as np
+import pprint
+import sys
+from abc import ABCMeta, abstractmethod
+from fvcore.transforms.transform import (
+    BlendTransform,
+    CropTransform,
+    HFlipTransform,
+    NoOpTransform,
+    Transform,
+    TransformList,
+    VFlipTransform,
+)
+from PIL import Image
+
+from .transform import ExtentTransform, ResizeTransform, RotationTransform
+
+__all__ = [
+    "RandomApply",
+    "RandomBrightness",
+    "RandomContrast",
+    "RandomCrop",
+    "RandomExtent",
+    "RandomFlip",
+    "RandomSaturation",
+    "RandomLighting",
+    "RandomRotation",
+    "Resize",
+    "ResizeShortestEdge",
+    "TransformGen",
+    "apply_transform_gens",
+]
+
+
+def check_dtype(img):
+    assert isinstance(img, np.ndarray), "[TransformGen] Needs an numpy array, but got a {}!".format(
+        type(img)
+    )
+    assert not isinstance(img.dtype, np.integer) or (
+        img.dtype == np.uint8
+    ), "[TransformGen] Got image of type {}, use uint8 or floating points instead!".format(
+        img.dtype
+    )
+    assert img.ndim in [2, 3], img.ndim
+
+
+class TransformGen(metaclass=ABCMeta):
+    """
+    TransformGen takes an image of type uint8 in range [0, 255], or
+    floating point in range [0, 1] or [0, 255] as input.
+
+    It creates a :class:`Transform` based on the given image, sometimes with randomness.
+    The transform can then be used to transform images
+    or other data (boxes, points, annotations, etc.) associated with it.
+
+    The assumption made in this class
+    is that the image itself is sufficient to instantiate a transform.
+    When this assumption is not true, you need to create the transforms by your own.
+
+    A list of `TransformGen` can be applied with :func:`apply_transform_gens`.
+    """
+
+    def _init(self, params=None):
+        if params:
+            for k, v in params.items():
+                if k != "self" and not k.startswith("_"):
+                    setattr(self, k, v)
+
+    @abstractmethod
+    def get_transform(self, img):
+        pass
+
+    def _rand_range(self, low=1.0, high=None, size=None):
+        """
+        Uniform float random number between low and high.
+        """
+        if high is None:
+            low, high = 0, low
+        if size is None:
+            size = []
+        return np.random.uniform(low, high, size)
+
+    def __repr__(self):
+        """
+        Produce something like:
+        "MyTransformGen(field1={self.field1}, field2={self.field2})"
+        """
+        try:
+            sig = inspect.signature(self.__init__)
+            classname = type(self).__name__
+            argstr = []
+            for name, param in sig.parameters.items():
+                assert (
+                    param.kind != param.VAR_POSITIONAL and param.kind != param.VAR_KEYWORD
+                ), "The default __repr__ doesn't support *args or **kwargs"
+                assert hasattr(self, name), (
+                    "Attribute {} not found! "
+                    "Default __repr__ only works if attributes match the constructor.".format(name)
+                )
+                attr = getattr(self, name)
+                default = param.default
+                if default is attr:
+                    continue
+                argstr.append("{}={}".format(name, pprint.pformat(attr)))
+            return "{}({})".format(classname, ", ".join(argstr))
+        except AssertionError:
+            return super().__repr__()
+
+    __str__ = __repr__
+
+
+class RandomApply(TransformGen):
+    """
+    Randomly apply the wrapper transformation with a given probability.
+    """
+
+    def __init__(self, transform, prob=0.5):
+        """
+        Args:
+            transform (Transform, TransformGen): the transform to be wrapped
+                by the `RandomApply`. The `transform` can either be a
+                `Transform` or `TransformGen` instance.
+            prob (float): probability between 0.0 and 1.0 that
+                the wrapper transformation is applied
+        """
+        super().__init__()
+        assert isinstance(transform, (Transform, TransformGen)), (
+            f"The given transform must either be a Transform or TransformGen instance. "
+            f"Not {type(transform)}"
+        )
+        assert 0.0 <= prob <= 1.0, f"Probablity must be between 0.0 and 1.0 (given: {prob})"
+        self.prob = prob
+        self.transform = transform
+
+    def get_transform(self, img):
+        do = self._rand_range() < self.prob
+        if do:
+            if isinstance(self.transform, TransformGen):
+                return self.transform.get_transform(img)
+            else:
+                return self.transform
+        else:
+            return NoOpTransform()
+
+
+class RandomFlip(TransformGen):
+    """
+    Flip the image horizontally or vertically with the given probability.
+    """
+
+    def __init__(self, prob=0.5, *, horizontal=True, vertical=False):
+        """
+        Args:
+            prob (float): probability of flip.
+            horizontal (boolean): whether to apply horizontal flipping
+            vertical (boolean): whether to apply vertical flipping
+        """
+        super().__init__()
+
+        if horizontal and vertical:
+            raise ValueError("Cannot do both horiz and vert. Please use two Flip instead.")
+        if not horizontal and not vertical:
+            raise ValueError("At least one of horiz or vert has to be True!")
+        self._init(locals())
+
+    def get_transform(self, img):
+        h, w = img.shape[:2]
+        do = self._rand_range() < self.prob
+        if do:
+            if self.horizontal:
+                return HFlipTransform(w)
+            elif self.vertical:
+                return VFlipTransform(h)
+        else:
+            return NoOpTransform()
+
+
+class Resize(TransformGen):
+    """ Resize image to a target size"""
+
+    def __init__(self, shape, interp=Image.BILINEAR):
+        """
+        Args:
+            shape: (h, w) tuple or a int
+            interp: PIL interpolation method
+        """
+        if isinstance(shape, int):
+            shape = (shape, shape)
+        shape = tuple(shape)
+        self._init(locals())
+
+    def get_transform(self, img):
+        return ResizeTransform(
+            img.shape[0], img.shape[1], self.shape[0], self.shape[1], self.interp
+        )
+
+
+class ResizeShortestEdge(TransformGen):
+    """
+    Scale the shorter edge to the given size, with a limit of `max_size` on the longer edge.
+    If `max_size` is reached, then downscale so that the longer edge does not exceed max_size.
+    """
+
+    def __init__(
+        self, short_edge_length, max_size=sys.maxsize, sample_style="range", interp=Image.BILINEAR
+    ):
+        """
+        Args:
+            short_edge_length (list[int]): If ``sample_style=="range"``,
+                a [min, max] interval from which to sample the shortest edge length.
+                If ``sample_style=="choice"``, a list of shortest edge lengths to sample from.
+            max_size (int): maximum allowed longest edge length.
+            sample_style (str): either "range" or "choice".
+        """
+        super().__init__()
+        assert sample_style in ["range", "choice"], sample_style
+
+        self.is_range = sample_style == "range"
+        if isinstance(short_edge_length, int):
+            short_edge_length = (short_edge_length, short_edge_length)
+        self._init(locals())
+
+    def get_transform(self, img):
+        h, w = img.shape[:2]
+
+        if self.is_range:
+            size = np.random.randint(self.short_edge_length[0], self.short_edge_length[1] + 1)
+        else:
+            size = np.random.choice(self.short_edge_length)
+        if size == 0:
+            return NoOpTransform()
+
+        scale = size * 1.0 / min(h, w)
+        if h < w:
+            newh, neww = size, scale * w
+        else:
+            newh, neww = scale * h, size
+        if max(newh, neww) > self.max_size:
+            scale = self.max_size * 1.0 / max(newh, neww)
+            newh = newh * scale
+            neww = neww * scale
+        neww = int(neww + 0.5)
+        newh = int(newh + 0.5)
+        return ResizeTransform(h, w, newh, neww, self.interp)
+
+
+class RandomRotation(TransformGen):
+    """
+    This method returns a copy of this image, rotated the given
+    number of degrees counter clockwise around the given center.
+    """
+
+    def __init__(self, angle, expand=True, center=None, sample_style="range", interp=None):
+        """
+        Args:
+            angle (list[float]): If ``sample_style=="range"``,
+                a [min, max] interval from which to sample the angle (in degrees).
+                If ``sample_style=="choice"``, a list of angles to sample from
+            expand (bool): choose if the image should be resized to fit the whole
+                rotated image (default), or simply cropped
+            center (list[[float, float]]):  If ``sample_style=="range"``,
+                a [[minx, miny], [maxx, maxy]] relative interval from which to sample the center,
+                [0, 0] being the top left of the image and [1, 1] the bottom right.
+                If ``sample_style=="choice"``, a list of centers to sample from
+                Default: None, which means that the center of rotation is the center of the image
+                center has no effect if expand=True because it only affects shifting
+        """
+        super().__init__()
+        assert sample_style in ["range", "choice"], sample_style
+        self.is_range = sample_style == "range"
+        if isinstance(angle, (float, int)):
+            angle = (angle, angle)
+        if center is not None and isinstance(center[0], (float, int)):
+            center = (center, center)
+        self._init(locals())
+
+    def get_transform(self, img):
+        h, w = img.shape[:2]
+        center = None
+        if self.is_range:
+            angle = np.random.uniform(self.angle[0], self.angle[1])
+            if self.center is not None:
+                center = (
+                    np.random.uniform(self.center[0][0], self.center[1][0]),
+                    np.random.uniform(self.center[0][1], self.center[1][1]),
+                )
+        else:
+            angle = np.random.choice(self.angle)
+            if self.center is not None:
+                center = np.random.choice(self.center)
+
+        if center is not None:
+            center = (w * center[0], h * center[1])  # Convert to absolute coordinates
+
+        return RotationTransform(h, w, angle, expand=self.expand, center=center, interp=self.interp)
+
+
+class RandomCrop(TransformGen):
+    """
+    Randomly crop a subimage out of an image.
+    """
+
+    def __init__(self, crop_type: str, crop_size):
+        """
+        Args:
+            crop_type (str): one of "relative_range", "relative", "absolute".
+                See `config/defaults.py` for explanation.
+            crop_size (tuple[float]): the relative ratio or absolute pixels of
+                height and width
+        """
+        super().__init__()
+        assert crop_type in ["relative_range", "relative", "absolute"]
+        self._init(locals())
+
+    def get_transform(self, img):
+        h, w = img.shape[:2]
+        croph, cropw = self.get_crop_size((h, w))
+        assert h >= croph and w >= cropw, "Shape computation in {} has bugs.".format(self)
+        h0 = np.random.randint(h - croph + 1)
+        w0 = np.random.randint(w - cropw + 1)
+        return CropTransform(w0, h0, cropw, croph)
+
+    def get_crop_size(self, image_size):
+        """
+        Args:
+            image_size (tuple): height, width
+
+        Returns:
+            crop_size (tuple): height, width in absolute pixels
+        """
+        h, w = image_size
+        if self.crop_type == "relative":
+            ch, cw = self.crop_size
+            return int(h * ch + 0.5), int(w * cw + 0.5)
+        elif self.crop_type == "relative_range":
+            crop_size = np.asarray(self.crop_size, dtype=np.float32)
+            ch, cw = crop_size + np.random.rand(2) * (1 - crop_size)
+            return int(h * ch + 0.5), int(w * cw + 0.5)
+        elif self.crop_type == "absolute":
+            return (min(self.crop_size[0], h), min(self.crop_size[1], w))
+        else:
+            NotImplementedError("Unknown crop type {}".format(self.crop_type))
+
+
+class RandomExtent(TransformGen):
+    """
+    Outputs an image by cropping a random "subrect" of the source image.
+
+    The subrect can be parameterized to include pixels outside the source image,
+    in which case they will be set to zeros (i.e. black). The size of the output
+    image will vary with the size of the random subrect.
+    """
+
+    def __init__(self, scale_range, shift_range):
+        """
+        Args:
+            output_size (h, w): Dimensions of output image
+            scale_range (l, h): Range of input-to-output size scaling factor
+            shift_range (x, y): Range of shifts of the cropped subrect. The rect
+                is shifted by [w / 2 * Uniform(-x, x), h / 2 * Uniform(-y, y)],
+                where (w, h) is the (width, height) of the input image. Set each
+                component to zero to crop at the image's center.
+        """
+        super().__init__()
+        self._init(locals())
+
+    def get_transform(self, img):
+        img_h, img_w = img.shape[:2]
+
+        # Initialize src_rect to fit the input image.
+        src_rect = np.array([-0.5 * img_w, -0.5 * img_h, 0.5 * img_w, 0.5 * img_h])
+
+        # Apply a random scaling to the src_rect.
+        src_rect *= np.random.uniform(self.scale_range[0], self.scale_range[1])
+
+        # Apply a random shift to the coordinates origin.
+        src_rect[0::2] += self.shift_range[0] * img_w * (np.random.rand() - 0.5)
+        src_rect[1::2] += self.shift_range[1] * img_h * (np.random.rand() - 0.5)
+
+        # Map src_rect coordinates into image coordinates (center at corner).
+        src_rect[0::2] += 0.5 * img_w
+        src_rect[1::2] += 0.5 * img_h
+
+        return ExtentTransform(
+            src_rect=(src_rect[0], src_rect[1], src_rect[2], src_rect[3]),
+            output_size=(int(src_rect[3] - src_rect[1]), int(src_rect[2] - src_rect[0])),
+        )
+
+
+class RandomContrast(TransformGen):
+    """
+    Randomly transforms image contrast.
+
+    Contrast intensity is uniformly sampled in (intensity_min, intensity_max).
+    - intensity < 1 will reduce contrast
+    - intensity = 1 will preserve the input image
+    - intensity > 1 will increase contrast
+
+    See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html
+    """
+
+    def __init__(self, intensity_min, intensity_max):
+        """
+        Args:
+            intensity_min (float): Minimum augmentation
+            intensity_max (float): Maximum augmentation
+        """
+        super().__init__()
+        self._init(locals())
+
+    def get_transform(self, img):
+        w = np.random.uniform(self.intensity_min, self.intensity_max)
+        return BlendTransform(src_image=img.mean(), src_weight=1 - w, dst_weight=w)
+
+
+class RandomBrightness(TransformGen):
+    """
+    Randomly transforms image brightness.
+
+    Brightness intensity is uniformly sampled in (intensity_min, intensity_max).
+    - intensity < 1 will reduce brightness
+    - intensity = 1 will preserve the input image
+    - intensity > 1 will increase brightness
+
+    See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html
+    """
+
+    def __init__(self, intensity_min, intensity_max):
+        """
+        Args:
+            intensity_min (float): Minimum augmentation
+            intensity_max (float): Maximum augmentation
+        """
+        super().__init__()
+        self._init(locals())
+
+    def get_transform(self, img):
+        w = np.random.uniform(self.intensity_min, self.intensity_max)
+        return BlendTransform(src_image=0, src_weight=1 - w, dst_weight=w)
+
+
+class RandomSaturation(TransformGen):
+    """
+    Randomly transforms image saturation.
+
+    Saturation intensity is uniformly sampled in (intensity_min, intensity_max).
+    - intensity < 1 will reduce saturation (make the image more grayscale)
+    - intensity = 1 will preserve the input image
+    - intensity > 1 will increase saturation
+
+    See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html
+    """
+
+    def __init__(self, intensity_min, intensity_max):
+        """
+        Args:
+            intensity_min (float): Minimum augmentation (1 preserves input).
+            intensity_max (float): Maximum augmentation (1 preserves input).
+        """
+        super().__init__()
+        self._init(locals())
+
+    def get_transform(self, img):
+        assert img.shape[-1] == 3, "Saturation only works on RGB images"
+        w = np.random.uniform(self.intensity_min, self.intensity_max)
+        grayscale = img.dot([0.299, 0.587, 0.114])[:, :, np.newaxis]
+        return BlendTransform(src_image=grayscale, src_weight=1 - w, dst_weight=w)
+
+
+class RandomLighting(TransformGen):
+    """
+    Randomly transforms image color using fixed PCA over ImageNet.
+
+    The degree of color jittering is randomly sampled via a normal distribution,
+    with standard deviation given by the scale parameter.
+    """
+
+    def __init__(self, scale):
+        """
+        Args:
+            scale (float): Standard deviation of principal component weighting.
+        """
+        super().__init__()
+        self._init(locals())
+        self.eigen_vecs = np.array(
+            [[-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203]]
+        )
+        self.eigen_vals = np.array([0.2175, 0.0188, 0.0045])
+
+    def get_transform(self, img):
+        assert img.shape[-1] == 3, "Saturation only works on RGB images"
+        weights = np.random.normal(scale=self.scale, size=3)
+        return BlendTransform(
+            src_image=self.eigen_vecs.dot(weights * self.eigen_vals), src_weight=1.0, dst_weight=1.0
+        )
+
+
+def apply_transform_gens(transform_gens, img):
+    """
+    Apply a list of :class:`TransformGen` or :class:`Transform` on the input image, and
+    returns the transformed image and a list of transforms.
+
+    We cannot simply create and return all transforms without
+    applying it to the image, because a subsequent transform may
+    need the output of the previous one.
+
+    Args:
+        transform_gens (list): list of :class:`TransformGen` or :class:`Transform` instance to
+            be applied.
+        img (ndarray): uint8 or floating point images with 1 or 3 channels.
+
+    Returns:
+        ndarray: the transformed image
+        TransformList: contain the transforms that's used.
+    """
+    for g in transform_gens:
+        assert isinstance(g, (Transform, TransformGen)), g
+
+    check_dtype(img)
+
+    tfms = []
+    for g in transform_gens:
+        tfm = g.get_transform(img) if isinstance(g, TransformGen) else g
+        assert isinstance(
+            tfm, Transform
+        ), "TransformGen {} must return an instance of Transform! Got {} instead".format(g, tfm)
+        img = tfm.apply_image(img)
+        tfms.append(tfm)
+    return img, TransformList(tfms)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .cityscapes_evaluation import CityscapesInstanceEvaluator, CityscapesSemSegEvaluator
+from .coco_evaluation import COCOEvaluator
+from .rotated_coco_evaluation import RotatedCOCOEvaluator
+from .evaluator import DatasetEvaluator, DatasetEvaluators, inference_context, inference_on_dataset
+from .lvis_evaluation import LVISEvaluator
+from .panoptic_evaluation import COCOPanopticEvaluator
+from .pascal_voc_evaluation import PascalVOCDetectionEvaluator
+from .sem_seg_evaluation import SemSegEvaluator
+from .testing import print_csv_format, verify_results
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/cityscapes_evaluation.py

@@ -0,0 +1,187 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import glob
+import logging
+import numpy as np
+import os
+import tempfile
+from collections import OrderedDict
+import torch
+from fvcore.common.file_io import PathManager
+from PIL import Image
+
+from detectron2.data import MetadataCatalog
+from detectron2.utils import comm
+
+from .evaluator import DatasetEvaluator
+
+
+class CityscapesEvaluator(DatasetEvaluator):
+    """
+    Base class for evaluation using cityscapes API.
+    """
+
+    def __init__(self, dataset_name):
+        """
+        Args:
+            dataset_name (str): the name of the dataset.
+                It must have the following metadata associated with it:
+                "thing_classes", "gt_dir".
+        """
+        self._metadata = MetadataCatalog.get(dataset_name)
+        self._cpu_device = torch.device("cpu")
+        self._logger = logging.getLogger(__name__)
+
+    def reset(self):
+        self._working_dir = tempfile.TemporaryDirectory(prefix="cityscapes_eval_")
+        self._temp_dir = self._working_dir.name
+        # All workers will write to the same results directory
+        # TODO this does not work in distributed training
+        self._temp_dir = comm.all_gather(self._temp_dir)[0]
+        if self._temp_dir != self._working_dir.name:
+            self._working_dir.cleanup()
+        self._logger.info(
+            "Writing cityscapes results to temporary directory {} ...".format(self._temp_dir)
+        )
+
+
+class CityscapesInstanceEvaluator(CityscapesEvaluator):
+    """
+    Evaluate instance segmentation results using cityscapes API.
+
+    Note:
+        * It does not work in multi-machine distributed training.
+        * It contains a synchronization, therefore has to be used on all ranks.
+        * Only the main process runs evaluation.
+    """
+
+    def process(self, inputs, outputs):
+        from cityscapesscripts.helpers.labels import name2label
+
+        for input, output in zip(inputs, outputs):
+            file_name = input["file_name"]
+            basename = os.path.splitext(os.path.basename(file_name))[0]
+            pred_txt = os.path.join(self._temp_dir, basename + "_pred.txt")
+
+            output = output["instances"].to(self._cpu_device)
+            num_instances = len(output)
+            with open(pred_txt, "w") as fout:
+                for i in range(num_instances):
+                    pred_class = output.pred_classes[i]
+                    classes = self._metadata.thing_classes[pred_class]
+                    class_id = name2label[classes].id
+                    score = output.scores[i]
+                    mask = output.pred_masks[i].numpy().astype("uint8")
+                    png_filename = os.path.join(
+                        self._temp_dir, basename + "_{}_{}.png".format(i, classes)
+                    )
+
+                    Image.fromarray(mask * 255).save(png_filename)
+                    fout.write("{} {} {}\n".format(os.path.basename(png_filename), class_id, score))
+
+    def evaluate(self):
+        """
+        Returns:
+            dict: has a key "segm", whose value is a dict of "AP" and "AP50".
+        """
+        comm.synchronize()
+        if comm.get_rank() > 0:
+            return
+        import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as cityscapes_eval
+
+        self._logger.info("Evaluating results under {} ...".format(self._temp_dir))
+
+        # set some global states in cityscapes evaluation API, before evaluating
+        cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir)
+        cityscapes_eval.args.predictionWalk = None
+        cityscapes_eval.args.JSONOutput = False
+        cityscapes_eval.args.colorized = False
+        cityscapes_eval.args.gtInstancesFile = os.path.join(self._temp_dir, "gtInstances.json")
+
+        # These lines are adopted from
+        # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalInstanceLevelSemanticLabeling.py # noqa
+        gt_dir = PathManager.get_local_path(self._metadata.gt_dir)
+        groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_instanceIds.png"))
+        assert len(
+            groundTruthImgList
+        ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format(
+            cityscapes_eval.args.groundTruthSearch
+        )
+        predictionImgList = []
+        for gt in groundTruthImgList:
+            predictionImgList.append(cityscapes_eval.getPrediction(gt, cityscapes_eval.args))
+        results = cityscapes_eval.evaluateImgLists(
+            predictionImgList, groundTruthImgList, cityscapes_eval.args
+        )["averages"]
+
+        ret = OrderedDict()
+        ret["segm"] = {"AP": results["allAp"] * 100, "AP50": results["allAp50%"] * 100}
+        self._working_dir.cleanup()
+        return ret
+
+
+class CityscapesSemSegEvaluator(CityscapesEvaluator):
+    """
+    Evaluate semantic segmentation results using cityscapes API.
+
+    Note:
+        * It does not work in multi-machine distributed training.
+        * It contains a synchronization, therefore has to be used on all ranks.
+        * Only the main process runs evaluation.
+    """
+
+    def process(self, inputs, outputs):
+        from cityscapesscripts.helpers.labels import trainId2label
+
+        for input, output in zip(inputs, outputs):
+            file_name = input["file_name"]
+            basename = os.path.splitext(os.path.basename(file_name))[0]
+            pred_filename = os.path.join(self._temp_dir, basename + "_pred.png")
+
+            output = output["sem_seg"].argmax(dim=0).to(self._cpu_device).numpy()
+            pred = 255 * np.ones(output.shape, dtype=np.uint8)
+            for train_id, label in trainId2label.items():
+                if label.ignoreInEval:
+                    continue
+                pred[output == train_id] = label.id
+            Image.fromarray(pred).save(pred_filename)
+
+    def evaluate(self):
+        comm.synchronize()
+        if comm.get_rank() > 0:
+            return
+        # Load the Cityscapes eval script *after* setting the required env var,
+        # since the script reads CITYSCAPES_DATASET into global variables at load time.
+        import cityscapesscripts.evaluation.evalPixelLevelSemanticLabeling as cityscapes_eval
+
+        self._logger.info("Evaluating results under {} ...".format(self._temp_dir))
+
+        # set some global states in cityscapes evaluation API, before evaluating
+        cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir)
+        cityscapes_eval.args.predictionWalk = None
+        cityscapes_eval.args.JSONOutput = False
+        cityscapes_eval.args.colorized = False
+
+        # These lines are adopted from
+        # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalPixelLevelSemanticLabeling.py # noqa
+        gt_dir = PathManager.get_local_path(self._metadata.gt_dir)
+        groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_labelIds.png"))
+        assert len(
+            groundTruthImgList
+        ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format(
+            cityscapes_eval.args.groundTruthSearch
+        )
+        predictionImgList = []
+        for gt in groundTruthImgList:
+            predictionImgList.append(cityscapes_eval.getPrediction(cityscapes_eval.args, gt))
+        results = cityscapes_eval.evaluateImgLists(
+            predictionImgList, groundTruthImgList, cityscapes_eval.args
+        )
+        ret = OrderedDict()
+        ret["sem_seg"] = {
+            "IoU": 100.0 * results["averageScoreClasses"],
+            "iIoU": 100.0 * results["averageScoreInstClasses"],
+            "IoU_sup": 100.0 * results["averageScoreCategories"],
+            "iIoU_sup": 100.0 * results["averageScoreInstCategories"],
+        }
+        self._working_dir.cleanup()
+        return ret

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/coco_evaluation.py

@@ -0,0 +1,512 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import contextlib
+import copy
+import io
+import itertools
+import json
+import logging
+import numpy as np
+import os
+import pickle
+from collections import OrderedDict
+import pycocotools.mask as mask_util
+import torch
+from fvcore.common.file_io import PathManager
+from pycocotools.coco import COCO
+from pycocotools.cocoeval import COCOeval
+from tabulate import tabulate
+
+import detectron2.utils.comm as comm
+from detectron2.data import MetadataCatalog
+from detectron2.data.datasets.coco import convert_to_coco_json
+from detectron2.structures import Boxes, BoxMode, pairwise_iou
+from detectron2.utils.logger import create_small_table
+
+from .evaluator import DatasetEvaluator
+
+
+class COCOEvaluator(DatasetEvaluator):
+    """
+    Evaluate object proposal, instance detection/segmentation, keypoint detection
+    outputs using COCO's metrics and APIs.
+    """
+
+    def __init__(self, dataset_name, cfg, distributed, output_dir=None):
+        """
+        Args:
+            dataset_name (str): name of the dataset to be evaluated.
+                It must have either the following corresponding metadata:
+
+                    "json_file": the path to the COCO format annotation
+
+                Or it must be in detectron2's standard dataset format
+                so it can be converted to COCO format automatically.
+            cfg (CfgNode): config instance
+            distributed (True): if True, will collect results from all ranks and run evaluation
+                in the main process.
+                Otherwise, will evaluate the results in the current process.
+            output_dir (str): optional, an output directory to dump all
+                results predicted on the dataset. The dump contains two files:
+
+                1. "instance_predictions.pth" a file in torch serialization
+                   format that contains all the raw original predictions.
+                2. "coco_instances_results.json" a json file in COCO's result
+                   format.
+        """
+        self._tasks = self._tasks_from_config(cfg)
+        self._distributed = distributed
+        self._output_dir = output_dir
+
+        self._cpu_device = torch.device("cpu")
+        self._logger = logging.getLogger(__name__)
+
+        self._metadata = MetadataCatalog.get(dataset_name)
+        if not hasattr(self._metadata, "json_file"):
+            self._logger.warning(
+                f"json_file was not found in MetaDataCatalog for '{dataset_name}'."
+                " Trying to convert it to COCO format ..."
+            )
+
+            cache_path = os.path.join(output_dir, f"{dataset_name}_coco_format.json")
+            self._metadata.json_file = cache_path
+            convert_to_coco_json(dataset_name, cache_path)
+
+        json_file = PathManager.get_local_path(self._metadata.json_file)
+        with contextlib.redirect_stdout(io.StringIO()):
+            self._coco_api = COCO(json_file)
+
+        self._kpt_oks_sigmas = cfg.TEST.KEYPOINT_OKS_SIGMAS
+        # Test set json files do not contain annotations (evaluation must be
+        # performed using the COCO evaluation server).
+        self._do_evaluation = "annotations" in self._coco_api.split_name
+
+    def reset(self):
+        self._predictions = []
+
+    def _tasks_from_config(self, cfg):
+        """
+        Returns:
+            tuple[str]: tasks that can be evaluated under the given configuration.
+        """
+        tasks = ("bbox",)
+        if cfg.MODEL.MASK_ON:
+            tasks = tasks + ("segm",)
+        if cfg.MODEL.KEYPOINT_ON:
+            tasks = tasks + ("keypoints",)
+        return tasks
+
+    def process(self, inputs, outputs):
+        """
+        Args:
+            inputs: the inputs to a COCO model (e.g., GeneralizedRCNN).
+                It is a list of dict. Each dict corresponds to an image and
+                contains keys like "height", "width", "file_name", "image_id".
+            outputs: the outputs of a COCO model. It is a list of dicts with key
+                "instances" that contains :class:`Instances`.
+        """
+        for input, output in zip(inputs, outputs):
+            prediction = {"image_id": input["image_id"]}
+
+            # TODO this is ugly
+            if "instances" in output:
+                instances = output["instances"].to(self._cpu_device)
+                prediction["instances"] = instances_to_coco_json(instances, input["image_id"])
+            if "proposals" in output:
+                prediction["proposals"] = output["proposals"].to(self._cpu_device)
+            self._predictions.append(prediction)
+
+    def evaluate(self):
+        if self._distributed:
+            comm.synchronize()
+            predictions = comm.gather(self._predictions, dst=0)
+            predictions = list(itertools.chain(*predictions))
+
+            if not comm.is_main_process():
+                return {}
+        else:
+            predictions = self._predictions
+
+        if len(predictions) == 0:
+            self._logger.warning("[COCOEvaluator] Did not receive valid predictions.")
+            return {}
+
+        if self._output_dir:
+            PathManager.mkdirs(self._output_dir)
+            file_path = os.path.join(self._output_dir, "instances_predictions.pth")
+            with PathManager.open(file_path, "wb") as f:
+                torch.save(predictions, f)
+
+        self._results = OrderedDict()
+        if "proposals" in predictions[0]:
+            self._eval_box_proposals(predictions)
+        if "instances" in predictions[0]:
+            self._eval_predictions(set(self._tasks), predictions)
+        # Copy so the caller can do whatever with results
+        return copy.deepcopy(self._results)
+
+    def _eval_predictions(self, tasks, predictions):
+        """
+        Evaluate predictions on the given tasks.
+        Fill self._results with the metrics of the tasks.
+        """
+        self._logger.info("Preparing results for COCO format ...")
+        coco_results = list(itertools.chain(*[x["instances"] for x in predictions]))
+
+        # unmap the category ids for COCO
+        if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
+            reverse_id_mapping = {
+                v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items()
+            }
+            for result in coco_results:
+                category_id = result["category_id"]
+                assert (
+                    category_id in reverse_id_mapping
+                ), "A prediction has category_id={}, which is not available in the dataset.".format(
+                    category_id
+                )
+                result["category_id"] = reverse_id_mapping[category_id]
+
+        if self._output_dir:
+            file_path = os.path.join(self._output_dir, "coco_instances_results.json")
+            self._logger.info("Saving results to {}".format(file_path))
+            with PathManager.open(file_path, "w") as f:
+                f.write(json.dumps(coco_results))
+                f.flush()
+
+        if not self._do_evaluation:
+            self._logger.info("Annotations are not available for evaluation.")
+            return
+
+        self._logger.info("Evaluating predictions ...")
+        for task in sorted(tasks):
+            coco_eval = (
+                _evaluate_predictions_on_coco(
+                    self._coco_api, coco_results, task, kpt_oks_sigmas=self._kpt_oks_sigmas
+                )
+                if len(coco_results) > 0
+                else None  # cocoapi does not handle empty results very well
+            )
+
+            res = self._derive_coco_results(
+                coco_eval, task, class_names=self._metadata.get("thing_classes")
+            )
+            self._results[task] = res
+
+    def _eval_box_proposals(self, predictions):
+        """
+        Evaluate the box proposals in predictions.
+        Fill self._results with the metrics for "box_proposals" task.
+        """
+        if self._output_dir:
+            # Saving generated box proposals to file.
+            # Predicted box_proposals are in XYXY_ABS mode.
+            bbox_mode = BoxMode.XYXY_ABS.value
+            ids, boxes, objectness_logits = [], [], []
+            for prediction in predictions:
+                ids.append(prediction["image_id"])
+                boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy())
+                objectness_logits.append(prediction["proposals"].objectness_logits.numpy())
+
+            proposal_data = {
+                "boxes": boxes,
+                "objectness_logits": objectness_logits,
+                "ids": ids,
+                "bbox_mode": bbox_mode,
+            }
+            with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f:
+                pickle.dump(proposal_data, f)
+
+        if not self._do_evaluation:
+            self._logger.info("Annotations are not available for evaluation.")
+            return
+
+        self._logger.info("Evaluating bbox proposals ...")
+        res = {}
+        areas = {"all": "", "small": "s", "medium": "m", "large": "l"}
+        for limit in [100, 1000]:
+            for area, suffix in areas.items():
+                stats = _evaluate_box_proposals(predictions, self._coco_api, area=area, limit=limit)
+                key = "AR{}@{:d}".format(suffix, limit)
+                res[key] = float(stats["ar"].item() * 100)
+        self._logger.info("Proposal metrics: \n" + create_small_table(res))
+        self._results["box_proposals"] = res
+
+    def _derive_coco_results(self, coco_eval, iou_type, class_names=None):
+        """
+        Derive the desired score numbers from summarized COCOeval.
+
+        Args:
+            coco_eval (None or COCOEval): None represents no predictions from model.
+            iou_type (str):
+            class_names (None or list[str]): if provided, will use it to predict
+                per-category AP.
+
+        Returns:
+            a dict of {metric name: score}
+        """
+
+        metrics = {
+            "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl"],
+            "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl"],
+            "keypoints": ["AP", "AP50", "AP75", "APm", "APl"],
+        }[iou_type]
+
+        if coco_eval is None:
+            self._logger.warn("No predictions from the model!")
+            return {metric: float("nan") for metric in metrics}
+
+        # the standard metrics
+        results = {
+            metric: float(coco_eval.stats[idx] * 100 if coco_eval.stats[idx] >= 0 else "nan")
+            for idx, metric in enumerate(metrics)
+        }
+        self._logger.info(
+            "Evaluation results for {}: \n".format(iou_type) + create_small_table(results)
+        )
+        if not np.isfinite(sum(results.values())):
+            self._logger.info("Note that some metrics cannot be computed.")
+
+        if class_names is None or len(class_names) <= 1:
+            return results
+        # Compute per-category AP
+        # from https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L222-L252 # noqa
+        precisions = coco_eval.eval["precision"]
+        # precision has dims (iou, recall, cls, area range, max dets)
+        assert len(class_names) == precisions.shape[2]
+
+        results_per_category = []
+        for idx, name in enumerate(class_names):
+            # area range index 0: all area ranges
+            # max dets index -1: typically 100 per image
+            precision = precisions[:, :, idx, 0, -1]
+            precision = precision[precision > -1]
+            ap = np.mean(precision) if precision.size else float("nan")
+            results_per_category.append(("{}".format(name), float(ap * 100)))
+
+        # tabulate it
+        N_COLS = min(6, len(results_per_category) * 2)
+        results_flatten = list(itertools.chain(*results_per_category))
+        results_2d = itertools.zip_longest(*[results_flatten[i::N_COLS] for i in range(N_COLS)])
+        table = tabulate(
+            results_2d,
+            tablefmt="pipe",
+            floatfmt=".3f",
+            headers=["category", "AP"] * (N_COLS // 2),
+            numalign="left",
+        )
+        self._logger.info("Per-category {} AP: \n".format(iou_type) + table)
+
+        results.update({"AP-" + name: ap for name, ap in results_per_category})
+        return results
+
+
+def instances_to_coco_json(instances, img_id):
+    """
+    Dump an "Instances" object to a COCO-format json that's used for evaluation.
+
+    Args:
+        instances (Instances):
+        img_id (int): the image id
+
+    Returns:
+        list[dict]: list of json annotations in COCO format.
+    """
+    num_instance = len(instances)
+    if num_instance == 0:
+        return []
+
+    boxes = instances.pred_boxes.tensor.numpy()
+    boxes = BoxMode.convert(boxes, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+    boxes = boxes.tolist()
+    scores = instances.scores.tolist()
+    classes = instances.pred_classes.tolist()
+
+    has_mask = instances.has("pred_masks")
+    if has_mask:
+        # use RLE to encode the masks, because they are too large and takes memory
+        # since this evaluator stores outputs of the entire dataset
+        rles = [
+            mask_util.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0]
+            for mask in instances.pred_masks
+        ]
+        for rle in rles:
+            # "counts" is an array encoded by mask_util as a byte-stream. Python3's
+            # json writer which always produces strings cannot serialize a bytestream
+            # unless you decode it. Thankfully, utf-8 works out (which is also what
+            # the pycocotools/_mask.pyx does).
+            rle["counts"] = rle["counts"].decode("utf-8")
+
+    has_keypoints = instances.has("pred_keypoints")
+    if has_keypoints:
+        keypoints = instances.pred_keypoints
+
+    results = []
+    for k in range(num_instance):
+        result = {
+            "image_id": img_id,
+            "category_id": classes[k],
+            "bbox": boxes[k],
+            "score": scores[k],
+        }
+        if has_mask:
+            result["segmentation"] = rles[k]
+        if has_keypoints:
+            # In COCO annotations,
+            # keypoints coordinates are pixel indices.
+            # However our predictions are floating point coordinates.
+            # Therefore we subtract 0.5 to be consistent with the annotation format.
+            # This is the inverse of data loading logic in `data/coco.py`.
+            keypoints[k][:, :2] -= 0.5
+            result["keypoints"] = keypoints[k].flatten().tolist()
+        results.append(result)
+    return results
+
+
+# inspired from Detectron:
+# https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa
+def _evaluate_box_proposals(dataset_predictions, coco_api, thresholds=None, area="all", limit=None):
+    """
+    Evaluate detection proposal recall metrics. This function is a much
+    faster alternative to the official COCO API recall evaluation code. However,
+    it produces slightly different results.
+    """
+    # Record max overlap value for each gt box
+    # Return vector of overlap values
+    areas = {
+        "all": 0,
+        "small": 1,
+        "medium": 2,
+        "large": 3,
+        "96-128": 4,
+        "128-256": 5,
+        "256-512": 6,
+        "512-inf": 7,
+    }
+    area_ranges = [
+        [0 ** 2, 1e5 ** 2],  # all
+        [0 ** 2, 32 ** 2],  # small
+        [32 ** 2, 96 ** 2],  # medium
+        [96 ** 2, 1e5 ** 2],  # large
+        [96 ** 2, 128 ** 2],  # 96-128
+        [128 ** 2, 256 ** 2],  # 128-256
+        [256 ** 2, 512 ** 2],  # 256-512
+        [512 ** 2, 1e5 ** 2],
+    ]  # 512-inf
+    assert area in areas, "Unknown area range: {}".format(area)
+    area_range = area_ranges[areas[area]]
+    gt_overlaps = []
+    num_pos = 0
+
+    for prediction_dict in dataset_predictions:
+        predictions = prediction_dict["proposals"]
+
+        # sort predictions in descending order
+        # TODO maybe remove this and make it explicit in the documentation
+        inds = predictions.objectness_logits.sort(descending=True)[1]
+        predictions = predictions[inds]
+
+        ann_ids = coco_api.getAnnIds(imgIds=prediction_dict["image_id"])
+        anno = coco_api.loadAnns(ann_ids)
+        gt_boxes = [
+            BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS)
+            for obj in anno
+            if obj["iscrowd"] == 0
+        ]
+        gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4)  # guard against no boxes
+        gt_boxes = Boxes(gt_boxes)
+        gt_areas = torch.as_tensor([obj["area"] for obj in anno if obj["iscrowd"] == 0])
+
+        if len(gt_boxes) == 0 or len(predictions) == 0:
+            continue
+
+        valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1])
+        gt_boxes = gt_boxes[valid_gt_inds]
+
+        num_pos += len(gt_boxes)
+
+        if len(gt_boxes) == 0:
+            continue
+
+        if limit is not None and len(predictions) > limit:
+            predictions = predictions[:limit]
+
+        overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes)
+
+        _gt_overlaps = torch.zeros(len(gt_boxes))
+        for j in range(min(len(predictions), len(gt_boxes))):
+            # find which proposal box maximally covers each gt box
+            # and get the iou amount of coverage for each gt box
+            max_overlaps, argmax_overlaps = overlaps.max(dim=0)
+
+            # find which gt box is 'best' covered (i.e. 'best' = most iou)
+            gt_ovr, gt_ind = max_overlaps.max(dim=0)
+            assert gt_ovr >= 0
+            # find the proposal box that covers the best covered gt box
+            box_ind = argmax_overlaps[gt_ind]
+            # record the iou coverage of this gt box
+            _gt_overlaps[j] = overlaps[box_ind, gt_ind]
+            assert _gt_overlaps[j] == gt_ovr
+            # mark the proposal box and the gt box as used
+            overlaps[box_ind, :] = -1
+            overlaps[:, gt_ind] = -1
+
+        # append recorded iou coverage level
+        gt_overlaps.append(_gt_overlaps)
+    gt_overlaps = (
+        torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32)
+    )
+    gt_overlaps, _ = torch.sort(gt_overlaps)
+
+    if thresholds is None:
+        step = 0.05
+        thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
+    recalls = torch.zeros_like(thresholds)
+    # compute recall for each iou threshold
+    for i, t in enumerate(thresholds):
+        recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
+    # ar = 2 * np.trapz(recalls, thresholds)
+    ar = recalls.mean()
+    return {
+        "ar": ar,
+        "recalls": recalls,
+        "thresholds": thresholds,
+        "gt_overlaps": gt_overlaps,
+        "num_pos": num_pos,
+    }
+
+
+def _evaluate_predictions_on_coco(coco_gt, coco_results, iou_type, kpt_oks_sigmas=None):
+    """
+    Evaluate the coco results using COCOEval API.
+    """
+    assert len(coco_results) > 0
+
+    if iou_type == "segm":
+        coco_results = copy.deepcopy(coco_results)
+        # When evaluating mask AP, if the results contain bbox, cocoapi will
+        # use the box area as the area of the instance, instead of the mask area.
+        # This leads to a different definition of small/medium/large.
+        # We remove the bbox field to let mask AP use mask area.
+        for c in coco_results:
+            c.pop("bbox", None)
+
+    coco_dt = coco_gt.loadRes(coco_results)
+    coco_eval = COCOeval(coco_gt, coco_dt, iou_type)
+    # Use the COCO default keypoint OKS sigmas unless overrides are specified
+    if kpt_oks_sigmas:
+        coco_eval.params.kpt_oks_sigmas = np.array(kpt_oks_sigmas)
+
+    if iou_type == "keypoints":
+        num_keypoints = len(coco_results[0]["keypoints"]) // 3
+        assert len(coco_eval.params.kpt_oks_sigmas) == num_keypoints, (
+            "[COCOEvaluator] The length of cfg.TEST.KEYPOINT_OKS_SIGMAS (default: 17) "
+            "must be equal to the number of keypoints. However the prediction has {} "
+            "keypoints! For more information please refer to "
+            "http://cocodataset.org/#keypoints-eval.".format(num_keypoints)
+        )
+
+    coco_eval.evaluate()
+    coco_eval.accumulate()
+    coco_eval.summarize()
+
+    return coco_eval

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/evaluator.py

@@ -0,0 +1,196 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import datetime
+import logging
+import time
+from collections import OrderedDict
+from contextlib import contextmanager
+import torch
+
+from detectron2.utils.comm import get_world_size, is_main_process
+from detectron2.utils.logger import log_every_n_seconds
+
+
+class DatasetEvaluator:
+    """
+    Base class for a dataset evaluator.
+
+    The function :func:`inference_on_dataset` runs the model over
+    all samples in the dataset, and have a DatasetEvaluator to process the inputs/outputs.
+
+    This class will accumulate information of the inputs/outputs (by :meth:`process`),
+    and produce evaluation results in the end (by :meth:`evaluate`).
+    """
+
+    def reset(self):
+        """
+        Preparation for a new round of evaluation.
+        Should be called before starting a round of evaluation.
+        """
+        pass
+
+    def process(self, inputs, outputs):
+        """
+        Process the pair of inputs and outputs.
+        If they contain batches, the pairs can be consumed one-by-one using `zip`:
+
+        .. code-block:: python
+
+            for input_, output in zip(inputs, outputs):
+                # do evaluation on single input/output pair
+                ...
+
+        Args:
+            inputs (list): the inputs that's used to call the model.
+            outputs (list): the return value of `model(inputs)`
+        """
+        pass
+
+    def evaluate(self):
+        """
+        Evaluate/summarize the performance, after processing all input/output pairs.
+
+        Returns:
+            dict:
+                A new evaluator class can return a dict of arbitrary format
+                as long as the user can process the results.
+                In our train_net.py, we expect the following format:
+
+                * key: the name of the task (e.g., bbox)
+                * value: a dict of {metric name: score}, e.g.: {"AP50": 80}
+        """
+        pass
+
+
+class DatasetEvaluators(DatasetEvaluator):
+    """
+    Wrapper class to combine multiple :class:`DatasetEvaluator` instances.
+
+    This class dispatches every evaluation call to
+    all of its :class:`DatasetEvaluator`.
+    """
+
+    def __init__(self, evaluators):
+        """
+        Args:
+            evaluators (list): the evaluators to combine.
+        """
+        super().__init__()
+        self._evaluators = evaluators
+
+    def reset(self):
+        for evaluator in self._evaluators:
+            evaluator.reset()
+
+    def process(self, inputs, outputs):
+        for evaluator in self._evaluators:
+            evaluator.process(inputs, outputs)
+
+    def evaluate(self):
+        results = OrderedDict()
+        for evaluator in self._evaluators:
+            result = evaluator.evaluate()
+            if is_main_process() and result is not None:
+                for k, v in result.items():
+                    assert (
+                        k not in results
+                    ), "Different evaluators produce results with the same key {}".format(k)
+                    results[k] = v
+        return results
+
+
+def inference_on_dataset(model, data_loader, evaluator):
+    """
+    Run model on the data_loader and evaluate the metrics with evaluator.
+    Also benchmark the inference speed of `model.forward` accurately.
+    The model will be used in eval mode.
+
+    Args:
+        model (nn.Module): a module which accepts an object from
+            `data_loader` and returns some outputs. It will be temporarily set to `eval` mode.
+
+            If you wish to evaluate a model in `training` mode instead, you can
+            wrap the given model and override its behavior of `.eval()` and `.train()`.
+        data_loader: an iterable object with a length.
+            The elements it generates will be the inputs to the model.
+        evaluator (DatasetEvaluator): the evaluator to run. Use `None` if you only want
+            to benchmark, but don't want to do any evaluation.
+
+    Returns:
+        The return value of `evaluator.evaluate()`
+    """
+    num_devices = get_world_size()
+    logger = logging.getLogger(__name__)
+    logger.info("Start inference on {} images".format(len(data_loader)))
+
+    total = len(data_loader)  # inference data loader must have a fixed length
+    if evaluator is None:
+        # create a no-op evaluator
+        evaluator = DatasetEvaluators([])
+    evaluator.reset()
+
+    num_warmup = min(5, total - 1)
+    start_time = time.perf_counter()
+    total_compute_time = 0
+    with inference_context(model), torch.no_grad():
+        for idx, inputs in enumerate(data_loader):
+            if idx == num_warmup:
+                start_time = time.perf_counter()
+                total_compute_time = 0
+
+            start_compute_time = time.perf_counter()
+            outputs = model(inputs)
+            if torch.cuda.is_available():
+                torch.cuda.synchronize()
+            total_compute_time += time.perf_counter() - start_compute_time
+            evaluator.process(inputs, outputs)
+
+            iters_after_start = idx + 1 - num_warmup * int(idx >= num_warmup)
+            seconds_per_img = total_compute_time / iters_after_start
+            if idx >= num_warmup * 2 or seconds_per_img > 5:
+                total_seconds_per_img = (time.perf_counter() - start_time) / iters_after_start
+                eta = datetime.timedelta(seconds=int(total_seconds_per_img * (total - idx - 1)))
+                log_every_n_seconds(
+                    logging.INFO,
+                    "Inference done {}/{}. {:.4f} s / demo. ETA={}".format(
+                        idx + 1, total, seconds_per_img, str(eta)
+                    ),
+                    n=5,
+                )
+
+    # Measure the time only for this worker (before the synchronization barrier)
+    total_time = time.perf_counter() - start_time
+    total_time_str = str(datetime.timedelta(seconds=total_time))
+    # NOTE this format is parsed by grep
+    logger.info(
+        "Total inference time: {} ({:.6f} s / demo per device, on {} devices)".format(
+            total_time_str, total_time / (total - num_warmup), num_devices
+        )
+    )
+    total_compute_time_str = str(datetime.timedelta(seconds=int(total_compute_time)))
+    logger.info(
+        "Total inference pure compute time: {} ({:.6f} s / demo per device, on {} devices)".format(
+            total_compute_time_str, total_compute_time / (total - num_warmup), num_devices
+        )
+    )
+
+    results = evaluator.evaluate()
+    # An evaluator may return None when not in main process.
+    # Replace it by an empty dict instead to make it easier for downstream code to handle
+    if results is None:
+        results = {}
+    return results
+
+
+@contextmanager
+def inference_context(model):
+    """
+    A context where the model is temporarily changed to eval mode,
+    and restored to previous mode afterwards.
+
+    Args:
+        model: a torch Module
+    """
+    training_mode = model.training
+    model.eval()
+    yield
+    model.train(training_mode)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/lvis_evaluation.py

@@ -0,0 +1,350 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import itertools
+import json
+import logging
+import os
+import pickle
+from collections import OrderedDict
+import torch
+from fvcore.common.file_io import PathManager
+
+import detectron2.utils.comm as comm
+from detectron2.data import MetadataCatalog
+from detectron2.structures import Boxes, BoxMode, pairwise_iou
+from detectron2.utils.logger import create_small_table
+
+from .coco_evaluation import instances_to_coco_json
+from .evaluator import DatasetEvaluator
+
+
+class LVISEvaluator(DatasetEvaluator):
+    """
+    Evaluate object proposal and instance detection/segmentation outputs using
+    LVIS's metrics and evaluation API.
+    """
+
+    def __init__(self, dataset_name, cfg, distributed, output_dir=None):
+        """
+        Args:
+            dataset_name (str): name of the dataset to be evaluated.
+                It must have the following corresponding metadata:
+                "json_file": the path to the LVIS format annotation
+            cfg (CfgNode): config instance
+            distributed (True): if True, will collect results from all ranks for evaluation.
+                Otherwise, will evaluate the results in the current process.
+            output_dir (str): optional, an output directory to dump results.
+        """
+        from lvis import LVIS
+
+        self._tasks = self._tasks_from_config(cfg)
+        self._distributed = distributed
+        self._output_dir = output_dir
+
+        self._cpu_device = torch.device("cpu")
+        self._logger = logging.getLogger(__name__)
+
+        self._metadata = MetadataCatalog.get(dataset_name)
+        json_file = PathManager.get_local_path(self._metadata.json_file)
+        self._lvis_api = LVIS(json_file)
+        # Test set json files do not contain annotations (evaluation must be
+        # performed using the LVIS evaluation server).
+        self._do_evaluation = len(self._lvis_api.get_ann_ids()) > 0
+
+    def reset(self):
+        self._predictions = []
+
+    def _tasks_from_config(self, cfg):
+        """
+        Returns:
+            tuple[str]: tasks that can be evaluated under the given configuration.
+        """
+        tasks = ("bbox",)
+        if cfg.MODEL.MASK_ON:
+            tasks = tasks + ("segm",)
+        return tasks
+
+    def process(self, inputs, outputs):
+        """
+        Args:
+            inputs: the inputs to a LVIS model (e.g., GeneralizedRCNN).
+                It is a list of dict. Each dict corresponds to an image and
+                contains keys like "height", "width", "file_name", "image_id".
+            outputs: the outputs of a LVIS model. It is a list of dicts with key
+                "instances" that contains :class:`Instances`.
+        """
+        for input, output in zip(inputs, outputs):
+            prediction = {"image_id": input["image_id"]}
+
+            if "instances" in output:
+                instances = output["instances"].to(self._cpu_device)
+                prediction["instances"] = instances_to_coco_json(instances, input["image_id"])
+            if "proposals" in output:
+                prediction["proposals"] = output["proposals"].to(self._cpu_device)
+            self._predictions.append(prediction)
+
+    def evaluate(self):
+        if self._distributed:
+            comm.synchronize()
+            predictions = comm.gather(self._predictions, dst=0)
+            predictions = list(itertools.chain(*predictions))
+
+            if not comm.is_main_process():
+                return
+        else:
+            predictions = self._predictions
+
+        if len(predictions) == 0:
+            self._logger.warning("[LVISEvaluator] Did not receive valid predictions.")
+            return {}
+
+        if self._output_dir:
+            PathManager.mkdirs(self._output_dir)
+            file_path = os.path.join(self._output_dir, "instances_predictions.pth")
+            with PathManager.open(file_path, "wb") as f:
+                torch.save(predictions, f)
+
+        self._results = OrderedDict()
+        if "proposals" in predictions[0]:
+            self._eval_box_proposals(predictions)
+        if "instances" in predictions[0]:
+            self._eval_predictions(set(self._tasks), predictions)
+        # Copy so the caller can do whatever with results
+        return copy.deepcopy(self._results)
+
+    def _eval_predictions(self, tasks, predictions):
+        """
+        Evaluate predictions on the given tasks.
+        Fill self._results with the metrics of the tasks.
+
+        Args:
+            predictions (list[dict]): list of outputs from the model
+        """
+        self._logger.info("Preparing results in the LVIS format ...")
+        lvis_results = list(itertools.chain(*[x["instances"] for x in predictions]))
+
+        # LVIS evaluator can be used to evaluate results for COCO dataset categories.
+        # In this case `_metadata` variable will have a field with COCO-specific category mapping.
+        if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
+            reverse_id_mapping = {
+                v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items()
+            }
+            for result in lvis_results:
+                result["category_id"] = reverse_id_mapping[result["category_id"]]
+        else:
+            # unmap the category ids for LVIS (from 0-indexed to 1-indexed)
+            for result in lvis_results:
+                result["category_id"] += 1
+
+        if self._output_dir:
+            file_path = os.path.join(self._output_dir, "lvis_instances_results.json")
+            self._logger.info("Saving results to {}".format(file_path))
+            with PathManager.open(file_path, "w") as f:
+                f.write(json.dumps(lvis_results))
+                f.flush()
+
+        if not self._do_evaluation:
+            self._logger.info("Annotations are not available for evaluation.")
+            return
+
+        self._logger.info("Evaluating predictions ...")
+        for task in sorted(tasks):
+            res = _evaluate_predictions_on_lvis(
+                self._lvis_api, lvis_results, task, class_names=self._metadata.get("thing_classes")
+            )
+            self._results[task] = res
+
+    def _eval_box_proposals(self, predictions):
+        """
+        Evaluate the box proposals in predictions.
+        Fill self._results with the metrics for "box_proposals" task.
+        """
+        if self._output_dir:
+            # Saving generated box proposals to file.
+            # Predicted box_proposals are in XYXY_ABS mode.
+            bbox_mode = BoxMode.XYXY_ABS.value
+            ids, boxes, objectness_logits = [], [], []
+            for prediction in predictions:
+                ids.append(prediction["image_id"])
+                boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy())
+                objectness_logits.append(prediction["proposals"].objectness_logits.numpy())
+
+            proposal_data = {
+                "boxes": boxes,
+                "objectness_logits": objectness_logits,
+                "ids": ids,
+                "bbox_mode": bbox_mode,
+            }
+            with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f:
+                pickle.dump(proposal_data, f)
+
+        if not self._do_evaluation:
+            self._logger.info("Annotations are not available for evaluation.")
+            return
+
+        self._logger.info("Evaluating bbox proposals ...")
+        res = {}
+        areas = {"all": "", "small": "s", "medium": "m", "large": "l"}
+        for limit in [100, 1000]:
+            for area, suffix in areas.items():
+                stats = _evaluate_box_proposals(predictions, self._lvis_api, area=area, limit=limit)
+                key = "AR{}@{:d}".format(suffix, limit)
+                res[key] = float(stats["ar"].item() * 100)
+        self._logger.info("Proposal metrics: \n" + create_small_table(res))
+        self._results["box_proposals"] = res
+
+
+# inspired from Detectron:
+# https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa
+def _evaluate_box_proposals(dataset_predictions, lvis_api, thresholds=None, area="all", limit=None):
+    """
+    Evaluate detection proposal recall metrics. This function is a much
+    faster alternative to the official LVIS API recall evaluation code. However,
+    it produces slightly different results.
+    """
+    # Record max overlap value for each gt box
+    # Return vector of overlap values
+    areas = {
+        "all": 0,
+        "small": 1,
+        "medium": 2,
+        "large": 3,
+        "96-128": 4,
+        "128-256": 5,
+        "256-512": 6,
+        "512-inf": 7,
+    }
+    area_ranges = [
+        [0 ** 2, 1e5 ** 2],  # all
+        [0 ** 2, 32 ** 2],  # small
+        [32 ** 2, 96 ** 2],  # medium
+        [96 ** 2, 1e5 ** 2],  # large
+        [96 ** 2, 128 ** 2],  # 96-128
+        [128 ** 2, 256 ** 2],  # 128-256
+        [256 ** 2, 512 ** 2],  # 256-512
+        [512 ** 2, 1e5 ** 2],
+    ]  # 512-inf
+    assert area in areas, "Unknown area range: {}".format(area)
+    area_range = area_ranges[areas[area]]
+    gt_overlaps = []
+    num_pos = 0
+
+    for prediction_dict in dataset_predictions:
+        predictions = prediction_dict["proposals"]
+
+        # sort predictions in descending order
+        # TODO maybe remove this and make it explicit in the documentation
+        inds = predictions.objectness_logits.sort(descending=True)[1]
+        predictions = predictions[inds]
+
+        ann_ids = lvis_api.get_ann_ids(img_ids=[prediction_dict["image_id"]])
+        anno = lvis_api.load_anns(ann_ids)
+        gt_boxes = [
+            BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS) for obj in anno
+        ]
+        gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4)  # guard against no boxes
+        gt_boxes = Boxes(gt_boxes)
+        gt_areas = torch.as_tensor([obj["area"] for obj in anno])
+
+        if len(gt_boxes) == 0 or len(predictions) == 0:
+            continue
+
+        valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1])
+        gt_boxes = gt_boxes[valid_gt_inds]
+
+        num_pos += len(gt_boxes)
+
+        if len(gt_boxes) == 0:
+            continue
+
+        if limit is not None and len(predictions) > limit:
+            predictions = predictions[:limit]
+
+        overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes)
+
+        _gt_overlaps = torch.zeros(len(gt_boxes))
+        for j in range(min(len(predictions), len(gt_boxes))):
+            # find which proposal box maximally covers each gt box
+            # and get the iou amount of coverage for each gt box
+            max_overlaps, argmax_overlaps = overlaps.max(dim=0)
+
+            # find which gt box is 'best' covered (i.e. 'best' = most iou)
+            gt_ovr, gt_ind = max_overlaps.max(dim=0)
+            assert gt_ovr >= 0
+            # find the proposal box that covers the best covered gt box
+            box_ind = argmax_overlaps[gt_ind]
+            # record the iou coverage of this gt box
+            _gt_overlaps[j] = overlaps[box_ind, gt_ind]
+            assert _gt_overlaps[j] == gt_ovr
+            # mark the proposal box and the gt box as used
+            overlaps[box_ind, :] = -1
+            overlaps[:, gt_ind] = -1
+
+        # append recorded iou coverage level
+        gt_overlaps.append(_gt_overlaps)
+    gt_overlaps = (
+        torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32)
+    )
+    gt_overlaps, _ = torch.sort(gt_overlaps)
+
+    if thresholds is None:
+        step = 0.05
+        thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
+    recalls = torch.zeros_like(thresholds)
+    # compute recall for each iou threshold
+    for i, t in enumerate(thresholds):
+        recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
+    # ar = 2 * np.trapz(recalls, thresholds)
+    ar = recalls.mean()
+    return {
+        "ar": ar,
+        "recalls": recalls,
+        "thresholds": thresholds,
+        "gt_overlaps": gt_overlaps,
+        "num_pos": num_pos,
+    }
+
+
+def _evaluate_predictions_on_lvis(lvis_gt, lvis_results, iou_type, class_names=None):
+    """
+    Args:
+        iou_type (str):
+        kpt_oks_sigmas (list[float]):
+        class_names (None or list[str]): if provided, will use it to predict
+            per-category AP.
+
+    Returns:
+        a dict of {metric name: score}
+    """
+    metrics = {
+        "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
+        "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
+    }[iou_type]
+
+    logger = logging.getLogger(__name__)
+
+    if len(lvis_results) == 0:  # TODO: check if needed
+        logger.warn("No predictions from the model!")
+        return {metric: float("nan") for metric in metrics}
+
+    if iou_type == "segm":
+        lvis_results = copy.deepcopy(lvis_results)
+        # When evaluating mask AP, if the results contain bbox, LVIS API will
+        # use the box area as the area of the instance, instead of the mask area.
+        # This leads to a different definition of small/medium/large.
+        # We remove the bbox field to let mask AP use mask area.
+        for c in lvis_results:
+            c.pop("bbox", None)
+
+    from lvis import LVISEval, LVISResults
+
+    lvis_results = LVISResults(lvis_gt, lvis_results)
+    lvis_eval = LVISEval(lvis_gt, lvis_results, iou_type)
+    lvis_eval.run()
+    lvis_eval.print_results()
+
+    # Pull the standard metrics from the LVIS results
+    results = lvis_eval.get_results()
+    results = {metric: float(results[metric] * 100) for metric in metrics}
+    logger.info("Evaluation results for {}: \n".format(iou_type) + create_small_table(results))
+    return results

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/panoptic_evaluation.py

@@ -0,0 +1,167 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import contextlib
+import io
+import itertools
+import json
+import logging
+import os
+import tempfile
+from collections import OrderedDict
+from fvcore.common.file_io import PathManager
+from PIL import Image
+from tabulate import tabulate
+
+from detectron2.data import MetadataCatalog
+from detectron2.utils import comm
+
+from .evaluator import DatasetEvaluator
+
+logger = logging.getLogger(__name__)
+
+
+class COCOPanopticEvaluator(DatasetEvaluator):
+    """
+    Evaluate Panoptic Quality metrics on COCO using PanopticAPI.
+    It saves panoptic segmentation prediction in `output_dir`
+
+    It contains a synchronize call and has to be called from all workers.
+    """
+
+    def __init__(self, dataset_name, output_dir):
+        """
+        Args:
+            dataset_name (str): name of the dataset
+            output_dir (str): output directory to save results for evaluation
+        """
+        self._metadata = MetadataCatalog.get(dataset_name)
+        self._thing_contiguous_id_to_dataset_id = {
+            v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items()
+        }
+        self._stuff_contiguous_id_to_dataset_id = {
+            v: k for k, v in self._metadata.stuff_dataset_id_to_contiguous_id.items()
+        }
+
+        self._predictions_json = os.path.join(output_dir, "predictions.json")
+
+    def reset(self):
+        self._predictions = []
+
+    def _convert_category_id(self, segment_info):
+        isthing = segment_info.pop("isthing", None)
+        if isthing is None:
+            # the model produces panoptic category id directly. No more conversion needed
+            return segment_info
+        if isthing is True:
+            segment_info["category_id"] = self._thing_contiguous_id_to_dataset_id[
+                segment_info["category_id"]
+            ]
+        else:
+            segment_info["category_id"] = self._stuff_contiguous_id_to_dataset_id[
+                segment_info["category_id"]
+            ]
+        return segment_info
+
+    def process(self, inputs, outputs):
+        from panopticapi.utils import id2rgb
+
+        for input, output in zip(inputs, outputs):
+            panoptic_img, segments_info = output["panoptic_seg"]
+            panoptic_img = panoptic_img.cpu().numpy()
+
+            file_name = os.path.basename(input["file_name"])
+            file_name_png = os.path.splitext(file_name)[0] + ".png"
+            with io.BytesIO() as out:
+                Image.fromarray(id2rgb(panoptic_img)).save(out, format="PNG")
+                segments_info = [self._convert_category_id(x) for x in segments_info]
+                self._predictions.append(
+                    {
+                        "image_id": input["image_id"],
+                        "file_name": file_name_png,
+                        "png_string": out.getvalue(),
+                        "segments_info": segments_info,
+                    }
+                )
+
+    def evaluate(self):
+        comm.synchronize()
+
+        self._predictions = comm.gather(self._predictions)
+        self._predictions = list(itertools.chain(*self._predictions))
+        if not comm.is_main_process():
+            return
+
+        # PanopticApi requires local files
+        gt_json = PathManager.get_local_path(self._metadata.panoptic_json)
+        gt_folder = PathManager.get_local_path(self._metadata.panoptic_root)
+
+        with tempfile.TemporaryDirectory(prefix="panoptic_eval") as pred_dir:
+            logger.info("Writing all panoptic predictions to {} ...".format(pred_dir))
+            for p in self._predictions:
+                with open(os.path.join(pred_dir, p["file_name"]), "wb") as f:
+                    f.write(p.pop("png_string"))
+
+            with open(gt_json, "r") as f:
+                json_data = json.load(f)
+            json_data["annotations"] = self._predictions
+            with PathManager.open(self._predictions_json, "w") as f:
+                f.write(json.dumps(json_data))
+
+            from panopticapi.evaluation import pq_compute
+
+            with contextlib.redirect_stdout(io.StringIO()):
+                pq_res = pq_compute(
+                    gt_json,
+                    PathManager.get_local_path(self._predictions_json),
+                    gt_folder=gt_folder,
+                    pred_folder=pred_dir,
+                )
+
+        res = {}
+        res["PQ"] = 100 * pq_res["All"]["pq"]
+        res["SQ"] = 100 * pq_res["All"]["sq"]
+        res["RQ"] = 100 * pq_res["All"]["rq"]
+        res["PQ_th"] = 100 * pq_res["Things"]["pq"]
+        res["SQ_th"] = 100 * pq_res["Things"]["sq"]
+        res["RQ_th"] = 100 * pq_res["Things"]["rq"]
+        res["PQ_st"] = 100 * pq_res["Stuff"]["pq"]
+        res["SQ_st"] = 100 * pq_res["Stuff"]["sq"]
+        res["RQ_st"] = 100 * pq_res["Stuff"]["rq"]
+
+        results = OrderedDict({"panoptic_seg": res})
+        _print_panoptic_results(pq_res)
+
+        return results
+
+
+def _print_panoptic_results(pq_res):
+    headers = ["", "PQ", "SQ", "RQ", "#categories"]
+    data = []
+    for name in ["All", "Things", "Stuff"]:
+        row = [name] + [pq_res[name][k] * 100 for k in ["pq", "sq", "rq"]] + [pq_res[name]["n"]]
+        data.append(row)
+    table = tabulate(
+        data, headers=headers, tablefmt="pipe", floatfmt=".3f", stralign="center", numalign="center"
+    )
+    logger.info("Panoptic Evaluation Results:\n" + table)
+
+
+if __name__ == "__main__":
+    from detectron2.utils.logger import setup_logger
+
+    logger = setup_logger()
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--gt-json")
+    parser.add_argument("--gt-dir")
+    parser.add_argument("--pred-json")
+    parser.add_argument("--pred-dir")
+    args = parser.parse_args()
+
+    from panopticapi.evaluation import pq_compute
+
+    with contextlib.redirect_stdout(io.StringIO()):
+        pq_res = pq_compute(
+            args.gt_json, args.pred_json, gt_folder=args.gt_dir, pred_folder=args.pred_dir
+        )
+        _print_panoptic_results(pq_res)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/pascal_voc_evaluation.py

@@ -0,0 +1,294 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import logging
+import numpy as np
+import os
+import tempfile
+import xml.etree.ElementTree as ET
+from collections import OrderedDict, defaultdict
+from functools import lru_cache
+import torch
+from fvcore.common.file_io import PathManager
+
+from detectron2.data import MetadataCatalog
+from detectron2.utils import comm
+
+from .evaluator import DatasetEvaluator
+
+
+class PascalVOCDetectionEvaluator(DatasetEvaluator):
+    """
+    Evaluate Pascal VOC AP.
+    It contains a synchronization, therefore has to be called from all ranks.
+
+    Note that this is a rewrite of the official Matlab API.
+    The results should be similar, but not identical to the one produced by
+    the official API.
+    """
+
+    def __init__(self, dataset_name):
+        """
+        Args:
+            dataset_name (str): name of the dataset, e.g., "voc_2007_test"
+        """
+        self._dataset_name = dataset_name
+        meta = MetadataCatalog.get(dataset_name)
+        self._anno_file_template = os.path.join(meta.dirname, "Annotations", "{}.xml")
+        self._image_set_path = os.path.join(meta.dirname, "ImageSets", "Main", meta.split + ".txt")
+        self._class_names = meta.thing_classes
+        assert meta.year in [2007, 2012], meta.year
+        self._is_2007 = meta.year == 2007
+        self._cpu_device = torch.device("cpu")
+        self._logger = logging.getLogger(__name__)
+
+    def reset(self):
+        self._predictions = defaultdict(list)  # class name -> list of prediction strings
+
+    def process(self, inputs, outputs):
+        for input, output in zip(inputs, outputs):
+            image_id = input["image_id"]
+            instances = output["instances"].to(self._cpu_device)
+            boxes = instances.pred_boxes.tensor.numpy()
+            scores = instances.scores.tolist()
+            classes = instances.pred_classes.tolist()
+            for box, score, cls in zip(boxes, scores, classes):
+                xmin, ymin, xmax, ymax = box
+                # The inverse of data loading logic in `data/pascal_voc.py`
+                xmin += 1
+                ymin += 1
+                self._predictions[cls].append(
+                    f"{image_id} {score:.3f} {xmin:.1f} {ymin:.1f} {xmax:.1f} {ymax:.1f}"
+                )
+
+    def evaluate(self):
+        """
+        Returns:
+            dict: has a key "segm", whose value is a dict of "AP", "AP50", and "AP75".
+        """
+        all_predictions = comm.gather(self._predictions, dst=0)
+        if not comm.is_main_process():
+            return
+        predictions = defaultdict(list)
+        for predictions_per_rank in all_predictions:
+            for clsid, lines in predictions_per_rank.items():
+                predictions[clsid].extend(lines)
+        del all_predictions
+
+        self._logger.info(
+            "Evaluating {} using {} metric. "
+            "Note that results do not use the official Matlab API.".format(
+                self._dataset_name, 2007 if self._is_2007 else 2012
+            )
+        )
+
+        with tempfile.TemporaryDirectory(prefix="pascal_voc_eval_") as dirname:
+            res_file_template = os.path.join(dirname, "{}.txt")
+
+            aps = defaultdict(list)  # iou -> ap per class
+            for cls_id, cls_name in enumerate(self._class_names):
+                lines = predictions.get(cls_id, [""])
+
+                with open(res_file_template.format(cls_name), "w") as f:
+                    f.write("\n".join(lines))
+
+                for thresh in range(50, 100, 5):
+                    rec, prec, ap = voc_eval(
+                        res_file_template,
+                        self._anno_file_template,
+                        self._image_set_path,
+                        cls_name,
+                        ovthresh=thresh / 100.0,
+                        use_07_metric=self._is_2007,
+                    )
+                    aps[thresh].append(ap * 100)
+
+        ret = OrderedDict()
+        mAP = {iou: np.mean(x) for iou, x in aps.items()}
+        ret["bbox"] = {"AP": np.mean(list(mAP.values())), "AP50": mAP[50], "AP75": mAP[75]}
+        return ret
+
+
+##############################################################################
+#
+# Below code is modified from
+# https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/datasets/voc_eval.py
+# --------------------------------------------------------
+# Fast/er R-CNN
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Bharath Hariharan
+# --------------------------------------------------------
+
+"""Python implementation of the PASCAL VOC devkit's AP evaluation code."""
+
+
+@lru_cache(maxsize=None)
+def parse_rec(filename):
+    """Parse a PASCAL VOC xml file."""
+    with PathManager.open(filename) as f:
+        tree = ET.parse(f)
+    objects = []
+    for obj in tree.findall("object"):
+        obj_struct = {}
+        obj_struct["name"] = obj.find("name").text
+        obj_struct["pose"] = obj.find("pose").text
+        obj_struct["truncated"] = int(obj.find("truncated").text)
+        obj_struct["difficult"] = int(obj.find("difficult").text)
+        bbox = obj.find("bndbox")
+        obj_struct["bbox"] = [
+            int(bbox.find("xmin").text),
+            int(bbox.find("ymin").text),
+            int(bbox.find("xmax").text),
+            int(bbox.find("ymax").text),
+        ]
+        objects.append(obj_struct)
+
+    return objects
+
+
+def voc_ap(rec, prec, use_07_metric=False):
+    """Compute VOC AP given precision and recall. If use_07_metric is true, uses
+    the VOC 07 11-point method (default:False).
+    """
+    if use_07_metric:
+        # 11 point metric
+        ap = 0.0
+        for t in np.arange(0.0, 1.1, 0.1):
+            if np.sum(rec >= t) == 0:
+                p = 0
+            else:
+                p = np.max(prec[rec >= t])
+            ap = ap + p / 11.0
+    else:
+        # correct AP calculation
+        # first append sentinel values at the end
+        mrec = np.concatenate(([0.0], rec, [1.0]))
+        mpre = np.concatenate(([0.0], prec, [0.0]))
+
+        # compute the precision envelope
+        for i in range(mpre.size - 1, 0, -1):
+            mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i])
+
+        # to calculate area under PR curve, look for points
+        # where X axis (recall) changes value
+        i = np.where(mrec[1:] != mrec[:-1])[0]
+
+        # and sum (\Delta recall) * prec
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])
+    return ap
+
+
+def voc_eval(detpath, annopath, imagesetfile, classname, ovthresh=0.5, use_07_metric=False):
+    """rec, prec, ap = voc_eval(detpath,
+                                annopath,
+                                imagesetfile,
+                                classname,
+                                [ovthresh],
+                                [use_07_metric])
+
+    Top level function that does the PASCAL VOC evaluation.
+
+    detpath: Path to detections
+        detpath.format(classname) should produce the detection results file.
+    annopath: Path to annotations
+        annopath.format(imagename) should be the xml annotations file.
+    imagesetfile: Text file containing the list of images, one image per line.
+    classname: Category name (duh)
+    [ovthresh]: Overlap threshold (default = 0.5)
+    [use_07_metric]: Whether to use VOC07's 11 point AP computation
+        (default False)
+    """
+    # assumes detections are in detpath.format(classname)
+    # assumes annotations are in annopath.format(imagename)
+    # assumes imagesetfile is a text file with each line an image name
+
+    # first load gt
+    # read list of images
+    with PathManager.open(imagesetfile, "r") as f:
+        lines = f.readlines()
+    imagenames = [x.strip() for x in lines]
+
+    # load annots
+    recs = {}
+    for imagename in imagenames:
+        recs[imagename] = parse_rec(annopath.format(imagename))
+
+    # extract gt objects for this class
+    class_recs = {}
+    npos = 0
+    for imagename in imagenames:
+        R = [obj for obj in recs[imagename] if obj["name"] == classname]
+        bbox = np.array([x["bbox"] for x in R])
+        difficult = np.array([x["difficult"] for x in R]).astype(np.bool)
+        # difficult = np.array([False for x in R]).astype(np.bool)  # treat all "difficult" as GT
+        det = [False] * len(R)
+        npos = npos + sum(~difficult)
+        class_recs[imagename] = {"bbox": bbox, "difficult": difficult, "det": det}
+
+    # read dets
+    detfile = detpath.format(classname)
+    with open(detfile, "r") as f:
+        lines = f.readlines()
+
+    splitlines = [x.strip().split(" ") for x in lines]
+    image_ids = [x[0] for x in splitlines]
+    confidence = np.array([float(x[1]) for x in splitlines])
+    BB = np.array([[float(z) for z in x[2:]] for x in splitlines]).reshape(-1, 4)
+
+    # sort by confidence
+    sorted_ind = np.argsort(-confidence)
+    BB = BB[sorted_ind, :]
+    image_ids = [image_ids[x] for x in sorted_ind]
+
+    # go down dets and mark TPs and FPs
+    nd = len(image_ids)
+    tp = np.zeros(nd)
+    fp = np.zeros(nd)
+    for d in range(nd):
+        R = class_recs[image_ids[d]]
+        bb = BB[d, :].astype(float)
+        ovmax = -np.inf
+        BBGT = R["bbox"].astype(float)
+
+        if BBGT.size > 0:
+            # compute overlaps
+            # intersection
+            ixmin = np.maximum(BBGT[:, 0], bb[0])
+            iymin = np.maximum(BBGT[:, 1], bb[1])
+            ixmax = np.minimum(BBGT[:, 2], bb[2])
+            iymax = np.minimum(BBGT[:, 3], bb[3])
+            iw = np.maximum(ixmax - ixmin + 1.0, 0.0)
+            ih = np.maximum(iymax - iymin + 1.0, 0.0)
+            inters = iw * ih
+
+            # union
+            uni = (
+                (bb[2] - bb[0] + 1.0) * (bb[3] - bb[1] + 1.0)
+                + (BBGT[:, 2] - BBGT[:, 0] + 1.0) * (BBGT[:, 3] - BBGT[:, 1] + 1.0)
+                - inters
+            )
+
+            overlaps = inters / uni
+            ovmax = np.max(overlaps)
+            jmax = np.argmax(overlaps)
+
+        if ovmax > ovthresh:
+            if not R["difficult"][jmax]:
+                if not R["det"][jmax]:
+                    tp[d] = 1.0
+                    R["det"][jmax] = 1
+                else:
+                    fp[d] = 1.0
+        else:
+            fp[d] = 1.0
+
+    # compute precision recall
+    fp = np.cumsum(fp)
+    tp = np.cumsum(tp)
+    rec = tp / float(npos)
+    # avoid divide by zero in case the first detection matches a difficult
+    # ground truth
+    prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps)
+    ap = voc_ap(rec, prec, use_07_metric)
+
+    return rec, prec, ap

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/rotated_coco_evaluation.py

@@ -0,0 +1,204 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import itertools
+import json
+import numpy as np
+import os
+import torch
+from fvcore.common.file_io import PathManager
+from pycocotools.cocoeval import COCOeval, maskUtils
+
+from detectron2.structures import BoxMode, RotatedBoxes, pairwise_iou_rotated
+
+from .coco_evaluation import COCOEvaluator
+
+
+class RotatedCOCOeval(COCOeval):
+    @staticmethod
+    def is_rotated(box_list):
+        if type(box_list) == np.ndarray:
+            return box_list.shape[1] == 5
+        elif type(box_list) == list:
+            if box_list == []:  # cannot decide the box_dim
+                return False
+            return np.all(
+                np.array(
+                    [
+                        (len(obj) == 5) and ((type(obj) == list) or (type(obj) == np.ndarray))
+                        for obj in box_list
+                    ]
+                )
+            )
+        return False
+
+    @staticmethod
+    def boxlist_to_tensor(boxlist, output_box_dim):
+        if type(boxlist) == np.ndarray:
+            box_tensor = torch.from_numpy(boxlist)
+        elif type(boxlist) == list:
+            if boxlist == []:
+                return torch.zeros((0, output_box_dim), dtype=torch.float32)
+            else:
+                box_tensor = torch.FloatTensor(boxlist)
+        else:
+            raise Exception("Unrecognized boxlist type")
+
+        input_box_dim = box_tensor.shape[1]
+        if input_box_dim != output_box_dim:
+            if input_box_dim == 4 and output_box_dim == 5:
+                box_tensor = BoxMode.convert(box_tensor, BoxMode.XYWH_ABS, BoxMode.XYWHA_ABS)
+            else:
+                raise Exception(
+                    "Unable to convert from {}-dim box to {}-dim box".format(
+                        input_box_dim, output_box_dim
+                    )
+                )
+        return box_tensor
+
+    def compute_iou_dt_gt(self, dt, gt, is_crowd):
+        if self.is_rotated(dt) or self.is_rotated(gt):
+            # TODO: take is_crowd into consideration
+            assert all(c == 0 for c in is_crowd)
+            dt = RotatedBoxes(self.boxlist_to_tensor(dt, output_box_dim=5))
+            gt = RotatedBoxes(self.boxlist_to_tensor(gt, output_box_dim=5))
+            return pairwise_iou_rotated(dt, gt)
+        else:
+            # This is the same as the classical COCO evaluation
+            return maskUtils.iou(dt, gt, is_crowd)
+
+    def computeIoU(self, imgId, catId):
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return []
+        inds = np.argsort([-d["score"] for d in dt], kind="mergesort")
+        dt = [dt[i] for i in inds]
+        if len(dt) > p.maxDets[-1]:
+            dt = dt[0 : p.maxDets[-1]]
+
+        assert p.iouType == "bbox", "unsupported iouType for iou computation"
+
+        g = [g["bbox"] for g in gt]
+        d = [d["bbox"] for d in dt]
+
+        # compute iou between each dt and gt region
+        iscrowd = [int(o["iscrowd"]) for o in gt]
+
+        # Note: this function is copied from cocoeval.py in cocoapi
+        # and the major difference is here.
+        ious = self.compute_iou_dt_gt(d, g, iscrowd)
+        return ious
+
+
+class RotatedCOCOEvaluator(COCOEvaluator):
+    """
+    Evaluate object proposal/instance detection outputs using COCO-like metrics and APIs,
+    with rotated boxes support.
+    Note: this uses IOU only and does not consider angle differences.
+    """
+
+    def process(self, inputs, outputs):
+        """
+        Args:
+            inputs: the inputs to a COCO model (e.g., GeneralizedRCNN).
+                It is a list of dict. Each dict corresponds to an image and
+                contains keys like "height", "width", "file_name", "image_id".
+            outputs: the outputs of a COCO model. It is a list of dicts with key
+                "instances" that contains :class:`Instances`.
+        """
+        for input, output in zip(inputs, outputs):
+            prediction = {"image_id": input["image_id"]}
+
+            if "instances" in output:
+                instances = output["instances"].to(self._cpu_device)
+
+                prediction["instances"] = self.instances_to_json(instances, input["image_id"])
+            if "proposals" in output:
+                prediction["proposals"] = output["proposals"].to(self._cpu_device)
+            self._predictions.append(prediction)
+
+    def instances_to_json(self, instances, img_id):
+        num_instance = len(instances)
+        if num_instance == 0:
+            return []
+
+        boxes = instances.pred_boxes.tensor.numpy()
+        if boxes.shape[1] == 4:
+            boxes = BoxMode.convert(boxes, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+        boxes = boxes.tolist()
+        scores = instances.scores.tolist()
+        classes = instances.pred_classes.tolist()
+
+        results = []
+        for k in range(num_instance):
+            result = {
+                "image_id": img_id,
+                "category_id": classes[k],
+                "bbox": boxes[k],
+                "score": scores[k],
+            }
+
+            results.append(result)
+        return results
+
+    def _eval_predictions(self, tasks, predictions):
+        """
+        Evaluate predictions on the given tasks.
+        Fill self._results with the metrics of the tasks.
+        """
+        self._logger.info("Preparing results for COCO format ...")
+        coco_results = list(itertools.chain(*[x["instances"] for x in predictions]))
+
+        # unmap the category ids for COCO
+        if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
+            reverse_id_mapping = {
+                v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items()
+            }
+            for result in coco_results:
+                result["category_id"] = reverse_id_mapping[result["category_id"]]
+
+        if self._output_dir:
+            file_path = os.path.join(self._output_dir, "coco_instances_results.json")
+            self._logger.info("Saving results to {}".format(file_path))
+            with PathManager.open(file_path, "w") as f:
+                f.write(json.dumps(coco_results))
+                f.flush()
+
+        if not self._do_evaluation:
+            self._logger.info("Annotations are not available for evaluation.")
+            return
+
+        self._logger.info("Evaluating predictions ...")
+        for task in sorted(tasks):
+            assert task == "bbox", "Task {} is not supported".format(task)
+            coco_eval = (
+                self._evaluate_predictions_on_coco(self._coco_api, coco_results)
+                if len(coco_results) > 0
+                else None  # cocoapi does not handle empty results very well
+            )
+
+            res = self._derive_coco_results(
+                coco_eval, task, class_names=self._metadata.get("thing_classes")
+            )
+            self._results[task] = res
+
+    def _evaluate_predictions_on_coco(self, coco_gt, coco_results):
+        """
+        Evaluate the coco results using COCOEval API.
+        """
+        assert len(coco_results) > 0
+
+        coco_dt = coco_gt.loadRes(coco_results)
+
+        # Only bbox is supported for now
+        coco_eval = RotatedCOCOeval(coco_gt, coco_dt, iouType="bbox")
+
+        coco_eval.evaluate()
+        coco_eval.accumulate()
+        coco_eval.summarize()
+
+        return coco_eval

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/sem_seg_evaluation.py

@@ -0,0 +1,168 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import itertools
+import json
+import logging
+import numpy as np
+import os
+from collections import OrderedDict
+import PIL.Image as Image
+import pycocotools.mask as mask_util
+import torch
+from fvcore.common.file_io import PathManager
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+from detectron2.utils.comm import all_gather, is_main_process, synchronize
+
+from .evaluator import DatasetEvaluator
+
+
+class SemSegEvaluator(DatasetEvaluator):
+    """
+    Evaluate semantic segmentation
+    """
+
+    def __init__(self, dataset_name, distributed, num_classes, ignore_label=255, output_dir=None):
+        """
+        Args:
+            dataset_name (str): name of the dataset to be evaluated.
+            distributed (True): if True, will collect results from all ranks for evaluation.
+                Otherwise, will evaluate the results in the current process.
+            num_classes (int): number of classes
+            ignore_label (int): value in semantic segmentation ground truth. Predictions for the
+            corresponding pixels should be ignored.
+            output_dir (str): an output directory to dump results.
+        """
+        self._dataset_name = dataset_name
+        self._distributed = distributed
+        self._output_dir = output_dir
+        self._num_classes = num_classes
+        self._ignore_label = ignore_label
+        self._N = num_classes + 1
+
+        self._cpu_device = torch.device("cpu")
+        self._logger = logging.getLogger(__name__)
+
+        self.input_file_to_gt_file = {
+            dataset_record["file_name"]: dataset_record["sem_seg_file_name"]
+            for dataset_record in DatasetCatalog.get(dataset_name)
+        }
+
+        meta = MetadataCatalog.get(dataset_name)
+        # Dict that maps contiguous training ids to COCO category ids
+        try:
+            c2d = meta.stuff_dataset_id_to_contiguous_id
+            self._contiguous_id_to_dataset_id = {v: k for k, v in c2d.items()}
+        except AttributeError:
+            self._contiguous_id_to_dataset_id = None
+        self._class_names = meta.stuff_classes
+
+    def reset(self):
+        self._conf_matrix = np.zeros((self._N, self._N), dtype=np.int64)
+        self._predictions = []
+
+    def process(self, inputs, outputs):
+        """
+        Args:
+            inputs: the inputs to a model.
+                It is a list of dicts. Each dict corresponds to an image and
+                contains keys like "height", "width", "file_name".
+            outputs: the outputs of a model. It is either list of semantic segmentation predictions
+                (Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic
+                segmentation prediction in the same format.
+        """
+        for input, output in zip(inputs, outputs):
+            output = output["sem_seg"].argmax(dim=0).to(self._cpu_device)
+            pred = np.array(output, dtype=np.int)
+            with PathManager.open(self.input_file_to_gt_file[input["file_name"]], "rb") as f:
+                gt = np.array(Image.open(f), dtype=np.int)
+
+            gt[gt == self._ignore_label] = self._num_classes
+
+            self._conf_matrix += np.bincount(
+                self._N * pred.reshape(-1) + gt.reshape(-1), minlength=self._N ** 2
+            ).reshape(self._N, self._N)
+
+            self._predictions.extend(self.encode_json_sem_seg(pred, input["file_name"]))
+
+    def evaluate(self):
+        """
+        Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval):
+
+        * Mean intersection-over-union averaged across classes (mIoU)
+        * Frequency Weighted IoU (fwIoU)
+        * Mean pixel accuracy averaged across classes (mACC)
+        * Pixel Accuracy (pACC)
+        """
+        if self._distributed:
+            synchronize()
+            conf_matrix_list = all_gather(self._conf_matrix)
+            self._predictions = all_gather(self._predictions)
+            self._predictions = list(itertools.chain(*self._predictions))
+            if not is_main_process():
+                return
+
+            self._conf_matrix = np.zeros_like(self._conf_matrix)
+            for conf_matrix in conf_matrix_list:
+                self._conf_matrix += conf_matrix
+
+        if self._output_dir:
+            PathManager.mkdirs(self._output_dir)
+            file_path = os.path.join(self._output_dir, "sem_seg_predictions.json")
+            with PathManager.open(file_path, "w") as f:
+                f.write(json.dumps(self._predictions))
+
+        acc = np.full(self._num_classes, np.nan, dtype=np.float)
+        iou = np.full(self._num_classes, np.nan, dtype=np.float)
+        tp = self._conf_matrix.diagonal()[:-1].astype(np.float)
+        pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float)
+        class_weights = pos_gt / np.sum(pos_gt)
+        pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float)
+        acc_valid = pos_gt > 0
+        acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid]
+        iou_valid = (pos_gt + pos_pred) > 0
+        union = pos_gt + pos_pred - tp
+        iou[acc_valid] = tp[acc_valid] / union[acc_valid]
+        macc = np.sum(acc[acc_valid]) / np.sum(acc_valid)
+        miou = np.sum(iou[acc_valid]) / np.sum(iou_valid)
+        fiou = np.sum(iou[acc_valid] * class_weights[acc_valid])
+        pacc = np.sum(tp) / np.sum(pos_gt)
+
+        res = {}
+        res["mIoU"] = 100 * miou
+        res["fwIoU"] = 100 * fiou
+        for i, name in enumerate(self._class_names):
+            res["IoU-{}".format(name)] = 100 * iou[i]
+        res["mACC"] = 100 * macc
+        res["pACC"] = 100 * pacc
+        for i, name in enumerate(self._class_names):
+            res["ACC-{}".format(name)] = 100 * acc[i]
+
+        if self._output_dir:
+            file_path = os.path.join(self._output_dir, "sem_seg_evaluation.pth")
+            with PathManager.open(file_path, "wb") as f:
+                torch.save(res, f)
+        results = OrderedDict({"sem_seg": res})
+        self._logger.info(results)
+        return results
+
+    def encode_json_sem_seg(self, sem_seg, input_file_name):
+        """
+        Convert semantic segmentation to COCO stuff format with segments encoded as RLEs.
+        See http://cocodataset.org/#format-results
+        """
+        json_list = []
+        for label in np.unique(sem_seg):
+            if self._contiguous_id_to_dataset_id is not None:
+                assert (
+                    label in self._contiguous_id_to_dataset_id
+                ), "Label {} is not in the metadata info for {}".format(label, self._dataset_name)
+                dataset_id = self._contiguous_id_to_dataset_id[label]
+            else:
+                dataset_id = int(label)
+            mask = (sem_seg == label).astype(np.uint8)
+            mask_rle = mask_util.encode(np.array(mask[:, :, None], order="F"))[0]
+            mask_rle["counts"] = mask_rle["counts"].decode("utf-8")
+            json_list.append(
+                {"file_name": input_file_name, "category_id": dataset_id, "segmentation": mask_rle}
+            )
+        return json_list

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/evaluation/testing.py

@@ -0,0 +1,78 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
+import numpy as np
+import pprint
+import sys
+from collections import OrderedDict
+from collections.abc import Mapping
+
+
+def print_csv_format(results):
+    """
+    Print main metrics in a format similar to Detectron,
+    so that they are easy to copypaste into a spreadsheet.
+
+    Args:
+        results (OrderedDict[dict]): task_name -> {metric -> score}
+    """
+    assert isinstance(results, OrderedDict), results  # unordered results cannot be properly printed
+    logger = logging.getLogger(__name__)
+    for task, res in results.items():
+        # Don't print "AP-category" metrics since they are usually not tracked.
+        important_res = [(k, v) for k, v in res.items() if "-" not in k]
+        logger.info("copypaste: Task: {}".format(task))
+        logger.info("copypaste: " + ",".join([k[0] for k in important_res]))
+        logger.info("copypaste: " + ",".join(["{0:.4f}".format(k[1]) for k in important_res]))
+
+
+def verify_results(cfg, results):
+    """
+    Args:
+        results (OrderedDict[dict]): task_name -> {metric -> score}
+
+    Returns:
+        bool: whether the verification succeeds or not
+    """
+    expected_results = cfg.TEST.EXPECTED_RESULTS
+    if not len(expected_results):
+        return True
+
+    ok = True
+    for task, metric, expected, tolerance in expected_results:
+        actual = results[task][metric]
+        if not np.isfinite(actual):
+            ok = False
+        diff = abs(actual - expected)
+        if diff > tolerance:
+            ok = False
+
+    logger = logging.getLogger(__name__)
+    if not ok:
+        logger.error("Result verification failed!")
+        logger.error("Expected Results: " + str(expected_results))
+        logger.error("Actual Results: " + pprint.pformat(results))
+
+        sys.exit(1)
+    else:
+        logger.info("Results verification passed.")
+    return ok
+
+
+def flatten_results_dict(results):
+    """
+    Expand a hierarchical dict of scalars into a flat dict of scalars.
+    If results[k1][k2][k3] = v, the returned dict will have the entry
+    {"k1/k2/k3": v}.
+
+    Args:
+        results (dict):
+    """
+    r = {}
+    for k, v in results.items():
+        if isinstance(v, Mapping):
+            v = flatten_results_dict(v)
+            for kk, vv in v.items():
+                r[k + "/" + kk] = vv
+        else:
+            r[k] = v
+    return r

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/README.md

@@ -0,0 +1,10 @@
+
+This directory contains code to prepare a detectron2 model for deployment.
+Currently it supports exporting a detectron2 model to Caffe2 format through ONNX.
+
+Please see [documentation](https://detectron2.readthedocs.io/tutorials/deployment.html) for its usage.
+
+
+### Acknowledgements
+
+Thanks to Mobile Vision team at Facebook for developing the conversion tools.

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/__init__.py

@@ -0,0 +1,5 @@
+# -*- coding: utf-8 -*-
+
+from .api import *
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/api.py

@@ -0,0 +1,277 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+import copy
+import logging
+import os
+import torch
+from caffe2.proto import caffe2_pb2
+from torch import nn
+
+from detectron2.config import CfgNode as CN
+
+from .caffe2_export import export_caffe2_detection_model
+from .caffe2_export import export_onnx_model as export_onnx_model_impl
+from .caffe2_export import run_and_save_graph
+from .caffe2_inference import ProtobufDetectionModel
+from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format
+from .shared import get_pb_arg_vali, get_pb_arg_vals, save_graph
+
+__all__ = [
+    "add_export_config",
+    "export_caffe2_model",
+    "Caffe2Model",
+    "export_onnx_model",
+    "Caffe2Tracer",
+]
+
+
+def add_export_config(cfg):
+    """
+    Args:
+        cfg (CfgNode): a detectron2 config
+
+    Returns:
+        CfgNode: an updated config with new options that will be used
+            by :class:`Caffe2Tracer`.
+    """
+    is_frozen = cfg.is_frozen()
+    cfg.defrost()
+    cfg.EXPORT_CAFFE2 = CN()
+    cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT = False
+    if is_frozen:
+        cfg.freeze()
+    return cfg
+
+
+class Caffe2Tracer:
+    """
+    Make a detectron2 model traceable with caffe2 style.
+
+    An original detectron2 model may not be traceable, or
+    cannot be deployed directly after being traced, due to some reasons:
+    1. control flow in some ops
+    2. custom ops
+    3. complicated pre/post processing
+
+    This class provides a traceable version of a detectron2 model by:
+    1. Rewrite parts of the model using ops in caffe2. Note that some ops do
+       not have GPU implementation.
+    2. Define the inputs "after pre-processing" as inputs to the model
+    3. Remove post-processing and produce raw layer outputs
+
+    More specifically about inputs: all builtin models take two input tensors.
+    (1) NCHW float "data" which is an image (usually in [0, 255])
+    (2) Nx3 float "im_info", each row of which is (height, width, 1.0)
+
+    After making a traceable model, the class provide methods to export such a
+    model to different deployment formats.
+
+    The class currently only supports models using builtin meta architectures.
+    """
+
+    def __init__(self, cfg, model, inputs):
+        """
+        Args:
+            cfg (CfgNode): a detectron2 config, with extra export-related options
+                added by :func:`add_export_config`.
+            model (nn.Module): a model built by
+                :func:`detectron2.modeling.build_model`.
+            inputs: sample inputs that the given model takes for inference.
+                Will be used to trace the model.
+        """
+        assert isinstance(cfg, CN), cfg
+        assert isinstance(model, torch.nn.Module), type(model)
+        if "EXPORT_CAFFE2" not in cfg:
+            cfg = add_export_config(cfg)  # will just the defaults
+
+        self.cfg = cfg
+        self.model = model
+        self.inputs = inputs
+
+    def _get_traceable(self):
+        # TODO how to make it extensible to support custom models
+        C2MetaArch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[self.cfg.MODEL.META_ARCHITECTURE]
+        traceable_model = C2MetaArch(self.cfg, copy.deepcopy(self.model))
+        traceable_inputs = traceable_model.get_caffe2_inputs(self.inputs)
+        return traceable_model, traceable_inputs
+
+    def export_caffe2(self):
+        """
+        Export the model to Caffe2's protobuf format.
+        The returned object can be saved with `.save_protobuf()` method.
+        The result can be loaded and executed using Caffe2 runtime.
+
+        Returns:
+            Caffe2Model
+        """
+        model, inputs = self._get_traceable()
+        predict_net, init_net = export_caffe2_detection_model(model, inputs)
+        return Caffe2Model(predict_net, init_net)
+
+    def export_onnx(self):
+        """
+        Export the model to ONNX format.
+        Note that the exported model contains custom ops only available in caffe2, therefore it
+        cannot be directly executed by other runtime. Post-processing or transformation passes
+        may be applied on the model to accommodate different runtimes.
+
+        Returns:
+            onnx.ModelProto: an onnx model.
+        """
+        model, inputs = self._get_traceable()
+        return export_onnx_model_impl(model, (inputs,))
+
+    def export_torchscript(self):
+        """
+        Export the model to a `torch.jit.TracedModule` by tracing.
+        The returned object can be saved to a file by ".save()".
+
+        Returns:
+            torch.jit.TracedModule: a torch TracedModule
+        """
+        model, inputs = self._get_traceable()
+        logger = logging.getLogger(__name__)
+        logger.info("Tracing the model with torch.jit.trace ...")
+        with torch.no_grad():
+            return torch.jit.trace(model, (inputs,), optimize=True)
+
+
+def export_caffe2_model(cfg, model, inputs):
+    """
+    Export a detectron2 model to caffe2 format.
+
+    Args:
+        cfg (CfgNode): a detectron2 config, with extra export-related options
+            added by :func:`add_export_config`.
+        model (nn.Module): a model built by
+            :func:`detectron2.modeling.build_model`.
+            It will be modified by this function.
+        inputs: sample inputs that the given model takes for inference.
+            Will be used to trace the model.
+
+    Returns:
+        Caffe2Model
+    """
+    return Caffe2Tracer(cfg, model, inputs).export_caffe2()
+
+
+def export_onnx_model(cfg, model, inputs):
+    """
+    Export a detectron2 model to ONNX format.
+    Note that the exported model contains custom ops only available in caffe2, therefore it
+    cannot be directly executed by other runtime. Post-processing or transformation passes
+    may be applied on the model to accommodate different runtimes.
+    Args:
+        cfg (CfgNode): a detectron2 config, with extra export-related options
+            added by :func:`add_export_config`.
+        model (nn.Module): a model built by
+            :func:`detectron2.modeling.build_model`.
+            It will be modified by this function.
+        inputs: sample inputs that the given model takes for inference.
+            Will be used to trace the model.
+    Returns:
+        onnx.ModelProto: an onnx model.
+    """
+    return Caffe2Tracer(cfg, model, inputs).export_onnx()
+
+
+class Caffe2Model(nn.Module):
+    """
+    A wrapper around the traced model in caffe2's pb format.
+    """
+
+    def __init__(self, predict_net, init_net):
+        super().__init__()
+        self.eval()  # always in eval mode
+        self._predict_net = predict_net
+        self._init_net = init_net
+        self._predictor = None
+
+    @property
+    def predict_net(self):
+        """
+        Returns:
+            core.Net: the underlying caffe2 predict net
+        """
+        return self._predict_net
+
+    @property
+    def init_net(self):
+        """
+        Returns:
+            core.Net: the underlying caffe2 init net
+        """
+        return self._init_net
+
+    __init__.__HIDE_SPHINX_DOC__ = True
+
+    def save_protobuf(self, output_dir):
+        """
+        Save the model as caffe2's protobuf format.
+
+        Args:
+            output_dir (str): the output directory to save protobuf files.
+        """
+        logger = logging.getLogger(__name__)
+        logger.info("Saving model to {} ...".format(output_dir))
+        os.makedirs(output_dir, exist_ok=True)
+
+        with open(os.path.join(output_dir, "model.pb"), "wb") as f:
+            f.write(self._predict_net.SerializeToString())
+        with open(os.path.join(output_dir, "model.pbtxt"), "w") as f:
+            f.write(str(self._predict_net))
+        with open(os.path.join(output_dir, "model_init.pb"), "wb") as f:
+            f.write(self._init_net.SerializeToString())
+
+    def save_graph(self, output_file, inputs=None):
+        """
+        Save the graph as SVG format.
+
+        Args:
+            output_file (str): a SVG file
+            inputs: optional inputs given to the model.
+                If given, the inputs will be used to run the graph to record
+                shape of every tensor. The shape information will be
+                saved together with the graph.
+        """
+        if inputs is None:
+            save_graph(self._predict_net, output_file, op_only=False)
+        else:
+            size_divisibility = get_pb_arg_vali(self._predict_net, "size_divisibility", 0)
+            device = get_pb_arg_vals(self._predict_net, "device", b"cpu").decode("ascii")
+            inputs = convert_batched_inputs_to_c2_format(inputs, size_divisibility, device)
+            inputs = [x.cpu().numpy() for x in inputs]
+            run_and_save_graph(self._predict_net, self._init_net, inputs, output_file)
+
+    @staticmethod
+    def load_protobuf(dir):
+        """
+        Args:
+            dir (str): a directory used to save Caffe2Model with
+                :meth:`save_protobuf`.
+                The files "model.pb" and "model_init.pb" are needed.
+
+        Returns:
+            Caffe2Model: the caffe2 model loaded from this directory.
+        """
+        predict_net = caffe2_pb2.NetDef()
+        with open(os.path.join(dir, "model.pb"), "rb") as f:
+            predict_net.ParseFromString(f.read())
+
+        init_net = caffe2_pb2.NetDef()
+        with open(os.path.join(dir, "model_init.pb"), "rb") as f:
+            init_net.ParseFromString(f.read())
+
+        return Caffe2Model(predict_net, init_net)
+
+    def __call__(self, inputs):
+        """
+        An interface that wraps around a caffe2 model and mimics detectron2's models'
+        input & output format. This is used to compare the outputs of caffe2 model
+        with its original torch model.
+
+        Due to the extra conversion between torch/caffe2,
+        this method is not meant for benchmark.
+        """
+        if self._predictor is None:
+            self._predictor = ProtobufDetectionModel(self._predict_net, self._init_net)
+        return self._predictor(inputs)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/c10.py

@@ -0,0 +1,503 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import math
+import torch
+import torch.nn.functional as F
+
+from detectron2.layers import cat
+from detectron2.layers.roi_align_rotated import ROIAlignRotated
+from detectron2.modeling import poolers
+from detectron2.modeling.proposal_generator import rpn
+from detectron2.modeling.roi_heads.mask_head import mask_rcnn_inference
+from detectron2.structures import Boxes, ImageList, Instances, Keypoints
+
+from .shared import alias, to_device
+
+
+"""
+This file contains caffe2-compatible implementation of several detectrno2 components.
+"""
+
+
+class Caffe2Boxes(Boxes):
+    """
+    Representing a list of detectron2.structures.Boxes from minibatch, each box
+    is represented by a 5d vector (batch index + 4 coordinates), or a 6d vector
+    (batch index + 5 coordinates) for RotatedBoxes.
+    """
+
+    def __init__(self, tensor):
+        assert isinstance(tensor, torch.Tensor)
+        assert tensor.dim() == 2 and tensor.size(-1) in [4, 5, 6], tensor.size()
+        # TODO: make tensor immutable when dim is Nx5 for Boxes,
+        # and Nx6 for RotatedBoxes?
+        self.tensor = tensor
+
+
+# TODO clean up this class, maybe just extend Instances
+class InstancesList(object):
+    """
+    Tensor representation of a list of Instances object for a batch of images.
+
+    When dealing with a batch of images with Caffe2 ops, a list of bboxes
+    (instances) are usually represented by single Tensor with size
+    (sigma(Ni), 5) or (sigma(Ni), 4) plus a batch split Tensor. This class is
+    for providing common functions to convert between these two representations.
+    """
+
+    def __init__(self, im_info, indices, extra_fields=None):
+        # [N, 3] -> (H, W, Scale)
+        self.im_info = im_info
+        # [N,] -> indice of batch to which the instance belongs
+        self.indices = indices
+        # [N, ...]
+        self.batch_extra_fields = extra_fields or {}
+
+        self.image_size = self.im_info
+
+    def get_fields(self):
+        """ like `get_fields` in the Instances object,
+        but return each field in tensor representations """
+        ret = {}
+        for k, v in self.batch_extra_fields.items():
+            # if isinstance(v, torch.Tensor):
+            #     tensor_rep = v
+            # elif isinstance(v, (Boxes, Keypoints)):
+            #     tensor_rep = v.tensor
+            # else:
+            #     raise ValueError("Can't find tensor representation for: {}".format())
+            ret[k] = v
+        return ret
+
+    def has(self, name):
+        return name in self.batch_extra_fields
+
+    def set(self, name, value):
+        data_len = len(value)
+        if len(self.batch_extra_fields):
+            assert (
+                len(self) == data_len
+            ), "Adding a field of length {} to a Instances of length {}".format(data_len, len(self))
+        self.batch_extra_fields[name] = value
+
+    def __setattr__(self, name, val):
+        if name in ["im_info", "indices", "batch_extra_fields", "image_size"]:
+            super().__setattr__(name, val)
+        else:
+            self.set(name, val)
+
+    def __getattr__(self, name):
+        if name not in self.batch_extra_fields:
+            raise AttributeError("Cannot find field '{}' in the given Instances!".format(name))
+        return self.batch_extra_fields[name]
+
+    def __len__(self):
+        return len(self.indices)
+
+    def flatten(self):
+        ret = []
+        for _, v in self.batch_extra_fields.items():
+            if isinstance(v, (Boxes, Keypoints)):
+                ret.append(v.tensor)
+            else:
+                ret.append(v)
+        return ret
+
+    @staticmethod
+    def to_d2_instances_list(instances_list):
+        """
+        Convert InstancesList to List[Instances]. The input `instances_list` can
+        also be a List[Instances], in this case this method is a non-op.
+        """
+        if not isinstance(instances_list, InstancesList):
+            assert all(isinstance(x, Instances) for x in instances_list)
+            return instances_list
+
+        ret = []
+        for i, info in enumerate(instances_list.im_info):
+            instances = Instances(torch.Size([int(info[0].item()), int(info[1].item())]))
+
+            ids = instances_list.indices == i
+            for k, v in instances_list.batch_extra_fields.items():
+                if isinstance(v, torch.Tensor):
+                    instances.set(k, v[ids])
+                    continue
+                elif isinstance(v, Boxes):
+                    instances.set(k, v[ids, -4:])
+                    continue
+
+                target_type, tensor_source = v
+                assert isinstance(tensor_source, torch.Tensor)
+                assert tensor_source.shape[0] == instances_list.indices.shape[0]
+                tensor_source = tensor_source[ids]
+
+                if issubclass(target_type, Boxes):
+                    instances.set(k, Boxes(tensor_source[:, -4:]))
+                elif issubclass(target_type, Keypoints):
+                    instances.set(k, Keypoints(tensor_source))
+                elif issubclass(target_type, torch.Tensor):
+                    instances.set(k, tensor_source)
+                else:
+                    raise ValueError("Can't handle targe type: {}".format(target_type))
+
+            ret.append(instances)
+        return ret
+
+
+class Caffe2Compatible(object):
+    def _get_tensor_mode(self):
+        return self._tensor_mode
+
+    def _set_tensor_mode(self, v):
+        self._tensor_mode = v
+
+    tensor_mode = property(_get_tensor_mode, _set_tensor_mode)
+    """
+    If true, the model expects C2-style tensor only inputs/outputs format.
+    """
+
+
+class Caffe2RPN(Caffe2Compatible, rpn.RPN):
+    def forward(self, images, features, gt_instances=None):
+        assert not self.training
+
+        features = [features[f] for f in self.in_features]
+        objectness_logits_pred, anchor_deltas_pred = self.rpn_head(features)
+
+        assert isinstance(images, ImageList)
+        if self.tensor_mode:
+            im_info = images.image_sizes
+        else:
+            im_info = torch.Tensor(
+                [[im_sz[0], im_sz[1], torch.Tensor([1.0])] for im_sz in images.image_sizes]
+            ).to(images.tensor.device)
+        assert isinstance(im_info, torch.Tensor)
+
+        rpn_rois_list = []
+        rpn_roi_probs_list = []
+        for scores, bbox_deltas, cell_anchors_tensor, feat_stride in zip(
+            objectness_logits_pred,
+            anchor_deltas_pred,
+            iter(self.anchor_generator.cell_anchors),
+            self.anchor_generator.strides,
+        ):
+            scores = scores.detach()
+            bbox_deltas = bbox_deltas.detach()
+
+            rpn_rois, rpn_roi_probs = torch.ops._caffe2.GenerateProposals(
+                scores,
+                bbox_deltas,
+                im_info,
+                cell_anchors_tensor,
+                spatial_scale=1.0 / feat_stride,
+                pre_nms_topN=self.pre_nms_topk[self.training],
+                post_nms_topN=self.post_nms_topk[self.training],
+                nms_thresh=self.nms_thresh,
+                min_size=self.min_box_side_len,
+                # correct_transform_coords=True,  # deprecated argument
+                angle_bound_on=True,  # Default
+                angle_bound_lo=-180,
+                angle_bound_hi=180,
+                clip_angle_thresh=1.0,  # Default
+                legacy_plus_one=False,
+            )
+            rpn_rois_list.append(rpn_rois)
+            rpn_roi_probs_list.append(rpn_roi_probs)
+
+        # For FPN in D2, in RPN all proposals from different levels are concated
+        # together, ranked and picked by top post_nms_topk. Then in ROIPooler
+        # it calculates level_assignments and calls the RoIAlign from
+        # the corresponding level.
+
+        if len(objectness_logits_pred) == 1:
+            rpn_rois = rpn_rois_list[0]
+            rpn_roi_probs = rpn_roi_probs_list[0]
+        else:
+            assert len(rpn_rois_list) == len(rpn_roi_probs_list)
+            rpn_post_nms_topN = self.post_nms_topk[self.training]
+
+            device = rpn_rois_list[0].device
+            input_list = [to_device(x, "cpu") for x in (rpn_rois_list + rpn_roi_probs_list)]
+
+            # TODO remove this after confirming rpn_max_level/rpn_min_level
+            # is not needed in CollectRpnProposals.
+            feature_strides = list(self.anchor_generator.strides)
+            rpn_min_level = int(math.log2(feature_strides[0]))
+            rpn_max_level = int(math.log2(feature_strides[-1]))
+            assert (rpn_max_level - rpn_min_level + 1) == len(
+                rpn_rois_list
+            ), "CollectRpnProposals requires continuous levels"
+
+            rpn_rois = torch.ops._caffe2.CollectRpnProposals(
+                input_list,
+                # NOTE: in current implementation, rpn_max_level and rpn_min_level
+                # are not needed, only the subtraction of two matters and it
+                # can be infer from the number of inputs. Keep them now for
+                # consistency.
+                rpn_max_level=2 + len(rpn_rois_list) - 1,
+                rpn_min_level=2,
+                rpn_post_nms_topN=rpn_post_nms_topN,
+            )
+            rpn_rois = to_device(rpn_rois, device)
+            rpn_roi_probs = []
+
+        proposals = self.c2_postprocess(im_info, rpn_rois, rpn_roi_probs, self.tensor_mode)
+        return proposals, {}
+
+    @staticmethod
+    def c2_postprocess(im_info, rpn_rois, rpn_roi_probs, tensor_mode):
+        proposals = InstancesList(
+            im_info=im_info,
+            indices=rpn_rois[:, 0],
+            extra_fields={
+                "proposal_boxes": Caffe2Boxes(rpn_rois),
+                "objectness_logits": (torch.Tensor, rpn_roi_probs),
+            },
+        )
+        if not tensor_mode:
+            proposals = InstancesList.to_d2_instances_list(proposals)
+        else:
+            proposals = [proposals]
+        return proposals
+
+
+class Caffe2ROIPooler(Caffe2Compatible, poolers.ROIPooler):
+    @staticmethod
+    def c2_preprocess(box_lists):
+        assert all(isinstance(x, Boxes) for x in box_lists)
+        if all(isinstance(x, Caffe2Boxes) for x in box_lists):
+            # input is pure-tensor based
+            assert len(box_lists) == 1
+            pooler_fmt_boxes = box_lists[0].tensor
+        else:
+            pooler_fmt_boxes = poolers.convert_boxes_to_pooler_format(box_lists)
+        return pooler_fmt_boxes
+
+    def forward(self, x, box_lists):
+        assert not self.training
+
+        pooler_fmt_boxes = self.c2_preprocess(box_lists)
+        num_level_assignments = len(self.level_poolers)
+
+        if num_level_assignments == 1:
+            if isinstance(self.level_poolers[0], ROIAlignRotated):
+                c2_roi_align = torch.ops._caffe2.RoIAlignRotated
+                aligned = True
+            else:
+                c2_roi_align = torch.ops._caffe2.RoIAlign
+                aligned = self.level_poolers[0].aligned
+
+            out = c2_roi_align(
+                x[0],
+                pooler_fmt_boxes,
+                order="NCHW",
+                spatial_scale=float(self.level_poolers[0].spatial_scale),
+                pooled_h=int(self.output_size[0]),
+                pooled_w=int(self.output_size[1]),
+                sampling_ratio=int(self.level_poolers[0].sampling_ratio),
+                aligned=aligned,
+            )
+            return out
+
+        device = pooler_fmt_boxes.device
+        assert (
+            self.max_level - self.min_level + 1 == 4
+        ), "Currently DistributeFpnProposals only support 4 levels"
+        fpn_outputs = torch.ops._caffe2.DistributeFpnProposals(
+            to_device(pooler_fmt_boxes, "cpu"),
+            roi_canonical_scale=self.canonical_box_size,
+            roi_canonical_level=self.canonical_level,
+            roi_max_level=self.max_level,
+            roi_min_level=self.min_level,
+            legacy_plus_one=False,
+        )
+        fpn_outputs = [to_device(x, device) for x in fpn_outputs]
+
+        rois_fpn_list = fpn_outputs[:-1]
+        rois_idx_restore_int32 = fpn_outputs[-1]
+
+        roi_feat_fpn_list = []
+        for roi_fpn, x_level, pooler in zip(rois_fpn_list, x, self.level_poolers):
+            if isinstance(pooler, ROIAlignRotated):
+                c2_roi_align = torch.ops._caffe2.RoIAlignRotated
+                aligned = True
+            else:
+                c2_roi_align = torch.ops._caffe2.RoIAlign
+                aligned = bool(pooler.aligned)
+
+            roi_feat_fpn = c2_roi_align(
+                x_level,
+                roi_fpn,
+                order="NCHW",
+                spatial_scale=float(pooler.spatial_scale),
+                pooled_h=int(self.output_size[0]),
+                pooled_w=int(self.output_size[1]),
+                sampling_ratio=int(pooler.sampling_ratio),
+                aligned=aligned,
+            )
+            roi_feat_fpn_list.append(roi_feat_fpn)
+
+        roi_feat_shuffled = cat(roi_feat_fpn_list, dim=0)
+        roi_feat = torch.ops._caffe2.BatchPermutation(roi_feat_shuffled, rois_idx_restore_int32)
+        return roi_feat
+
+
+class Caffe2FastRCNNOutputsInference:
+    def __init__(self, tensor_mode):
+        self.tensor_mode = tensor_mode  # whether the output is caffe2 tensor mode
+
+    def __call__(self, box_predictor, predictions, proposals):
+        """ equivalent to FastRCNNOutputLayers.inference """
+        score_thresh = box_predictor.test_score_thresh
+        nms_thresh = box_predictor.test_nms_thresh
+        topk_per_image = box_predictor.test_topk_per_image
+        is_rotated = len(box_predictor.box2box_transform.weights) == 5
+
+        if is_rotated:
+            box_dim = 5
+            assert box_predictor.box2box_transform.weights[4] == 1, (
+                "The weights for Rotated BBoxTransform in C2 have only 4 dimensions,"
+                + " thus enforcing the angle weight to be 1 for now"
+            )
+            box2box_transform_weights = box_predictor.box2box_transform.weights[:4]
+        else:
+            box_dim = 4
+            box2box_transform_weights = box_predictor.box2box_transform.weights
+
+        class_logits, box_regression = predictions
+        class_prob = F.softmax(class_logits, -1)
+
+        assert box_regression.shape[1] % box_dim == 0
+        cls_agnostic_bbox_reg = box_regression.shape[1] // box_dim == 1
+
+        input_tensor_mode = proposals[0].proposal_boxes.tensor.shape[1] == box_dim + 1
+
+        rois = type(proposals[0].proposal_boxes).cat([p.proposal_boxes for p in proposals])
+        device, dtype = rois.tensor.device, rois.tensor.dtype
+        if input_tensor_mode:
+            im_info = proposals[0].image_size
+            rois = rois.tensor
+        else:
+            im_info = torch.Tensor(
+                [[sz[0], sz[1], 1.0] for sz in [x.image_size for x in proposals]]
+            )
+            batch_ids = cat(
+                [
+                    torch.full((b, 1), i, dtype=dtype, device=device)
+                    for i, b in enumerate(len(p) for p in proposals)
+                ],
+                dim=0,
+            )
+            rois = torch.cat([batch_ids, rois.tensor], dim=1)
+
+        roi_pred_bbox, roi_batch_splits = torch.ops._caffe2.BBoxTransform(
+            to_device(rois, "cpu"),
+            to_device(box_regression, "cpu"),
+            to_device(im_info, "cpu"),
+            weights=box2box_transform_weights,
+            apply_scale=True,
+            rotated=is_rotated,
+            angle_bound_on=True,
+            angle_bound_lo=-180,
+            angle_bound_hi=180,
+            clip_angle_thresh=1.0,
+            legacy_plus_one=False,
+        )
+        roi_pred_bbox = to_device(roi_pred_bbox, device)
+        roi_batch_splits = to_device(roi_batch_splits, device)
+
+        nms_outputs = torch.ops._caffe2.BoxWithNMSLimit(
+            to_device(class_prob, "cpu"),
+            to_device(roi_pred_bbox, "cpu"),
+            to_device(roi_batch_splits, "cpu"),
+            score_thresh=float(score_thresh),
+            nms=float(nms_thresh),
+            detections_per_im=int(topk_per_image),
+            soft_nms_enabled=False,
+            soft_nms_method="linear",
+            soft_nms_sigma=0.5,
+            soft_nms_min_score_thres=0.001,
+            rotated=is_rotated,
+            cls_agnostic_bbox_reg=cls_agnostic_bbox_reg,
+            input_boxes_include_bg_cls=False,
+            output_classes_include_bg_cls=False,
+            legacy_plus_one=False,
+        )
+        roi_score_nms = to_device(nms_outputs[0], device)
+        roi_bbox_nms = to_device(nms_outputs[1], device)
+        roi_class_nms = to_device(nms_outputs[2], device)
+        roi_batch_splits_nms = to_device(nms_outputs[3], device)
+        roi_keeps_nms = to_device(nms_outputs[4], device)
+        roi_keeps_size_nms = to_device(nms_outputs[5], device)
+        if not self.tensor_mode:
+            roi_class_nms = roi_class_nms.to(torch.int64)
+
+        roi_batch_ids = cat(
+            [
+                torch.full((b, 1), i, dtype=dtype, device=device)
+                for i, b in enumerate(int(x.item()) for x in roi_batch_splits_nms)
+            ],
+            dim=0,
+        )
+
+        roi_class_nms = alias(roi_class_nms, "class_nms")
+        roi_score_nms = alias(roi_score_nms, "score_nms")
+        roi_bbox_nms = alias(roi_bbox_nms, "bbox_nms")
+        roi_batch_splits_nms = alias(roi_batch_splits_nms, "batch_splits_nms")
+        roi_keeps_nms = alias(roi_keeps_nms, "keeps_nms")
+        roi_keeps_size_nms = alias(roi_keeps_size_nms, "keeps_size_nms")
+
+        results = InstancesList(
+            im_info=im_info,
+            indices=roi_batch_ids[:, 0],
+            extra_fields={
+                "pred_boxes": Caffe2Boxes(roi_bbox_nms),
+                "scores": roi_score_nms,
+                "pred_classes": roi_class_nms,
+            },
+        )
+
+        if not self.tensor_mode:
+            results = InstancesList.to_d2_instances_list(results)
+            batch_splits = roi_batch_splits_nms.int().tolist()
+            kept_indices = list(roi_keeps_nms.to(torch.int64).split(batch_splits))
+        else:
+            results = [results]
+            kept_indices = [roi_keeps_nms]
+
+        return results, kept_indices
+
+
+class Caffe2MaskRCNNInference:
+    def __call__(self, pred_mask_logits, pred_instances):
+        """ equivalent to mask_head.mask_rcnn_inference """
+        if all(isinstance(x, InstancesList) for x in pred_instances):
+            assert len(pred_instances) == 1
+            mask_probs_pred = pred_mask_logits.sigmoid()
+            mask_probs_pred = alias(mask_probs_pred, "mask_fcn_probs")
+            pred_instances[0].pred_masks = mask_probs_pred
+        else:
+            mask_rcnn_inference(pred_mask_logits, pred_instances)
+
+
+class Caffe2KeypointRCNNInference:
+    def __init__(self, use_heatmap_max_keypoint):
+        self.use_heatmap_max_keypoint = use_heatmap_max_keypoint
+
+    def __call__(self, pred_keypoint_logits, pred_instances):
+        # just return the keypoint heatmap for now,
+        # there will be option to call HeatmapMaxKeypointOp
+        output = alias(pred_keypoint_logits, "kps_score")
+        if all(isinstance(x, InstancesList) for x in pred_instances):
+            assert len(pred_instances) == 1
+            if self.use_heatmap_max_keypoint:
+                device = output.device
+                output = torch.ops._caffe2.HeatmapMaxKeypoint(
+                    to_device(output, "cpu"),
+                    pred_instances[0].pred_boxes.tensor,
+                    should_output_softmax=True,  # worth make it configerable?
+                )
+                output = to_device(output, device)
+                output = alias(output, "keypoints_out")
+            pred_instances[0].pred_keypoints = output
+        return pred_keypoint_logits

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/caffe2_export.py

@@ -0,0 +1,204 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import copy
+import io
+import logging
+import numpy as np
+from typing import List
+import onnx
+import torch
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core
+from caffe2.python.onnx.backend import Caffe2Backend
+from tabulate import tabulate
+from termcolor import colored
+from torch.onnx import OperatorExportTypes
+
+from .shared import (
+    ScopedWS,
+    construct_init_net_from_params,
+    fuse_alias_placeholder,
+    fuse_copy_between_cpu_and_gpu,
+    get_params_from_init_net,
+    group_norm_replace_aten_with_caffe2,
+    infer_device_type,
+    remove_dead_end_ops,
+    remove_reshape_for_fc,
+    save_graph,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def export_onnx_model(model, inputs):
+    """
+    Trace and export a model to onnx format.
+
+    Args:
+        model (nn.Module):
+        inputs (tuple[args]): the model will be called by `model(*inputs)`
+
+    Returns:
+        an onnx model
+    """
+    assert isinstance(model, torch.nn.Module)
+
+    # make sure all modules are in eval mode, onnx may change the training state
+    # of the module if the states are not consistent
+    def _check_eval(module):
+        assert not module.training
+
+    model.apply(_check_eval)
+
+    # Export the model to ONNX
+    with torch.no_grad():
+        with io.BytesIO() as f:
+            torch.onnx.export(
+                model,
+                inputs,
+                f,
+                operator_export_type=OperatorExportTypes.ONNX_ATEN_FALLBACK,
+                # verbose=True,  # NOTE: uncomment this for debugging
+                # export_params=True,
+            )
+            onnx_model = onnx.load_from_string(f.getvalue())
+
+    # Apply ONNX's Optimization
+    all_passes = onnx.optimizer.get_available_passes()
+    passes = ["fuse_bn_into_conv"]
+    assert all(p in all_passes for p in passes)
+    onnx_model = onnx.optimizer.optimize(onnx_model, passes)
+    return onnx_model
+
+
+def _op_stats(net_def):
+    type_count = {}
+    for t in [op.type for op in net_def.op]:
+        type_count[t] = type_count.get(t, 0) + 1
+    type_count_list = sorted(type_count.items(), key=lambda kv: kv[0])  # alphabet
+    type_count_list = sorted(type_count_list, key=lambda kv: -kv[1])  # count
+    return "\n".join("{:>4}x {}".format(count, name) for name, count in type_count_list)
+
+
+def _assign_device_option(
+    predict_net: caffe2_pb2.NetDef, init_net: caffe2_pb2.NetDef, tensor_inputs: List[torch.Tensor]
+):
+    """
+    ONNX exported network doesn't have concept of device, assign necessary
+    device option for each op in order to make it runable on GPU runtime.
+    """
+
+    def _get_device_type(torch_tensor):
+        assert torch_tensor.device.type in ["cpu", "cuda"]
+        assert torch_tensor.device.index == 0
+        return torch_tensor.device.type
+
+    def _assign_op_device_option(net_proto, net_ssa, blob_device_types):
+        for op, ssa_i in zip(net_proto.op, net_ssa):
+            if op.type in ["CopyCPUToGPU", "CopyGPUToCPU"]:
+                op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0))
+            else:
+                devices = [blob_device_types[b] for b in ssa_i[0] + ssa_i[1]]
+                assert all(d == devices[0] for d in devices)
+                if devices[0] == "cuda":
+                    op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0))
+
+    # update ops in predict_net
+    predict_net_input_device_types = {
+        (name, 0): _get_device_type(tensor)
+        for name, tensor in zip(predict_net.external_input, tensor_inputs)
+    }
+    predict_net_device_types = infer_device_type(
+        predict_net, known_status=predict_net_input_device_types, device_name_style="pytorch"
+    )
+    predict_net_ssa, _ = core.get_ssa(predict_net)
+    _assign_op_device_option(predict_net, predict_net_ssa, predict_net_device_types)
+
+    # update ops in init_net
+    init_net_ssa, versions = core.get_ssa(init_net)
+    init_net_output_device_types = {
+        (name, versions[name]): predict_net_device_types[(name, 0)]
+        for name in init_net.external_output
+    }
+    init_net_device_types = infer_device_type(
+        init_net, known_status=init_net_output_device_types, device_name_style="pytorch"
+    )
+    _assign_op_device_option(init_net, init_net_ssa, init_net_device_types)
+
+
+def export_caffe2_detection_model(model: torch.nn.Module, tensor_inputs: List[torch.Tensor]):
+    """
+    Export a caffe2-compatible Detectron2 model to caffe2 format via ONNX.
+
+    Arg:
+        model: a caffe2-compatible version of detectron2 model, defined in caffe2_modeling.py
+        tensor_inputs: a list of tensors that caffe2 model takes as input.
+    """
+    model = copy.deepcopy(model)
+    assert isinstance(model, torch.nn.Module)
+    assert hasattr(model, "encode_additional_info")
+
+    # Export via ONNX
+    logger.info("Exporting a {} model via ONNX ...".format(type(model).__name__))
+    onnx_model = export_onnx_model(model, (tensor_inputs,))
+    # Convert ONNX model to Caffe2 protobuf
+    init_net, predict_net = Caffe2Backend.onnx_graph_to_caffe2_net(onnx_model)
+    ops_table = [[op.type, op.input, op.output] for op in predict_net.op]
+    table = tabulate(ops_table, headers=["type", "input", "output"], tablefmt="pipe")
+    logger.info(
+        "ONNX export Done. Exported predict_net (before optimizations):\n" + colored(table, "cyan")
+    )
+
+    # Apply protobuf optimization
+    fuse_alias_placeholder(predict_net, init_net)
+    if any(t.device.type != "cpu" for t in tensor_inputs):
+        fuse_copy_between_cpu_and_gpu(predict_net)
+        remove_dead_end_ops(init_net)
+        _assign_device_option(predict_net, init_net, tensor_inputs)
+    params, device_options = get_params_from_init_net(init_net)
+    predict_net, params = remove_reshape_for_fc(predict_net, params)
+    init_net = construct_init_net_from_params(params, device_options)
+    group_norm_replace_aten_with_caffe2(predict_net)
+
+    # Record necessary information for running the pb model in Detectron2 system.
+    model.encode_additional_info(predict_net, init_net)
+
+    logger.info("Operators used in predict_net: \n{}".format(_op_stats(predict_net)))
+    logger.info("Operators used in init_net: \n{}".format(_op_stats(init_net)))
+
+    return predict_net, init_net
+
+
+def run_and_save_graph(predict_net, init_net, tensor_inputs, graph_save_path):
+    """
+    Run the caffe2 model on given inputs, recording the shape and draw the graph.
+
+    predict_net/init_net: caffe2 model.
+    tensor_inputs: a list of tensors that caffe2 model takes as input.
+    graph_save_path: path for saving graph of exported model.
+    """
+
+    logger.info("Saving graph of ONNX exported model to {} ...".format(graph_save_path))
+    save_graph(predict_net, graph_save_path, op_only=False)
+
+    # Run the exported Caffe2 net
+    logger.info("Running ONNX exported model ...")
+    with ScopedWS("__ws_tmp__", True) as ws:
+        ws.RunNetOnce(init_net)
+        initialized_blobs = set(ws.Blobs())
+        uninitialized = [inp for inp in predict_net.external_input if inp not in initialized_blobs]
+        for name, blob in zip(uninitialized, tensor_inputs):
+            ws.FeedBlob(name, blob)
+
+        try:
+            ws.RunNetOnce(predict_net)
+        except RuntimeError as e:
+            logger.warning("Encountered RuntimeError: \n{}".format(str(e)))
+
+        ws_blobs = {b: ws.FetchBlob(b) for b in ws.Blobs()}
+        blob_sizes = {b: ws_blobs[b].shape for b in ws_blobs if isinstance(ws_blobs[b], np.ndarray)}
+
+        logger.info("Saving graph with blob shapes to {} ...".format(graph_save_path))
+        save_graph(predict_net, graph_save_path, op_only=False, blob_sizes=blob_sizes)
+
+        return ws_blobs

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/caffe2_inference.py

@@ -0,0 +1,136 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import collections
+import logging
+import numpy as np
+import torch
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core
+
+from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format
+from .shared import ScopedWS, get_pb_arg_vali, get_pb_arg_vals, infer_device_type
+
+logger = logging.getLogger(__name__)
+
+
+class ProtobufModel(torch.nn.Module):
+    """
+    A class works just like nn.Module in terms of inference, but running
+    caffe2 model under the hood. Input/Output are Dict[str, tensor] whose keys
+    are in external_input/output.
+    """
+
+    def __init__(self, predict_net, init_net):
+        logger.info("Initializing ProtobufModel ...")
+        super().__init__()
+        assert isinstance(predict_net, caffe2_pb2.NetDef)
+        assert isinstance(init_net, caffe2_pb2.NetDef)
+        self.ws_name = "__ws_tmp__"
+        self.net = core.Net(predict_net)
+
+        with ScopedWS(self.ws_name, is_reset=True, is_cleanup=False) as ws:
+            ws.RunNetOnce(init_net)
+            for blob in self.net.Proto().external_input:
+                if blob not in ws.Blobs():
+                    ws.CreateBlob(blob)
+            ws.CreateNet(self.net)
+
+        self._error_msgs = set()
+
+    def forward(self, inputs_dict):
+        assert all(inp in self.net.Proto().external_input for inp in inputs_dict)
+        with ScopedWS(self.ws_name, is_reset=False, is_cleanup=False) as ws:
+            for b, tensor in inputs_dict.items():
+                ws.FeedBlob(b, tensor)
+            try:
+                ws.RunNet(self.net.Proto().name)
+            except RuntimeError as e:
+                if not str(e) in self._error_msgs:
+                    self._error_msgs.add(str(e))
+                    logger.warning("Encountered new RuntimeError: \n{}".format(str(e)))
+                logger.warning("Catch the error and use partial results.")
+
+            outputs_dict = collections.OrderedDict(
+                [(b, ws.FetchBlob(b)) for b in self.net.Proto().external_output]
+            )
+            # Remove outputs of current run, this is necessary in order to
+            # prevent fetching the result from previous run if the model fails
+            # in the middle.
+            for b in self.net.Proto().external_output:
+                # Needs to create uninitialized blob to make the net runable.
+                # This is "equivalent" to: ws.RemoveBlob(b) then ws.CreateBlob(b),
+                # but there'no such API.
+                ws.FeedBlob(b, "{}, a C++ native class of type nullptr (uninitialized).".format(b))
+
+        return outputs_dict
+
+
+class ProtobufDetectionModel(torch.nn.Module):
+    """
+    A class works just like a pytorch meta arch in terms of inference, but running
+    caffe2 model under the hood.
+    """
+
+    def __init__(self, predict_net, init_net, *, convert_outputs=None):
+        """
+        Args:
+            predict_net, init_net (core.Net): caffe2 nets
+            convert_outptus (callable): a function that converts caffe2
+                outputs to the same format of the original pytorch model.
+                By default, use the one defined in the caffe2 meta_arch.
+        """
+        super().__init__()
+        self.protobuf_model = ProtobufModel(predict_net, init_net)
+        self.size_divisibility = get_pb_arg_vali(predict_net, "size_divisibility", 0)
+        self.device = get_pb_arg_vals(predict_net, "device", b"cpu").decode("ascii")
+
+        if convert_outputs is None:
+            meta_arch = get_pb_arg_vals(predict_net, "meta_architecture", b"GeneralizedRCNN")
+            meta_arch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[meta_arch.decode("ascii")]
+            self._convert_outputs = meta_arch.get_outputs_converter(predict_net, init_net)
+        else:
+            self._convert_outputs = convert_outputs
+
+    def _infer_output_devices(self, inputs_dict):
+        def _get_device_type(torch_tensor):
+            assert torch_tensor.device.type in ["cpu", "cuda"]
+            assert torch_tensor.device.index == 0
+            return torch_tensor.device.type
+
+        predict_net = self.protobuf_model.net.Proto()
+        input_device_types = {
+            (name, 0): _get_device_type(tensor) for name, tensor in inputs_dict.items()
+        }
+        device_type_map = infer_device_type(
+            predict_net, known_status=input_device_types, device_name_style="pytorch"
+        )
+        ssa, versions = core.get_ssa(predict_net)
+        versioned_outputs = [(name, versions[name]) for name in predict_net.external_output]
+        output_devices = [device_type_map[outp] for outp in versioned_outputs]
+        return output_devices
+
+    def _convert_inputs(self, batched_inputs):
+        # currently all models convert inputs in the same way
+        data, im_info = convert_batched_inputs_to_c2_format(
+            batched_inputs, self.size_divisibility, self.device
+        )
+        return {"data": data, "im_info": im_info}
+
+    def forward(self, batched_inputs):
+        c2_inputs = self._convert_inputs(batched_inputs)
+        c2_results = self.protobuf_model(c2_inputs)
+
+        if any(t.device.type != "cpu" for _, t in c2_inputs.items()):
+            output_devices = self._infer_output_devices(c2_inputs)
+        else:
+            output_devices = ["cpu" for _ in self.protobuf_model.net.Proto().external_output]
+
+        def _cast_caffe2_blob_to_torch_tensor(blob, device):
+            return torch.Tensor(blob).to(device) if isinstance(blob, np.ndarray) else None
+
+        c2_results = {
+            name: _cast_caffe2_blob_to_torch_tensor(c2_results[name], device)
+            for name, device in zip(self.protobuf_model.net.Proto().external_output, output_devices)
+        }
+
+        return self._convert_outputs(batched_inputs, c2_inputs, c2_results)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/caffe2_modeling.py

@@ -0,0 +1,493 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import functools
+import io
+import struct
+import types
+import torch
+
+from detectron2.modeling import meta_arch
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.meta_arch.panoptic_fpn import combine_semantic_and_instance_outputs
+from detectron2.modeling.postprocessing import detector_postprocess, sem_seg_postprocess
+from detectron2.modeling.roi_heads import keypoint_head
+from detectron2.structures import Boxes, ImageList, Instances, RotatedBoxes
+
+from .c10 import Caffe2Compatible
+from .patcher import ROIHeadsPatcher, patch_generalized_rcnn
+from .shared import (
+    alias,
+    check_set_pb_arg,
+    get_pb_arg_floats,
+    get_pb_arg_valf,
+    get_pb_arg_vali,
+    get_pb_arg_vals,
+    mock_torch_nn_functional_interpolate,
+)
+
+
+def assemble_rcnn_outputs_by_name(image_sizes, tensor_outputs, force_mask_on=False):
+    """
+    A function to assemble caffe2 model's outputs (i.e. Dict[str, Tensor])
+    to detectron2's format (i.e. list of Instances instance).
+    This only works when the model follows the Caffe2 detectron's naming convention.
+
+    Args:
+        image_sizes (List[List[int, int]]): [H, W] of every image.
+        tensor_outputs (Dict[str, Tensor]): external_output to its tensor.
+
+        force_mask_on (Bool): if true, the it make sure there'll be pred_masks even
+            if the mask is not found from tensor_outputs (usually due to model crash)
+    """
+
+    results = [Instances(image_size) for image_size in image_sizes]
+
+    batch_splits = tensor_outputs.get("batch_splits", None)
+    if batch_splits:
+        raise NotImplementedError()
+    assert len(image_sizes) == 1
+    result = results[0]
+
+    bbox_nms = tensor_outputs["bbox_nms"]
+    score_nms = tensor_outputs["score_nms"]
+    class_nms = tensor_outputs["class_nms"]
+    # Detection will always success because Conv support 0-batch
+    assert bbox_nms is not None
+    assert score_nms is not None
+    assert class_nms is not None
+    if bbox_nms.shape[1] == 5:
+        result.pred_boxes = RotatedBoxes(bbox_nms)
+    else:
+        result.pred_boxes = Boxes(bbox_nms)
+    result.scores = score_nms
+    result.pred_classes = class_nms.to(torch.int64)
+
+    mask_fcn_probs = tensor_outputs.get("mask_fcn_probs", None)
+    if mask_fcn_probs is not None:
+        # finish the mask pred
+        mask_probs_pred = mask_fcn_probs
+        num_masks = mask_probs_pred.shape[0]
+        class_pred = result.pred_classes
+        indices = torch.arange(num_masks, device=class_pred.device)
+        mask_probs_pred = mask_probs_pred[indices, class_pred][:, None]
+        result.pred_masks = mask_probs_pred
+    elif force_mask_on:
+        # NOTE: there's no way to know the height/width of mask here, it won't be
+        # used anyway when batch size is 0, so just set them to 0.
+        result.pred_masks = torch.zeros([0, 1, 0, 0], dtype=torch.uint8)
+
+    keypoints_out = tensor_outputs.get("keypoints_out", None)
+    kps_score = tensor_outputs.get("kps_score", None)
+    if keypoints_out is not None:
+        # keypoints_out: [N, 4, #kypoints], where 4 is in order of (x, y, score, prob)
+        keypoints_tensor = keypoints_out
+        # NOTE: it's possible that prob is not calculated if "should_output_softmax"
+        # is set to False in HeatmapMaxKeypoint, so just using raw score, seems
+        # it doesn't affect mAP. TODO: check more carefully.
+        keypoint_xyp = keypoints_tensor.transpose(1, 2)[:, :, [0, 1, 2]]
+        result.pred_keypoints = keypoint_xyp
+    elif kps_score is not None:
+        # keypoint heatmap to sparse data structure
+        pred_keypoint_logits = kps_score
+        keypoint_head.keypoint_rcnn_inference(pred_keypoint_logits, [result])
+
+    return results
+
+
+def _cast_to_f32(f64):
+    return struct.unpack("f", struct.pack("f", f64))[0]
+
+
+def set_caffe2_compatible_tensor_mode(model, enable=True):
+    def _fn(m):
+        if isinstance(m, Caffe2Compatible):
+            m.tensor_mode = enable
+
+    model.apply(_fn)
+
+
+def convert_batched_inputs_to_c2_format(batched_inputs, size_divisibility, device):
+    """
+    See get_caffe2_inputs() below.
+    """
+    assert all(isinstance(x, dict) for x in batched_inputs)
+    assert all(x["image"].dim() == 3 for x in batched_inputs)
+
+    images = [x["image"] for x in batched_inputs]
+    images = ImageList.from_tensors(images, size_divisibility)
+
+    im_info = []
+    for input_per_image, image_size in zip(batched_inputs, images.image_sizes):
+        target_height = input_per_image.get("height", image_size[0])
+        target_width = input_per_image.get("width", image_size[1])  # noqa
+        # NOTE: The scale inside im_info is kept as convention and for providing
+        # post-processing information if further processing is needed. For
+        # current Caffe2 model definitions that don't include post-processing inside
+        # the model, this number is not used.
+        # NOTE: There can be a slight difference between width and height
+        # scales, using a single number can results in numerical difference
+        # compared with D2's post-processing.
+        scale = target_height / image_size[0]
+        im_info.append([image_size[0], image_size[1], scale])
+    im_info = torch.Tensor(im_info)
+
+    return images.tensor.to(device), im_info.to(device)
+
+
+class Caffe2MetaArch(Caffe2Compatible, torch.nn.Module):
+    """
+    Base class for caffe2-compatible implementation of a meta architecture.
+    The forward is traceable and its traced graph can be converted to caffe2
+    graph through ONNX.
+    """
+
+    def __init__(self, cfg, torch_model):
+        """
+        Args:
+            cfg (CfgNode):
+            torch_model (nn.Module): the detectron2 model (meta_arch) to be
+                converted.
+        """
+        super().__init__()
+        self._wrapped_model = torch_model
+        self.eval()
+        set_caffe2_compatible_tensor_mode(self, True)
+
+    def get_caffe2_inputs(self, batched_inputs):
+        """
+        Convert pytorch-style structured inputs to caffe2-style inputs that
+        are tuples of tensors.
+
+        Args:
+            batched_inputs (list[dict]): inputs to a detectron2 model
+                in its standard format. Each dict has "image" (CHW tensor), and optionally
+                "height" and "width".
+
+        Returns:
+            tuple[Tensor]:
+                tuple of tensors that will be the inputs to the
+                :meth:`forward` method. For existing models, the first
+                is an NCHW tensor (padded and batched); the second is
+                a im_info Nx3 tensor, where the rows are
+                (height, width, unused legacy parameter)
+        """
+        return convert_batched_inputs_to_c2_format(
+            batched_inputs,
+            self._wrapped_model.backbone.size_divisibility,
+            self._wrapped_model.device,
+        )
+
+    def encode_additional_info(self, predict_net, init_net):
+        """
+        Save extra metadata that will be used by inference in the output protobuf.
+        """
+        pass
+
+    def forward(self, inputs):
+        """
+        Run the forward in caffe2-style. It has to use caffe2-compatible ops
+        and the method will be used for tracing.
+
+        Args:
+            inputs (tuple[Tensor]): inputs defined by :meth:`get_caffe2_input`.
+                They will be the inputs of the converted caffe2 graph.
+
+        Returns:
+            tuple[Tensor]: output tensors. They will be the outputs of the
+                converted caffe2 graph.
+        """
+        raise NotImplementedError
+
+    def _caffe2_preprocess_image(self, inputs):
+        """
+        Caffe2 implementation of preprocess_image, which is called inside each MetaArch's forward.
+        It normalizes the input images, and the final caffe2 graph assumes the
+        inputs have been batched already.
+        """
+        data, im_info = inputs
+        data = alias(data, "data")
+        im_info = alias(im_info, "im_info")
+        mean, std = self._wrapped_model.pixel_mean, self._wrapped_model.pixel_std
+        normalized_data = (data - mean) / std
+        normalized_data = alias(normalized_data, "normalized_data")
+
+        # Pack (data, im_info) into ImageList which is recognized by self.inference.
+        images = ImageList(tensor=normalized_data, image_sizes=im_info)
+        return images
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        """
+        Creates a function that converts outputs of the caffe2 model to
+        detectron2's standard format.
+        The function uses information in `predict_net` and `init_net` that are
+        available at inferene time. Therefore the function logic can be used in inference.
+
+        The returned function has the following signature:
+
+            def convert(batched_inputs, c2_inputs, c2_results) -> detectron2_outputs
+
+        Where
+
+            * batched_inputs (list[dict]): the original input format of the meta arch
+            * c2_inputs (dict[str, Tensor]): the caffe2 inputs.
+            * c2_results (dict[str, Tensor]): the caffe2 output format,
+                corresponding to the outputs of the :meth:`forward` function.
+            * detectron2_outputs: the original output format of the meta arch.
+
+        This function can be used to compare the outputs of the original meta arch and
+        the converted caffe2 graph.
+
+        Returns:
+            callable: a callable of the above signature.
+        """
+        raise NotImplementedError
+
+
+class Caffe2GeneralizedRCNN(Caffe2MetaArch):
+    def __init__(self, cfg, torch_model):
+        assert isinstance(torch_model, meta_arch.GeneralizedRCNN)
+        torch_model = patch_generalized_rcnn(torch_model)
+        super().__init__(cfg, torch_model)
+
+        self.roi_heads_patcher = ROIHeadsPatcher(cfg, self._wrapped_model.roi_heads)
+
+    def encode_additional_info(self, predict_net, init_net):
+        size_divisibility = self._wrapped_model.backbone.size_divisibility
+        check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility)
+        check_set_pb_arg(
+            predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii")
+        )
+        check_set_pb_arg(predict_net, "meta_architecture", "s", b"GeneralizedRCNN")
+
+    @mock_torch_nn_functional_interpolate()
+    def forward(self, inputs):
+        if not self.tensor_mode:
+            return self._wrapped_model.inference(inputs)
+        images = self._caffe2_preprocess_image(inputs)
+        features = self._wrapped_model.backbone(images.tensor)
+        proposals, _ = self._wrapped_model.proposal_generator(images, features)
+        with self.roi_heads_patcher.mock_roi_heads():
+            detector_results, _ = self._wrapped_model.roi_heads(images, features, proposals)
+        return tuple(detector_results[0].flatten())
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        def f(batched_inputs, c2_inputs, c2_results):
+            image_sizes = [[int(im[0]), int(im[1])] for im in c2_inputs["im_info"]]
+            results = assemble_rcnn_outputs_by_name(image_sizes, c2_results)
+            return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes)
+
+        return f
+
+
+class Caffe2PanopticFPN(Caffe2MetaArch):
+    def __init__(self, cfg, torch_model):
+        assert isinstance(torch_model, meta_arch.PanopticFPN)
+        torch_model = patch_generalized_rcnn(torch_model)
+        super().__init__(cfg, torch_model)
+
+        self.roi_heads_patcher = ROIHeadsPatcher(cfg, self._wrapped_model.roi_heads)
+
+    @mock_torch_nn_functional_interpolate()
+    def forward(self, inputs):
+        assert self.tensor_mode
+        images = self._caffe2_preprocess_image(inputs)
+        features = self._wrapped_model.backbone(images.tensor)
+
+        sem_seg_results, _ = self._wrapped_model.sem_seg_head(features)
+        sem_seg_results = alias(sem_seg_results, "sem_seg")
+
+        proposals, _ = self._wrapped_model.proposal_generator(images, features)
+
+        with self.roi_heads_patcher.mock_roi_heads(self.tensor_mode):
+            detector_results, _ = self._wrapped_model.roi_heads(images, features, proposals)
+
+        return tuple(detector_results[0].flatten()) + (sem_seg_results,)
+
+    def encode_additional_info(self, predict_net, init_net):
+        size_divisibility = self._wrapped_model.backbone.size_divisibility
+        check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility)
+        check_set_pb_arg(
+            predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii")
+        )
+        check_set_pb_arg(predict_net, "meta_architecture", "s", b"PanopticFPN")
+
+        # Inference parameters:
+        check_set_pb_arg(predict_net, "combine_on", "i", self._wrapped_model.combine_on)
+        check_set_pb_arg(
+            predict_net,
+            "combine_overlap_threshold",
+            "f",
+            _cast_to_f32(self._wrapped_model.combine_overlap_threshold),
+        )
+        check_set_pb_arg(
+            predict_net,
+            "combine_stuff_area_limit",
+            "i",
+            self._wrapped_model.combine_stuff_area_limit,
+        )
+        check_set_pb_arg(
+            predict_net,
+            "combine_instances_confidence_threshold",
+            "f",
+            _cast_to_f32(self._wrapped_model.combine_instances_confidence_threshold),
+        )
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        combine_on = get_pb_arg_vali(predict_net, "combine_on", None)
+        combine_overlap_threshold = get_pb_arg_valf(predict_net, "combine_overlap_threshold", None)
+        combine_stuff_area_limit = get_pb_arg_vali(predict_net, "combine_stuff_area_limit", None)
+        combine_instances_confidence_threshold = get_pb_arg_valf(
+            predict_net, "combine_instances_confidence_threshold", None
+        )
+
+        def f(batched_inputs, c2_inputs, c2_results):
+            image_sizes = [[int(im[0]), int(im[1])] for im in c2_inputs["im_info"]]
+            detector_results = assemble_rcnn_outputs_by_name(
+                image_sizes, c2_results, force_mask_on=True
+            )
+            sem_seg_results = c2_results["sem_seg"]
+
+            # copied from meta_arch/panoptic_fpn.py ...
+            processed_results = []
+            for sem_seg_result, detector_result, input_per_image, image_size in zip(
+                sem_seg_results, detector_results, batched_inputs, image_sizes
+            ):
+                height = input_per_image.get("height", image_size[0])
+                width = input_per_image.get("width", image_size[1])
+                sem_seg_r = sem_seg_postprocess(sem_seg_result, image_size, height, width)
+                detector_r = detector_postprocess(detector_result, height, width)
+
+                processed_results.append({"sem_seg": sem_seg_r, "instances": detector_r})
+
+                if combine_on:
+                    panoptic_r = combine_semantic_and_instance_outputs(
+                        detector_r,
+                        sem_seg_r.argmax(dim=0),
+                        combine_overlap_threshold,
+                        combine_stuff_area_limit,
+                        combine_instances_confidence_threshold,
+                    )
+                    processed_results[-1]["panoptic_seg"] = panoptic_r
+            return processed_results
+
+        return f
+
+
+class Caffe2RetinaNet(Caffe2MetaArch):
+    def __init__(self, cfg, torch_model):
+        assert isinstance(torch_model, meta_arch.RetinaNet)
+        super().__init__(cfg, torch_model)
+
+    @mock_torch_nn_functional_interpolate()
+    def forward(self, inputs):
+        assert self.tensor_mode
+        images = self._caffe2_preprocess_image(inputs)
+
+        # explicitly return the images sizes to avoid removing "im_info" by ONNX
+        # since it's not used in the forward path
+        return_tensors = [images.image_sizes]
+
+        features = self._wrapped_model.backbone(images.tensor)
+        features = [features[f] for f in self._wrapped_model.in_features]
+        for i, feature_i in enumerate(features):
+            features[i] = alias(feature_i, "feature_{}".format(i), is_backward=True)
+            return_tensors.append(features[i])
+
+        box_cls, box_delta = self._wrapped_model.head(features)
+        for i, (box_cls_i, box_delta_i) in enumerate(zip(box_cls, box_delta)):
+            return_tensors.append(alias(box_cls_i, "box_cls_{}".format(i)))
+            return_tensors.append(alias(box_delta_i, "box_delta_{}".format(i)))
+
+        return tuple(return_tensors)
+
+    def encode_additional_info(self, predict_net, init_net):
+        size_divisibility = self._wrapped_model.backbone.size_divisibility
+        check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility)
+        check_set_pb_arg(
+            predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii")
+        )
+        check_set_pb_arg(predict_net, "meta_architecture", "s", b"RetinaNet")
+
+        # Inference parameters:
+        check_set_pb_arg(
+            predict_net, "score_threshold", "f", _cast_to_f32(self._wrapped_model.score_threshold)
+        )
+        check_set_pb_arg(predict_net, "topk_candidates", "i", self._wrapped_model.topk_candidates)
+        check_set_pb_arg(
+            predict_net, "nms_threshold", "f", _cast_to_f32(self._wrapped_model.nms_threshold)
+        )
+        check_set_pb_arg(
+            predict_net,
+            "max_detections_per_image",
+            "i",
+            self._wrapped_model.max_detections_per_image,
+        )
+
+        check_set_pb_arg(
+            predict_net,
+            "bbox_reg_weights",
+            "floats",
+            [_cast_to_f32(w) for w in self._wrapped_model.box2box_transform.weights],
+        )
+        self._encode_anchor_generator_cfg(predict_net)
+
+    def _encode_anchor_generator_cfg(self, predict_net):
+        # serialize anchor_generator for future use
+        serialized_anchor_generator = io.BytesIO()
+        torch.save(self._wrapped_model.anchor_generator, serialized_anchor_generator)
+        # Ideally we can put anchor generating inside the model, then we don't
+        # need to store this information.
+        bytes = serialized_anchor_generator.getvalue()
+        check_set_pb_arg(predict_net, "serialized_anchor_generator", "s", bytes)
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        self = types.SimpleNamespace()
+        serialized_anchor_generator = io.BytesIO(
+            get_pb_arg_vals(predict_net, "serialized_anchor_generator", None)
+        )
+        self.anchor_generator = torch.load(serialized_anchor_generator)
+        bbox_reg_weights = get_pb_arg_floats(predict_net, "bbox_reg_weights", None)
+        self.box2box_transform = Box2BoxTransform(weights=tuple(bbox_reg_weights))
+        self.score_threshold = get_pb_arg_valf(predict_net, "score_threshold", None)
+        self.topk_candidates = get_pb_arg_vali(predict_net, "topk_candidates", None)
+        self.nms_threshold = get_pb_arg_valf(predict_net, "nms_threshold", None)
+        self.max_detections_per_image = get_pb_arg_vali(
+            predict_net, "max_detections_per_image", None
+        )
+
+        # hack to reuse inference code from RetinaNet
+        self.inference = functools.partial(meta_arch.RetinaNet.inference, self)
+        self.inference_single_image = functools.partial(
+            meta_arch.RetinaNet.inference_single_image, self
+        )
+
+        def f(batched_inputs, c2_inputs, c2_results):
+            image_sizes = [[int(im[0]), int(im[1])] for im in c2_inputs["im_info"]]
+
+            num_features = len([x for x in c2_results.keys() if x.startswith("box_cls_")])
+            box_cls = [c2_results["box_cls_{}".format(i)] for i in range(num_features)]
+            box_delta = [c2_results["box_delta_{}".format(i)] for i in range(num_features)]
+
+            # For each feature level, feature should have the same batch size and
+            # spatial dimension as the box_cls and box_delta.
+            dummy_features = [box_delta[i].clone()[:, 0:0, :, :] for i in range(num_features)]
+            anchors = self.anchor_generator(dummy_features)
+
+            # self.num_classess can be inferred
+            self.num_classes = box_cls[0].shape[1] // (box_delta[0].shape[1] // 4)
+
+            results = self.inference(box_cls, box_delta, anchors, image_sizes)
+            return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes)
+
+        return f
+
+
+META_ARCH_CAFFE2_EXPORT_TYPE_MAP = {
+    "GeneralizedRCNN": Caffe2GeneralizedRCNN,
+    "PanopticFPN": Caffe2PanopticFPN,
+    "RetinaNet": Caffe2RetinaNet,
+}

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/patcher.py

@@ -0,0 +1,153 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import contextlib
+import mock
+import torch
+
+from detectron2.modeling import poolers
+from detectron2.modeling.proposal_generator import rpn
+from detectron2.modeling.roi_heads import keypoint_head, mask_head
+from detectron2.modeling.roi_heads.fast_rcnn import FastRCNNOutputLayers
+
+from .c10 import (
+    Caffe2Compatible,
+    Caffe2FastRCNNOutputsInference,
+    Caffe2KeypointRCNNInference,
+    Caffe2MaskRCNNInference,
+    Caffe2ROIPooler,
+    Caffe2RPN,
+)
+
+
+class GenericMixin(object):
+    pass
+
+
+class Caffe2CompatibleConverter(object):
+    """
+    A GenericUpdater which implements the `create_from` interface, by modifying
+    module object and assign it with another class replaceCls.
+    """
+
+    def __init__(self, replaceCls):
+        self.replaceCls = replaceCls
+
+    def create_from(self, module):
+        # update module's class to the new class
+        assert isinstance(module, torch.nn.Module)
+        if issubclass(self.replaceCls, GenericMixin):
+            # replaceCls should act as mixin, create a new class on-the-fly
+            new_class = type(
+                "{}MixedWith{}".format(self.replaceCls.__name__, module.__class__.__name__),
+                (self.replaceCls, module.__class__),
+                {},  # {"new_method": lambda self: ...},
+            )
+            module.__class__ = new_class
+        else:
+            # replaceCls is complete class, this allow arbitrary class swap
+            module.__class__ = self.replaceCls
+
+        # initialize Caffe2Compatible
+        if isinstance(module, Caffe2Compatible):
+            module.tensor_mode = False
+
+        return module
+
+
+def patch(model, target, updater, *args, **kwargs):
+    """
+    recursively (post-order) update all modules with the target type and its
+    subclasses, make a initialization/composition/inheritance/... via the
+    updater.create_from.
+    """
+    for name, module in model.named_children():
+        model._modules[name] = patch(module, target, updater, *args, **kwargs)
+    if isinstance(model, target):
+        return updater.create_from(model, *args, **kwargs)
+    return model
+
+
+def patch_generalized_rcnn(model):
+    ccc = Caffe2CompatibleConverter
+    model = patch(model, rpn.RPN, ccc(Caffe2RPN))
+    model = patch(model, poolers.ROIPooler, ccc(Caffe2ROIPooler))
+
+    return model
+
+
+@contextlib.contextmanager
+def mock_fastrcnn_outputs_inference(
+    tensor_mode, check=True, box_predictor_type=FastRCNNOutputLayers
+):
+    with mock.patch.object(
+        box_predictor_type,
+        "inference",
+        autospec=True,
+        side_effect=Caffe2FastRCNNOutputsInference(tensor_mode),
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+@contextlib.contextmanager
+def mock_mask_rcnn_inference(tensor_mode, patched_module, check=True):
+    with mock.patch(
+        "{}.mask_rcnn_inference".format(patched_module), side_effect=Caffe2MaskRCNNInference()
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+@contextlib.contextmanager
+def mock_keypoint_rcnn_inference(tensor_mode, patched_module, use_heatmap_max_keypoint, check=True):
+    with mock.patch(
+        "{}.keypoint_rcnn_inference".format(patched_module),
+        side_effect=Caffe2KeypointRCNNInference(use_heatmap_max_keypoint),
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+class ROIHeadsPatcher:
+    def __init__(self, cfg, heads):
+        self.heads = heads
+
+        self.use_heatmap_max_keypoint = cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT
+
+    @contextlib.contextmanager
+    def mock_roi_heads(self, tensor_mode=True):
+        """
+        Patching several inference functions inside ROIHeads and its subclasses
+
+        Args:
+            tensor_mode (bool): whether the inputs/outputs are caffe2's tensor
+                format or not. Default to True.
+        """
+        # NOTE: this requries the `keypoint_rcnn_inference` and `mask_rcnn_inference`
+        # are called inside the same file as BaseXxxHead due to using mock.patch.
+        kpt_heads_mod = keypoint_head.BaseKeypointRCNNHead.__module__
+        mask_head_mod = mask_head.BaseMaskRCNNHead.__module__
+
+        mock_ctx_managers = [
+            mock_fastrcnn_outputs_inference(
+                tensor_mode=tensor_mode,
+                check=True,
+                box_predictor_type=type(self.heads.box_predictor),
+            )
+        ]
+        if getattr(self.heads, "keypoint_on", False):
+            mock_ctx_managers += [
+                mock_keypoint_rcnn_inference(
+                    tensor_mode, kpt_heads_mod, self.use_heatmap_max_keypoint
+                )
+            ]
+        if getattr(self.heads, "mask_on", False):
+            mock_ctx_managers += [mock_mask_rcnn_inference(tensor_mode, mask_head_mod)]
+
+        with contextlib.ExitStack() as stack:  # python 3.3+
+            for mgr in mock_ctx_managers:
+                stack.enter_context(mgr)
+            yield

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/export/shared.py

@@ -0,0 +1,1034 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import collections
+import contextlib
+import copy
+import functools
+import logging
+import mock
+import numpy as np
+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+import caffe2.python.utils as putils
+import torch
+import torch.nn.functional as F
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core, net_drawer, workspace
+from torch.nn.functional import interpolate as interp
+
+logger = logging.getLogger(__name__)
+
+
+# ==== torch/utils_toffee/cast.py =======================================
+
+
+def to_device(t, device_str):
+    """
+    This function is a replacement of .to(another_device) such that it allows the
+    casting to be traced properly by explicitly calling the underlying copy ops.
+    It also avoids introducing unncessary op when casting to the same device.
+    """
+    src = t.device
+    dst = torch.device(device_str)
+
+    if src == dst:
+        return t
+    elif src.type == "cuda" and dst.type == "cpu":
+        return torch.ops._caffe2.CopyGPUToCPU(t)
+    elif src.type == "cpu" and dst.type == "cuda":
+        return torch.ops._caffe2.CopyCPUToGPU(t)
+    else:
+        raise RuntimeError("Can't cast tensor from device {} to device {}".format(src, dst))
+
+
+# ==== torch/utils_toffee/interpolate.py =======================================
+
+
+# Note: borrowed from vision/detection/fair/detectron/detectron/modeling/detector.py
+def BilinearInterpolation(tensor_in, up_scale):
+    assert up_scale % 2 == 0, "Scale should be even"
+
+    def upsample_filt(size):
+        factor = (size + 1) // 2
+        if size % 2 == 1:
+            center = factor - 1
+        else:
+            center = factor - 0.5
+
+        og = np.ogrid[:size, :size]
+        return (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor)
+
+    kernel_size = int(up_scale) * 2
+    bil_filt = upsample_filt(kernel_size)
+
+    dim = int(tensor_in.shape[1])
+    kernel = np.zeros((dim, dim, kernel_size, kernel_size), dtype=np.float32)
+    kernel[range(dim), range(dim), :, :] = bil_filt
+
+    tensor_out = F.conv_transpose2d(
+        tensor_in,
+        weight=to_device(torch.Tensor(kernel), tensor_in.device),
+        bias=None,
+        stride=int(up_scale),
+        padding=int(up_scale / 2),
+    )
+
+    return tensor_out
+
+
+# NOTE: ONNX is incompatible with traced torch.nn.functional.interpolate if
+# using dynamic `scale_factor` rather than static `size`. (T43166860)
+# NOTE: Caffe2 Int8 conversion might not be able to quantize `size` properly.
+def onnx_compatibale_interpolate(
+    input, size=None, scale_factor=None, mode="nearest", align_corners=None
+):
+    # NOTE: The input dimensions are interpreted in the form:
+    # `mini-batch x channels x [optional depth] x [optional height] x width`.
+    if size is None and scale_factor is not None:
+        if input.dim() == 4:
+            if isinstance(scale_factor, (int, float)):
+                height_scale, width_scale = (scale_factor, scale_factor)
+            else:
+                assert isinstance(scale_factor, (tuple, list))
+                assert len(scale_factor) == 2
+                height_scale, width_scale = scale_factor
+
+            assert not align_corners, "No matching C2 op for align_corners == True"
+            if mode == "nearest":
+                return torch.ops._caffe2.ResizeNearest(
+                    input, order="NCHW", width_scale=width_scale, height_scale=height_scale
+                )
+            elif mode == "bilinear":
+                logger.warning(
+                    "Use F.conv_transpose2d for bilinear interpolate"
+                    " because there's no such C2 op, this may cause significant"
+                    " slowdown and the boundary pixels won't be as same as"
+                    " using F.interpolate due to padding."
+                )
+                assert height_scale == width_scale
+                return BilinearInterpolation(input, up_scale=height_scale)
+        logger.warning("Output size is not static, it might cause ONNX conversion issue")
+
+    return interp(input, size, scale_factor, mode, align_corners)
+
+
+@contextlib.contextmanager
+def mock_torch_nn_functional_interpolate():
+    if torch.onnx.is_in_onnx_export():
+        with mock.patch(
+            "torch.nn.functional.interpolate", side_effect=onnx_compatibale_interpolate
+        ):
+            yield
+    else:
+        yield
+
+
+# ==== torch/utils_caffe2/ws_utils.py ==========================================
+
+
+class ScopedWS(object):
+    def __init__(self, ws_name, is_reset, is_cleanup=False):
+        self.ws_name = ws_name
+        self.is_reset = is_reset
+        self.is_cleanup = is_cleanup
+        self.org_ws = ""
+
+    def __enter__(self):
+        self.org_ws = workspace.CurrentWorkspace()
+        if self.ws_name is not None:
+            workspace.SwitchWorkspace(self.ws_name, True)
+        if self.is_reset:
+            workspace.ResetWorkspace()
+
+        return workspace
+
+    def __exit__(self, *args):
+        if self.is_cleanup:
+            workspace.ResetWorkspace()
+        if self.ws_name is not None:
+            workspace.SwitchWorkspace(self.org_ws)
+
+
+def fetch_any_blob(name):
+    bb = None
+    try:
+        bb = workspace.FetchBlob(name)
+    except TypeError:
+        bb = workspace.FetchInt8Blob(name)
+    except Exception as e:
+        logger.error("Get blob {} error: {}".format(name, e))
+
+    return bb
+
+
+# ==== torch/utils_caffe2/protobuf.py ==========================================
+
+
+def get_pb_arg(pb, arg_name):
+    for x in pb.arg:
+        if x.name == arg_name:
+            return x
+    return None
+
+
+def get_pb_arg_valf(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.f if arg is not None else default_val
+
+
+def get_pb_arg_floats(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(map(float, arg.floats)) if arg is not None else default_val
+
+
+def get_pb_arg_ints(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(map(int, arg.ints)) if arg is not None else default_val
+
+
+def get_pb_arg_vali(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.i if arg is not None else default_val
+
+
+def get_pb_arg_vals(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.s if arg is not None else default_val
+
+
+def get_pb_arg_valstrings(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(arg.strings) if arg is not None else default_val
+
+
+def check_set_pb_arg(pb, arg_name, arg_attr, arg_value, allow_override=False):
+    arg = get_pb_arg(pb, arg_name)
+    if arg is None:
+        arg = putils.MakeArgument(arg_name, arg_value)
+        assert hasattr(arg, arg_attr)
+        pb.arg.extend([arg])
+    if allow_override and getattr(arg, arg_attr) != arg_value:
+        logger.warning(
+            "Override argument {}: {} -> {}".format(arg_name, getattr(arg, arg_attr), arg_value)
+        )
+        setattr(arg, arg_attr, arg_value)
+    else:
+        assert arg is not None
+        assert getattr(arg, arg_attr) == arg_value, "Existing value {}, new value {}".format(
+            getattr(arg, arg_attr), arg_value
+        )
+
+
+def _create_const_fill_op_from_numpy(name, tensor, device_option=None):
+    assert type(tensor) == np.ndarray
+    kTypeNameMapper = {
+        np.dtype("float32"): "GivenTensorFill",
+        np.dtype("int32"): "GivenTensorIntFill",
+        np.dtype("int64"): "GivenTensorInt64Fill",
+        np.dtype("uint8"): "GivenTensorStringFill",
+    }
+
+    args_dict = {}
+    if tensor.dtype == np.dtype("uint8"):
+        args_dict.update({"values": [str(tensor.data)], "shape": [1]})
+    else:
+        args_dict.update({"values": tensor, "shape": tensor.shape})
+
+    if device_option is not None:
+        args_dict["device_option"] = device_option
+
+    return core.CreateOperator(kTypeNameMapper[tensor.dtype], [], [name], **args_dict)
+
+
+def _create_const_fill_op_from_c2_int8_tensor(name, int8_tensor):
+    assert type(int8_tensor) == workspace.Int8Tensor
+    kTypeNameMapper = {
+        np.dtype("int32"): "Int8GivenIntTensorFill",
+        np.dtype("uint8"): "Int8GivenTensorFill",
+    }
+
+    tensor = int8_tensor.data
+    assert tensor.dtype in [np.dtype("uint8"), np.dtype("int32")]
+    values = tensor.tobytes() if tensor.dtype == np.dtype("uint8") else tensor
+
+    return core.CreateOperator(
+        kTypeNameMapper[tensor.dtype],
+        [],
+        [name],
+        values=values,
+        shape=tensor.shape,
+        Y_scale=int8_tensor.scale,
+        Y_zero_point=int8_tensor.zero_point,
+    )
+
+
+def create_const_fill_op(
+    name: str,
+    blob: Union[np.ndarray, workspace.Int8Tensor],
+    device_option: Optional[caffe2_pb2.DeviceOption] = None,
+) -> caffe2_pb2.OperatorDef:
+    """
+    Given a blob object, return the Caffe2 operator that creates this blob
+    as constant. Currently support NumPy tensor and Caffe2 Int8Tensor.
+    """
+
+    tensor_type = type(blob)
+    assert tensor_type in [
+        np.ndarray,
+        workspace.Int8Tensor,
+    ], 'Error when creating const fill op for "{}", unsupported blob type: {}'.format(
+        name, type(blob)
+    )
+
+    if tensor_type == np.ndarray:
+        return _create_const_fill_op_from_numpy(name, blob, device_option)
+    elif tensor_type == workspace.Int8Tensor:
+        assert device_option is None
+        return _create_const_fill_op_from_c2_int8_tensor(name, blob)
+
+
+def construct_init_net_from_params(
+    params: Dict[str, Any], device_options: Optional[Dict[str, caffe2_pb2.DeviceOption]] = None
+) -> caffe2_pb2.NetDef:
+    """
+    Construct the init_net from params dictionary
+    """
+    init_net = caffe2_pb2.NetDef()
+    device_options = device_options or {}
+    for name, blob in params.items():
+        if isinstance(blob, str):
+            logger.warning(
+                (
+                    "Blob {} with type {} is not supported in generating init net,"
+                    " skipped.".format(name, type(blob))
+                )
+            )
+            continue
+        init_net.op.extend(
+            [create_const_fill_op(name, blob, device_option=device_options.get(name, None))]
+        )
+        init_net.external_output.append(name)
+    return init_net
+
+
+def get_producer_map(ssa):
+    """
+    Return dict from versioned blob to (i, j),
+        where i is index of producer op, j is the index of output of that op.
+    """
+    producer_map = {}
+    for i in range(len(ssa)):
+        outputs = ssa[i][1]
+        for j, outp in enumerate(outputs):
+            producer_map[outp] = (i, j)
+    return producer_map
+
+
+def get_consumer_map(ssa):
+    """
+    Return dict from versioned blob to list of (i, j),
+        where i is index of consumer op, j is the index of input of that op.
+    """
+    consumer_map = collections.defaultdict(list)
+    for i in range(len(ssa)):
+        inputs = ssa[i][0]
+        for j, inp in enumerate(inputs):
+            consumer_map[inp].append((i, j))
+    return consumer_map
+
+
+def get_params_from_init_net(
+    init_net: caffe2_pb2.NetDef,
+) -> [Dict[str, Any], Dict[str, caffe2_pb2.DeviceOption]]:
+    """
+    Take the output blobs from init_net by running it.
+    Outputs:
+        params: dict from blob name to numpy array
+        device_options: dict from blob name to the device option of its creating op
+    """
+    # NOTE: this assumes that the params is determined by producer op with the
+    # only exception be CopyGPUToCPU which is CUDA op but returns CPU tensor.
+    def _get_device_option(producer_op):
+        if producer_op.type == "CopyGPUToCPU":
+            return caffe2_pb2.DeviceOption()
+        else:
+            return producer_op.device_option
+
+    with ScopedWS("__get_params_from_init_net__", is_reset=True, is_cleanup=True) as ws:
+        ws.RunNetOnce(init_net)
+        params = {b: fetch_any_blob(b) for b in init_net.external_output}
+    ssa, versions = core.get_ssa(init_net)
+    producer_map = get_producer_map(ssa)
+    device_options = {
+        b: _get_device_option(init_net.op[producer_map[(b, versions[b])][0]])
+        for b in init_net.external_output
+    }
+    return params, device_options
+
+
+def _updater_raise(op, input_types, output_types):
+    raise RuntimeError(
+        "Failed to apply updater for op {} given input_types {} and"
+        " output_types {}".format(op, input_types, output_types)
+    )
+
+
+def _generic_status_identifier(
+    predict_net: caffe2_pb2.NetDef,
+    status_updater: Callable,
+    known_status: Dict[Tuple[str, int], Any],
+) -> Dict[Tuple[str, int], Any]:
+    """
+    Statically infer the status of each blob, the status can be such as device type
+        (CPU/GPU), layout (NCHW/NHWC), data type (float32/int8), etc. "Blob" here
+        is versioned blob (Tuple[str, int]) in the format compatible with ssa.
+    Inputs:
+        predict_net: the caffe2 network
+        status_updater: a callable, given an op and the status of its input/output,
+            it returns the updated status of input/output. `None` is used for
+            representing unknown status.
+        known_status: a dict containing known status, used as initialization.
+    Outputs:
+        A dict mapping from versioned blob to its status
+    """
+    ssa, versions = core.get_ssa(predict_net)
+    versioned_ext_input = [(b, 0) for b in predict_net.external_input]
+    versioned_ext_output = [(b, versions[b]) for b in predict_net.external_output]
+    all_versioned_blobs = set().union(*[set(x[0] + x[1]) for x in ssa])
+
+    allowed_vbs = all_versioned_blobs.union(versioned_ext_input).union(versioned_ext_output)
+    assert all(k in allowed_vbs for k in known_status)
+    assert all(v is not None for v in known_status.values())
+    _known_status = copy.deepcopy(known_status)
+
+    def _check_and_update(key, value):
+        assert value is not None
+        if key in _known_status:
+            if not _known_status[key] == value:
+                raise RuntimeError(
+                    "Confilict status for {}, existing status {}, new status {}".format(
+                        key, _known_status[key], value
+                    )
+                )
+        _known_status[key] = value
+
+    def _update_i(op, ssa_i):
+        versioned_inputs = ssa_i[0]
+        versioned_outputs = ssa_i[1]
+
+        inputs_status = [_known_status.get(b, None) for b in versioned_inputs]
+        outputs_status = [_known_status.get(b, None) for b in versioned_outputs]
+
+        new_inputs_status, new_outputs_status = status_updater(op, inputs_status, outputs_status)
+
+        for versioned_blob, status in zip(
+            versioned_inputs + versioned_outputs, new_inputs_status + new_outputs_status
+        ):
+            if status is not None:
+                _check_and_update(versioned_blob, status)
+
+    for op, ssa_i in zip(predict_net.op, ssa):
+        _update_i(op, ssa_i)
+    for op, ssa_i in zip(reversed(predict_net.op), reversed(ssa)):
+        _update_i(op, ssa_i)
+
+    # NOTE: This strictly checks all the blob from predict_net must be assgined
+    # a known status. However sometimes it's impossible (eg. having deadend op),
+    # we may relax this constraint if
+    for k in all_versioned_blobs:
+        if k not in _known_status:
+            raise NotImplementedError(
+                "Can not infer the status for {}. Currently only support the case where"
+                " a single forward and backward pass can identify status for all blobs.".format(k)
+            )
+
+    return _known_status
+
+
+def infer_device_type(
+    predict_net: caffe2_pb2.NetDef,
+    known_status: Dict[Tuple[str, int], Any],
+    device_name_style: str = "caffe2",
+) -> Dict[Tuple[str, int], str]:
+    """ Return the device type ("cpu" or "gpu"/"cuda") of each (versioned) blob """
+
+    assert device_name_style in ["caffe2", "pytorch"]
+    _CPU_STR = "cpu"
+    _GPU_STR = "gpu" if device_name_style == "caffe2" else "cuda"
+
+    def _copy_cpu_to_gpu_updater(op, input_types, output_types):
+        if input_types[0] == _GPU_STR or output_types[0] == _CPU_STR:
+            _updater_raise(op, input_types, output_types)
+        return ([_CPU_STR], [_GPU_STR])
+
+    def _copy_gpu_to_cpu_updater(op, input_types, output_types):
+        if input_types[0] == _CPU_STR or output_types[0] == _GPU_STR:
+            _updater_raise(op, input_types, output_types)
+        return ([_GPU_STR], [_CPU_STR])
+
+    def _other_ops_updater(op, input_types, output_types):
+        non_none_types = [x for x in input_types + output_types if x is not None]
+        if len(non_none_types) > 0:
+            the_type = non_none_types[0]
+            if not all(x == the_type for x in non_none_types):
+                _updater_raise(op, input_types, output_types)
+        else:
+            the_type = None
+        return ([the_type for _ in op.input], [the_type for _ in op.output])
+
+    def _device_updater(op, *args, **kwargs):
+        return {
+            "CopyCPUToGPU": _copy_cpu_to_gpu_updater,
+            "CopyGPUToCPU": _copy_gpu_to_cpu_updater,
+        }.get(op.type, _other_ops_updater)(op, *args, **kwargs)
+
+    return _generic_status_identifier(predict_net, _device_updater, known_status)
+
+
+# ==== torch/utils_caffe2/vis.py ===============================================
+
+
+def _modify_blob_names(ops, blob_rename_f):
+    ret = []
+
+    def _replace_list(blob_list, replaced_list):
+        del blob_list[:]
+        blob_list.extend(replaced_list)
+
+    for x in ops:
+        cur = copy.deepcopy(x)
+        _replace_list(cur.input, list(map(blob_rename_f, cur.input)))
+        _replace_list(cur.output, list(map(blob_rename_f, cur.output)))
+        ret.append(cur)
+
+    return ret
+
+
+def _rename_blob(name, blob_sizes, blob_ranges):
+    def _list_to_str(bsize):
+        ret = ", ".join([str(x) for x in bsize])
+        ret = "[" + ret + "]"
+        return ret
+
+    ret = name
+    if blob_sizes is not None and name in blob_sizes:
+        ret += "\n" + _list_to_str(blob_sizes[name])
+    if blob_ranges is not None and name in blob_ranges:
+        ret += "\n" + _list_to_str(blob_ranges[name])
+
+    return ret
+
+
+# graph_name could not contain word 'graph'
+def save_graph(net, file_name, graph_name="net", op_only=True, blob_sizes=None, blob_ranges=None):
+    blob_rename_f = functools.partial(_rename_blob, blob_sizes=blob_sizes, blob_ranges=blob_ranges)
+    return save_graph_base(net, file_name, graph_name, op_only, blob_rename_f)
+
+
+def save_graph_base(net, file_name, graph_name="net", op_only=True, blob_rename_func=None):
+    graph = None
+    ops = net.op
+    if blob_rename_func is not None:
+        ops = _modify_blob_names(ops, blob_rename_func)
+    if not op_only:
+        graph = net_drawer.GetPydotGraph(ops, graph_name, rankdir="TB")
+    else:
+        graph = net_drawer.GetPydotGraphMinimal(
+            ops, graph_name, rankdir="TB", minimal_dependency=True
+        )
+
+    try:
+        par_dir = os.path.dirname(file_name)
+        if not os.path.exists(par_dir):
+            os.makedirs(par_dir)
+
+        format = os.path.splitext(os.path.basename(file_name))[-1]
+        if format == ".png":
+            graph.write_png(file_name)
+        elif format == ".pdf":
+            graph.write_pdf(file_name)
+        elif format == ".svg":
+            graph.write_svg(file_name)
+        else:
+            print("Incorrect format {}".format(format))
+    except Exception as e:
+        print("Error when writing graph to image {}".format(e))
+
+    return graph
+
+
+# ==== torch/utils_toffee/aten_to_caffe2.py ====================================
+
+
+def group_norm_replace_aten_with_caffe2(predict_net: caffe2_pb2.NetDef):
+    """
+    For ONNX exported model, GroupNorm will be represented as ATen op,
+        this can be a drop in replacement from ATen to GroupNorm
+    """
+    count = 0
+    for op in predict_net.op:
+        if op.type == "ATen":
+            op_name = get_pb_arg_vals(op, "operator", None)  # return byte in py3
+            if op_name and op_name.decode() == "group_norm":
+                op.arg.remove(get_pb_arg(op, "operator"))
+
+                if get_pb_arg_vali(op, "cudnn_enabled", None):
+                    op.arg.remove(get_pb_arg(op, "cudnn_enabled"))
+
+                num_groups = get_pb_arg_vali(op, "num_groups", None)
+                if num_groups is not None:
+                    op.arg.remove(get_pb_arg(op, "num_groups"))
+                    check_set_pb_arg(op, "group", "i", num_groups)
+
+                op.type = "GroupNorm"
+                count += 1
+    if count > 1:
+        logger.info("Replaced {} ATen operator to GroupNormOp".format(count))
+
+
+# ==== torch/utils_toffee/alias.py =============================================
+
+
+def alias(x, name, is_backward=False):
+    if not torch.onnx.is_in_onnx_export():
+        return x
+    assert isinstance(x, torch.Tensor)
+    return torch.ops._caffe2.AliasWithName(x, name, is_backward=is_backward)
+
+
+def fuse_alias_placeholder(predict_net, init_net):
+    """ Remove AliasWithName placeholder and rename the input/output of it """
+    # First we finish all the re-naming
+    for i, op in enumerate(predict_net.op):
+        if op.type == "AliasWithName":
+            assert len(op.input) == 1
+            assert len(op.output) == 1
+            name = get_pb_arg_vals(op, "name", None).decode()
+            is_backward = bool(get_pb_arg_vali(op, "is_backward", 0))
+            rename_op_input(predict_net, init_net, i, 0, name, from_producer=is_backward)
+            rename_op_output(predict_net, i, 0, name)
+
+    # Remove AliasWithName, should be very safe since it's a non-op
+    new_ops = []
+    for op in predict_net.op:
+        if op.type != "AliasWithName":
+            new_ops.append(op)
+        else:
+            # safety check
+            assert op.input == op.output
+            assert op.input[0] == op.arg[0].s.decode()
+    del predict_net.op[:]
+    predict_net.op.extend(new_ops)
+
+
+# ==== torch/utils_caffe2/graph_transform.py ===================================
+
+
+class IllegalGraphTransformError(ValueError):
+    """ When a graph transform function call can't be executed. """
+
+
+def _rename_versioned_blob_in_proto(
+    proto: caffe2_pb2.NetDef,
+    old_name: str,
+    new_name: str,
+    version: int,
+    ssa: List[Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]],
+    start_versions: Dict[str, int],
+    end_versions: Dict[str, int],
+):
+    """ In given proto, rename all blobs with matched version """
+    # Operater list
+    for op, i_th_ssa in zip(proto.op, ssa):
+        versioned_inputs, versioned_outputs = i_th_ssa
+        for i in range(len(op.input)):
+            if versioned_inputs[i] == (old_name, version):
+                op.input[i] = new_name
+        for i in range(len(op.output)):
+            if versioned_outputs[i] == (old_name, version):
+                op.output[i] = new_name
+    # external_input
+    if start_versions.get(old_name, 0) == version:
+        for i in range(len(proto.external_input)):
+            if proto.external_input[i] == old_name:
+                proto.external_input[i] = new_name
+    # external_output
+    if end_versions.get(old_name, 0) == version:
+        for i in range(len(proto.external_output)):
+            if proto.external_output[i] == old_name:
+                proto.external_output[i] = new_name
+
+
+def rename_op_input(
+    predict_net: caffe2_pb2.NetDef,
+    init_net: caffe2_pb2.NetDef,
+    op_id: int,
+    input_id: int,
+    new_name: str,
+    from_producer: bool = False,
+):
+    """
+    Rename the op_id-th operator in predict_net, change it's input_id-th input's
+        name to the new_name. It also does automatic re-route and change
+        external_input and init_net if necessary.
+    - It requires the input is only consumed by this op.
+    - This function modifies predict_net and init_net in-place.
+    - When from_producer is enable, this also updates other operators that consumes
+        the same input. Be cautious because may trigger unintended behavior.
+    """
+    assert isinstance(predict_net, caffe2_pb2.NetDef)
+    assert isinstance(init_net, caffe2_pb2.NetDef)
+
+    init_net_ssa, init_net_versions = core.get_ssa(init_net)
+    predict_net_ssa, predict_net_versions = core.get_ssa(
+        predict_net, copy.deepcopy(init_net_versions)
+    )
+
+    versioned_inputs, versioned_outputs = predict_net_ssa[op_id]
+    old_name, version = versioned_inputs[input_id]
+
+    if from_producer:
+        producer_map = get_producer_map(predict_net_ssa)
+        if not (old_name, version) in producer_map:
+            raise NotImplementedError(
+                "Can't find producer, the input {} is probably from"
+                " init_net, this is not supported yet.".format(old_name)
+            )
+        producer = producer_map[(old_name, version)]
+        rename_op_output(predict_net, producer[0], producer[1], new_name)
+        return
+
+    def contain_targets(op_ssa):
+        return (old_name, version) in op_ssa[0]
+
+    is_consumer = [contain_targets(op_ssa) for op_ssa in predict_net_ssa]
+    if sum(is_consumer) > 1:
+        raise IllegalGraphTransformError(
+            (
+                "Input '{}' of operator(#{}) are consumed by other ops, please use"
+                + " rename_op_output on the producer instead. Offending op: \n{}"
+            ).format(old_name, op_id, predict_net.op[op_id])
+        )
+
+    # update init_net
+    _rename_versioned_blob_in_proto(
+        init_net, old_name, new_name, version, init_net_ssa, {}, init_net_versions
+    )
+    # update predict_net
+    _rename_versioned_blob_in_proto(
+        predict_net,
+        old_name,
+        new_name,
+        version,
+        predict_net_ssa,
+        init_net_versions,
+        predict_net_versions,
+    )
+
+
+def rename_op_output(predict_net: caffe2_pb2.NetDef, op_id: int, output_id: int, new_name: str):
+    """
+    Rename the op_id-th operator in predict_net, change it's output_id-th input's
+        name to the new_name. It also does automatic re-route and change
+        external_output and if necessary.
+    - It allows multiple consumers of its output.
+    - This function modifies predict_net in-place, doesn't need init_net.
+    """
+    assert isinstance(predict_net, caffe2_pb2.NetDef)
+
+    ssa, blob_versions = core.get_ssa(predict_net)
+
+    versioned_inputs, versioned_outputs = ssa[op_id]
+    old_name, version = versioned_outputs[output_id]
+
+    # update predict_net
+    _rename_versioned_blob_in_proto(
+        predict_net, old_name, new_name, version, ssa, {}, blob_versions
+    )
+
+
+def get_sub_graph_external_input_output(
+    predict_net: caffe2_pb2.NetDef, sub_graph_op_indices: List[int]
+) -> Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]:
+    """
+    Return the list of external input/output of sub-graph,
+    each element is tuple of the name and corresponding version in predict_net.
+
+    external input/output is defined the same way as caffe2 NetDef.
+    """
+    ssa, versions = core.get_ssa(predict_net)
+
+    all_inputs = []
+    all_outputs = []
+    for op_id in sub_graph_op_indices:
+        all_inputs += [inp for inp in ssa[op_id][0] if inp not in all_inputs]
+        all_outputs += list(ssa[op_id][1])  # ssa output won't repeat
+
+    # for versioned blobs, external inputs are just those blob in all_inputs
+    # but not in all_outputs
+    ext_inputs = [inp for inp in all_inputs if inp not in all_outputs]
+
+    # external outputs are essentially outputs of this subgraph that are used
+    # outside of this sub-graph (including predict_net.external_output)
+    all_other_inputs = sum(
+        (ssa[i][0] for i in range(len(ssa)) if i not in sub_graph_op_indices),
+        [(outp, versions[outp]) for outp in predict_net.external_output],
+    )
+    ext_outputs = [outp for outp in all_outputs if outp in set(all_other_inputs)]
+
+    return ext_inputs, ext_outputs
+
+
+class DiGraph:
+    """ A DAG representation of caffe2 graph, each vertice is a versioned blob. """
+
+    def __init__(self):
+        self.vertices = set()
+        self.graph = collections.defaultdict(list)
+
+    def add_edge(self, u, v):
+        self.graph[u].append(v)
+        self.vertices.add(u)
+        self.vertices.add(v)
+
+    # grab from https://www.geeksforgeeks.org/find-paths-given-source-destination/
+    def get_all_paths(self, s, d):
+        visited = {k: False for k in self.vertices}
+        path = []
+        all_paths = []
+
+        def _get_all_paths_util(graph, u, d, visited, path):
+            visited[u] = True
+            path.append(u)
+            if u == d:
+                all_paths.append(copy.deepcopy(path))
+            else:
+                for i in graph[u]:
+                    if not visited[i]:
+                        _get_all_paths_util(graph, i, d, visited, path)
+            path.pop()
+            visited[u] = False
+
+        _get_all_paths_util(self.graph, s, d, visited, path)
+        return all_paths
+
+    @staticmethod
+    def from_ssa(ssa):
+        graph = DiGraph()
+        for op_id in range(len(ssa)):
+            for inp in ssa[op_id][0]:
+                for outp in ssa[op_id][1]:
+                    graph.add_edge(inp, outp)
+        return graph
+
+
+def _get_dependency_chain(ssa, versioned_target, versioned_source):
+    """
+    Return the index list of relevant operator to produce target blob from source blob,
+        if there's no dependency, return empty list.
+    """
+
+    # finding all paths between nodes can be O(N!), thus we can only search
+    # in the subgraph using the op starting from the first consumer of source blob
+    # to the producer of the target blob.
+    consumer_map = get_consumer_map(ssa)
+    producer_map = get_producer_map(ssa)
+    start_op = min(x[0] for x in consumer_map[versioned_source]) - 15
+    end_op = (
+        producer_map[versioned_target][0] + 15 if versioned_target in producer_map else start_op
+    )
+    sub_graph_ssa = ssa[start_op : end_op + 1]
+    if len(sub_graph_ssa) > 30:
+        logger.warning(
+            "Subgraph bebetween {} and {} is large (from op#{} to op#{}), it"
+            " might take non-trival time to find all paths between them.".format(
+                versioned_source, versioned_target, start_op, end_op
+            )
+        )
+
+    dag = DiGraph.from_ssa(sub_graph_ssa)
+    paths = dag.get_all_paths(versioned_source, versioned_target)  # include two ends
+    ops_in_paths = [[producer_map[blob][0] for blob in path[1:]] for path in paths]
+    return sorted(set().union(*[set(ops) for ops in ops_in_paths]))
+
+
+def identify_reshape_sub_graph(predict_net: caffe2_pb2.NetDef) -> List[List[int]]:
+    """
+    Idenfity the reshape sub-graph in a protobuf.
+    The reshape sub-graph is defined as matching the following pattern:
+
+    (input_blob) -> Op_1 -> ... -> Op_N -> (new_shape) -─┐
+        └-------------------------------------------> Reshape -> (output_blob)
+
+    Return:
+        List of sub-graphs, each sub-graph is represented as a list of indices
+        of the relavent ops, [Op_1, Op_2, ..., Op_N, Reshape]
+    """
+
+    ssa, _ = core.get_ssa(predict_net)
+
+    ret = []
+    for i, op in enumerate(predict_net.op):
+        if op.type == "Reshape":
+            assert len(op.input) == 2
+            input_ssa = ssa[i][0]
+            data_source = input_ssa[0]
+            shape_source = input_ssa[1]
+            op_indices = _get_dependency_chain(ssa, shape_source, data_source)
+            ret.append(op_indices + [i])
+    return ret
+
+
+def remove_reshape_for_fc(predict_net, params):
+    """
+    In PyTorch nn.Linear has to take 2D tensor, this often leads to reshape
+        a 4D tensor to 2D by calling .view(). However this (dynamic) reshaping
+        doesn't work well with ONNX and Int8 tools, and cause using extra
+        ops (eg. ExpandDims) that might not be available on mobile.
+    Luckily Caffe2 supports 4D tensor for FC, so we can remove those reshape
+        after exporting ONNX model.
+    """
+    from caffe2.python import core
+
+    # find all reshape sub-graph that can be removed, which is now all Reshape
+    # sub-graph whose output is only consumed by FC.
+    # TODO: to make it safer, we may need the actually value to better determine
+    # if a Reshape before FC is removable.
+    reshape_sub_graphs = identify_reshape_sub_graph(predict_net)
+    sub_graphs_to_remove = []
+    for reshape_sub_graph in reshape_sub_graphs:
+        reshape_op_id = reshape_sub_graph[-1]
+        assert predict_net.op[reshape_op_id].type == "Reshape"
+        ssa, _ = core.get_ssa(predict_net)
+        reshape_output = ssa[reshape_op_id][1][0]
+        consumers = [i for i in range(len(ssa)) if reshape_output in ssa[i][0]]
+        if all(predict_net.op[consumer].type == "FC" for consumer in consumers):
+            # safety check if the sub-graph is isolated, for this reshape sub-graph,
+            # it means it has one non-param external input and one external output.
+            ext_inputs, ext_outputs = get_sub_graph_external_input_output(
+                predict_net, reshape_sub_graph
+            )
+            non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0]
+            if len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1:
+                sub_graphs_to_remove.append(reshape_sub_graph)
+
+    # perform removing subgraph by:
+    # 1: rename the Reshape's output to its input, then the graph can be
+    #   seen as in-place itentify, meaning whose external input/output are the same.
+    # 2: simply remove those ops.
+    remove_op_ids = []
+    params_to_remove = []
+    for sub_graph in sub_graphs_to_remove:
+        logger.info(
+            "Remove Reshape sub-graph:\n{}".format(
+                "".join(["(#{:>4})\n{}".format(i, predict_net.op[i]) for i in sub_graph])
+            )
+        )
+        reshape_op_id = sub_graph[-1]
+        new_reshap_output = predict_net.op[reshape_op_id].input[0]
+        rename_op_output(predict_net, reshape_op_id, 0, new_reshap_output)
+        ext_inputs, ext_outputs = get_sub_graph_external_input_output(predict_net, sub_graph)
+        non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0]
+        params_ext_inputs = [inp for inp in ext_inputs if inp[1] == 0]
+        assert len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1
+        assert ext_outputs[0][0] == non_params_ext_inputs[0][0]
+        assert ext_outputs[0][1] == non_params_ext_inputs[0][1] + 1
+        remove_op_ids.extend(sub_graph)
+        params_to_remove.extend(params_ext_inputs)
+
+    predict_net = copy.deepcopy(predict_net)
+    new_ops = [op for i, op in enumerate(predict_net.op) if i not in remove_op_ids]
+    del predict_net.op[:]
+    predict_net.op.extend(new_ops)
+    for versioned_params in params_to_remove:
+        name = versioned_params[0]
+        logger.info("Remove params: {} from init_net and predict_net.external_input".format(name))
+        del params[name]
+        predict_net.external_input.remove(name)
+
+    return predict_net, params
+
+
+def fuse_copy_between_cpu_and_gpu(predict_net: caffe2_pb2.NetDef):
+    """
+    In-place fuse extra copy ops between cpu/gpu for the following case:
+        a -CopyAToB-> b -CopyBToA> c1 -NextOp1-> d1
+                        -CopyBToA> c2 -NextOp2-> d2
+    The fused network will look like:
+        a -NextOp1-> d1
+          -NextOp2-> d2
+    """
+
+    _COPY_OPS = ["CopyCPUToGPU", "CopyGPUToCPU"]
+
+    def _fuse_once(predict_net):
+        ssa, blob_versions = core.get_ssa(predict_net)
+        consumer_map = get_consumer_map(ssa)
+        versioned_external_output = [
+            (name, blob_versions[name]) for name in predict_net.external_output
+        ]
+
+        for op_id, op in enumerate(predict_net.op):
+            if op.type in _COPY_OPS:
+                fw_copy_versioned_output = ssa[op_id][1][0]
+                consumer_ids = [x[0] for x in consumer_map[fw_copy_versioned_output]]
+                reverse_op_type = _COPY_OPS[1 - _COPY_OPS.index(op.type)]
+
+                is_fusable = (
+                    len(consumer_ids) > 0
+                    and fw_copy_versioned_output not in versioned_external_output
+                    and all(
+                        predict_net.op[_op_id].type == reverse_op_type
+                        and ssa[_op_id][1][0] not in versioned_external_output
+                        for _op_id in consumer_ids
+                    )
+                )
+
+                if is_fusable:
+                    for rv_copy_op_id in consumer_ids:
+                        # making each NextOp uses "a" directly and removing Copy ops
+                        rs_copy_versioned_output = ssa[rv_copy_op_id][1][0]
+                        next_op_id, inp_id = consumer_map[rs_copy_versioned_output][0]
+                        predict_net.op[next_op_id].input[inp_id] = op.input[0]
+                    # remove CopyOps
+                    new_ops = [
+                        op
+                        for i, op in enumerate(predict_net.op)
+                        if i != op_id and i not in consumer_ids
+                    ]
+                    del predict_net.op[:]
+                    predict_net.op.extend(new_ops)
+                    return True
+
+        return False
+
+    # _fuse_once returns False is nothing can be fused
+    while _fuse_once(predict_net):
+        pass
+
+
+def remove_dead_end_ops(net_def: caffe2_pb2.NetDef):
+    """ remove ops if its output is not used or not in external_output """
+    ssa, versions = core.get_ssa(net_def)
+    versioned_external_output = [(name, versions[name]) for name in net_def.external_output]
+    consumer_map = get_consumer_map(ssa)
+    removed_op_ids = set()
+
+    def _is_dead_end(versioned_blob):
+        return not (
+            versioned_blob in versioned_external_output
+            or (
+                len(consumer_map[versioned_blob]) > 0
+                and all(x[0] not in removed_op_ids for x in consumer_map[versioned_blob])
+            )
+        )
+
+    for i, ssa_i in reversed(list(enumerate(ssa))):
+        versioned_outputs = ssa_i[1]
+        if all(_is_dead_end(outp) for outp in versioned_outputs):
+            removed_op_ids.add(i)
+
+    # simply removing those deadend ops should have no effect to external_output
+    new_ops = [op for i, op in enumerate(net_def.op) if i not in removed_op_ids]
+    del net_def.op[:]
+    net_def.op.extend(new_ops)

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated.h

@@ -0,0 +1,35 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+#pragma once
+#include <torch/types.h>
+
+namespace detectron2 {
+
+at::Tensor box_iou_rotated_cpu(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2);
+
+#ifdef WITH_CUDA
+at::Tensor box_iou_rotated_cuda(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2);
+#endif
+
+// Interface for Python
+// inline is needed to prevent multiple function definitions when this header is
+// included by different cpps
+inline at::Tensor box_iou_rotated(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2) {
+  assert(boxes1.device().is_cuda() == boxes2.device().is_cuda());
+  if (boxes1.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return box_iou_rotated_cuda(boxes1.contiguous(), boxes2.contiguous());
+#else
+    AT_ERROR("Not compiled with GPU support");
+#endif
+  }
+
+  return box_iou_rotated_cpu(boxes1.contiguous(), boxes2.contiguous());
+}
+
+} // namespace detectron2

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cpu.cpp

@@ -0,0 +1,39 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+#include "box_iou_rotated.h"
+#include "box_iou_rotated_utils.h"
+
+namespace detectron2 {
+
+template <typename T>
+void box_iou_rotated_cpu_kernel(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2,
+    at::Tensor& ious) {
+  auto num_boxes1 = boxes1.size(0);
+  auto num_boxes2 = boxes2.size(0);
+
+  for (int i = 0; i < num_boxes1; i++) {
+    for (int j = 0; j < num_boxes2; j++) {
+      ious[i * num_boxes2 + j] = single_box_iou_rotated<T>(
+          boxes1[i].data_ptr<T>(), boxes2[j].data_ptr<T>());
+    }
+  }
+}
+
+at::Tensor box_iou_rotated_cpu(
+    // input must be contiguous:
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2) {
+  auto num_boxes1 = boxes1.size(0);
+  auto num_boxes2 = boxes2.size(0);
+  at::Tensor ious =
+      at::empty({num_boxes1 * num_boxes2}, boxes1.options().dtype(at::kFloat));
+
+  box_iou_rotated_cpu_kernel<float>(boxes1, boxes2, ious);
+
+  // reshape from 1d array to 2d array
+  auto shape = std::vector<int64_t>{num_boxes1, num_boxes2};
+  return ious.reshape(shape);
+}
+
+} // namespace detectron2

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cuda.cu

@@ -0,0 +1,130 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+#include "box_iou_rotated_utils.h"
+
+namespace detectron2 {
+
+// 2D block with 32 * 16 = 512 threads per block
+const int BLOCK_DIM_X = 32;
+const int BLOCK_DIM_Y = 16;
+
+template <typename T>
+__global__ void box_iou_rotated_cuda_kernel(
+    const int n_boxes1,
+    const int n_boxes2,
+    const T* dev_boxes1,
+    const T* dev_boxes2,
+    T* dev_ious) {
+  const int row_start = blockIdx.x * blockDim.x;
+  const int col_start = blockIdx.y * blockDim.y;
+
+  const int row_size = min(n_boxes1 - row_start, blockDim.x);
+  const int col_size = min(n_boxes2 - col_start, blockDim.y);
+
+  __shared__ float block_boxes1[BLOCK_DIM_X * 5];
+  __shared__ float block_boxes2[BLOCK_DIM_Y * 5];
+
+  // It's safe to copy using threadIdx.x since BLOCK_DIM_X >= BLOCK_DIM_Y
+  if (threadIdx.x < row_size && threadIdx.y == 0) {
+    block_boxes1[threadIdx.x * 5 + 0] =
+        dev_boxes1[(row_start + threadIdx.x) * 5 + 0];
+    block_boxes1[threadIdx.x * 5 + 1] =
+        dev_boxes1[(row_start + threadIdx.x) * 5 + 1];
+    block_boxes1[threadIdx.x * 5 + 2] =
+        dev_boxes1[(row_start + threadIdx.x) * 5 + 2];
+    block_boxes1[threadIdx.x * 5 + 3] =
+        dev_boxes1[(row_start + threadIdx.x) * 5 + 3];
+    block_boxes1[threadIdx.x * 5 + 4] =
+        dev_boxes1[(row_start + threadIdx.x) * 5 + 4];
+  }
+
+  if (threadIdx.x < col_size && threadIdx.y == 0) {
+    block_boxes2[threadIdx.x * 5 + 0] =
+        dev_boxes2[(col_start + threadIdx.x) * 5 + 0];
+    block_boxes2[threadIdx.x * 5 + 1] =
+        dev_boxes2[(col_start + threadIdx.x) * 5 + 1];
+    block_boxes2[threadIdx.x * 5 + 2] =
+        dev_boxes2[(col_start + threadIdx.x) * 5 + 2];
+    block_boxes2[threadIdx.x * 5 + 3] =
+        dev_boxes2[(col_start + threadIdx.x) * 5 + 3];
+    block_boxes2[threadIdx.x * 5 + 4] =
+        dev_boxes2[(col_start + threadIdx.x) * 5 + 4];
+  }
+  __syncthreads();
+
+  if (threadIdx.x < row_size && threadIdx.y < col_size) {
+    int offset = (row_start + threadIdx.x) * n_boxes2 + col_start + threadIdx.y;
+    dev_ious[offset] = single_box_iou_rotated<T>(
+        block_boxes1 + threadIdx.x * 5, block_boxes2 + threadIdx.y * 5);
+  }
+}
+
+at::Tensor box_iou_rotated_cuda(
+    // input must be contiguous
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2) {
+  using scalar_t = float;
+  AT_ASSERTM(
+      boxes1.scalar_type() == at::kFloat, "boxes1 must be a float tensor");
+  AT_ASSERTM(
+      boxes2.scalar_type() == at::kFloat, "boxes2 must be a float tensor");
+  AT_ASSERTM(boxes1.is_cuda(), "boxes1 must be a CUDA tensor");
+  AT_ASSERTM(boxes2.is_cuda(), "boxes2 must be a CUDA tensor");
+  at::cuda::CUDAGuard device_guard(boxes1.device());
+
+  auto num_boxes1 = boxes1.size(0);
+  auto num_boxes2 = boxes2.size(0);
+
+  at::Tensor ious =
+      at::empty({num_boxes1 * num_boxes2}, boxes1.options().dtype(at::kFloat));
+
+  bool transpose = false;
+  if (num_boxes1 > 0 && num_boxes2 > 0) {
+    scalar_t *data1 = boxes1.data_ptr<scalar_t>(),
+             *data2 = boxes2.data_ptr<scalar_t>();
+
+    if (num_boxes2 > 65535 * BLOCK_DIM_Y) {
+      AT_ASSERTM(
+          num_boxes1 <= 65535 * BLOCK_DIM_Y,
+          "Too many boxes for box_iou_rotated_cuda!");
+      // x dim is allowed to be large, but y dim cannot,
+      // so we transpose the two to avoid "invalid configuration argument"
+      // error. We assume one of them is small. Otherwise the result is hard to
+      // fit in memory anyway.
+      std::swap(num_boxes1, num_boxes2);
+      std::swap(data1, data2);
+      transpose = true;
+    }
+
+    const int blocks_x =
+        at::cuda::ATenCeilDiv(static_cast<int>(num_boxes1), BLOCK_DIM_X);
+    const int blocks_y =
+        at::cuda::ATenCeilDiv(static_cast<int>(num_boxes2), BLOCK_DIM_Y);
+
+    dim3 blocks(blocks_x, blocks_y);
+    dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y);
+    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+    box_iou_rotated_cuda_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
+        num_boxes1,
+        num_boxes2,
+        data1,
+        data2,
+        (scalar_t*)ious.data_ptr<scalar_t>());
+
+    AT_CUDA_CHECK(cudaGetLastError());
+  }
+
+  // reshape from 1d array to 2d array
+  auto shape = std::vector<int64_t>{num_boxes1, num_boxes2};
+  if (transpose) {
+    return ious.view(shape).t();
+  } else {
+    return ious.view(shape);
+  }
+}
+
+} // namespace detectron2

Leffa/preprocess/humanparsing/mhp_extension/detectron2/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_utils.h

@@ -0,0 +1,363 @@
+// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+#pragma once
+
+#include <cassert>
+#include <cmath>
+
+#ifdef __CUDACC__
+// Designates functions callable from the host (CPU) and the device (GPU)
+#define HOST_DEVICE __host__ __device__
+#define HOST_DEVICE_INLINE HOST_DEVICE __forceinline__
+#else
+#include <algorithm>
+#define HOST_DEVICE
+#define HOST_DEVICE_INLINE HOST_DEVICE inline
+#endif
+
+namespace detectron2 {
+
+namespace {
+
+template <typename T>
+struct RotatedBox {
+  T x_ctr, y_ctr, w, h, a;
+};
+
+template <typename T>
+struct Point {
+  T x, y;
+  HOST_DEVICE_INLINE Point(const T& px = 0, const T& py = 0) : x(px), y(py) {}
+  HOST_DEVICE_INLINE Point operator+(const Point& p) const {
+    return Point(x + p.x, y + p.y);
+  }
+  HOST_DEVICE_INLINE Point& operator+=(const Point& p) {
+    x += p.x;
+    y += p.y;
+    return *this;
+  }
+  HOST_DEVICE_INLINE Point operator-(const Point& p) const {
+    return Point(x - p.x, y - p.y);
+  }
+  HOST_DEVICE_INLINE Point operator*(const T coeff) const {
+    return Point(x * coeff, y * coeff);
+  }
+};
+
+template <typename T>
+HOST_DEVICE_INLINE T dot_2d(const Point<T>& A, const Point<T>& B) {
+  return A.x * B.x + A.y * B.y;
+}
+
+// R: result type. can be different from input type
+template <typename T, typename R = T>
+HOST_DEVICE_INLINE R cross_2d(const Point<T>& A, const Point<T>& B) {
+  return static_cast<R>(A.x) * static_cast<R>(B.y) -
+      static_cast<R>(B.x) * static_cast<R>(A.y);
+}
+
+template <typename T>
+HOST_DEVICE_INLINE void get_rotated_vertices(
+    const RotatedBox<T>& box,
+    Point<T> (&pts)[4]) {
+  // M_PI / 180. == 0.01745329251
+  double theta = box.a * 0.01745329251;
+  T cosTheta2 = (T)cos(theta) * 0.5f;
+  T sinTheta2 = (T)sin(theta) * 0.5f;
+
+  // y: top --> down; x: left --> right
+  pts[0].x = box.x_ctr + sinTheta2 * box.h + cosTheta2 * box.w;
+  pts[0].y = box.y_ctr + cosTheta2 * box.h - sinTheta2 * box.w;
+  pts[1].x = box.x_ctr - sinTheta2 * box.h + cosTheta2 * box.w;
+  pts[1].y = box.y_ctr - cosTheta2 * box.h - sinTheta2 * box.w;
+  pts[2].x = 2 * box.x_ctr - pts[0].x;
+  pts[2].y = 2 * box.y_ctr - pts[0].y;
+  pts[3].x = 2 * box.x_ctr - pts[1].x;
+  pts[3].y = 2 * box.y_ctr - pts[1].y;
+}
+
+template <typename T>
+HOST_DEVICE_INLINE int get_intersection_points(
+    const Point<T> (&pts1)[4],
+    const Point<T> (&pts2)[4],
+    Point<T> (&intersections)[24]) {
+  // Line vector
+  // A line from p1 to p2 is: p1 + (p2-p1)*t, t=[0,1]
+  Point<T> vec1[4], vec2[4];
+  for (int i = 0; i < 4; i++) {
+    vec1[i] = pts1[(i + 1) % 4] - pts1[i];
+    vec2[i] = pts2[(i + 1) % 4] - pts2[i];
+  }
+
+  // Line test - test all line combos for intersection
+  int num = 0; // number of intersections
+  for (int i = 0; i < 4; i++) {
+    for (int j = 0; j < 4; j++) {
+      // Solve for 2x2 Ax=b
+      T det = cross_2d<T>(vec2[j], vec1[i]);
+
+      // This takes care of parallel lines
+      if (fabs(det) <= 1e-14) {
+        continue;
+      }
+
+      auto vec12 = pts2[j] - pts1[i];
+
+      T t1 = cross_2d<T>(vec2[j], vec12) / det;
+      T t2 = cross_2d<T>(vec1[i], vec12) / det;
+
+      if (t1 >= 0.0f && t1 <= 1.0f && t2 >= 0.0f && t2 <= 1.0f) {
+        intersections[num++] = pts1[i] + vec1[i] * t1;
+      }
+    }
+  }
+
+  // Check for vertices of rect1 inside rect2
+  {
+    const auto& AB = vec2[0];
+    const auto& DA = vec2[3];
+    auto ABdotAB = dot_2d<T>(AB, AB);
+    auto ADdotAD = dot_2d<T>(DA, DA);
+    for (int i = 0; i < 4; i++) {
+      // assume ABCD is the rectangle, and P is the point to be judged
+      // P is inside ABCD iff. P's projection on AB lies within AB
+      // and P's projection on AD lies within AD
+
+      auto AP = pts1[i] - pts2[0];
+
+      auto APdotAB = dot_2d<T>(AP, AB);
+      auto APdotAD = -dot_2d<T>(AP, DA);
+
+      if ((APdotAB >= 0) && (APdotAD >= 0) && (APdotAB <= ABdotAB) &&
+          (APdotAD <= ADdotAD)) {
+        intersections[num++] = pts1[i];
+      }
+    }
+  }
+
+  // Reverse the check - check for vertices of rect2 inside rect1
+  {
+    const auto& AB = vec1[0];
+    const auto& DA = vec1[3];
+    auto ABdotAB = dot_2d<T>(AB, AB);
+    auto ADdotAD = dot_2d<T>(DA, DA);
+    for (int i = 0; i < 4; i++) {
+      auto AP = pts2[i] - pts1[0];
+
+      auto APdotAB = dot_2d<T>(AP, AB);
+      auto APdotAD = -dot_2d<T>(AP, DA);
+
+      if ((APdotAB >= 0) && (APdotAD >= 0) && (APdotAB <= ABdotAB) &&
+          (APdotAD <= ADdotAD)) {
+        intersections[num++] = pts2[i];
+      }
+    }
+  }
+
+  return num;
+}
+
+template <typename T>
+HOST_DEVICE_INLINE int convex_hull_graham(
+    const Point<T> (&p)[24],
+    const int& num_in,
+    Point<T> (&q)[24],
+    bool shift_to_zero = false) {
+  assert(num_in >= 2);
+
+  // Step 1:
+  // Find point with minimum y
+  // if more than 1 points have the same minimum y,
+  // pick the one with the minimum x.
+  int t = 0;
+  for (int i = 1; i < num_in; i++) {
+    if (p[i].y < p[t].y || (p[i].y == p[t].y && p[i].x < p[t].x)) {
+      t = i;
+    }
+  }
+  auto& start = p[t]; // starting point
+
+  // Step 2:
+  // Subtract starting point from every points (for sorting in the next step)
+  for (int i = 0; i < num_in; i++) {
+    q[i] = p[i] - start;
+  }
+
+  // Swap the starting point to position 0
+  auto tmp = q[0];
+  q[0] = q[t];
+  q[t] = tmp;
+
+  // Step 3:
+  // Sort point 1 ~ num_in according to their relative cross-product values
+  // (essentially sorting according to angles)
+  // If the angles are the same, sort according to their distance to origin
+  T dist[24];
+#ifdef __CUDACC__
+  // compute distance to origin before sort, and sort them together with the
+  // points
+  for (int i = 0; i < num_in; i++) {
+    dist[i] = dot_2d<T>(q[i], q[i]);
+  }
+
+  // CUDA version
+  // In the future, we can potentially use thrust
+  // for sorting here to improve speed (though not guaranteed)
+  for (int i = 1; i < num_in - 1; i++) {
+    for (int j = i + 1; j < num_in; j++) {
+      T crossProduct = cross_2d<T>(q[i], q[j]);
+      if ((crossProduct < -1e-6) ||
+          (fabs(crossProduct) < 1e-6 && dist[i] > dist[j])) {
+        auto q_tmp = q[i];
+        q[i] = q[j];
+        q[j] = q_tmp;
+        auto dist_tmp = dist[i];
+        dist[i] = dist[j];
+        dist[j] = dist_tmp;
+      }
+    }
+  }
+#else
+  // CPU version
+  std::sort(
+      q + 1, q + num_in, [](const Point<T>& A, const Point<T>& B) -> bool {
+        T temp = cross_2d<T>(A, B);
+        if (fabs(temp) < 1e-6) {
+          return dot_2d<T>(A, A) < dot_2d<T>(B, B);
+        } else {
+          return temp > 0;
+        }
+      });
+  // compute distance to origin after sort, since the points are now different.
+  for (int i = 0; i < num_in; i++) {
+    dist[i] = dot_2d<T>(q[i], q[i]);
+  }
+#endif
+
+  // Step 4:
+  // Make sure there are at least 2 points (that don't overlap with each other)
+  // in the stack
+  int k; // index of the non-overlapped second point
+  for (k = 1; k < num_in; k++) {
+    if (dist[k] > 1e-8) {
+      break;
+    }
+  }
+  if (k == num_in) {
+    // We reach the end, which means the convex hull is just one point
+    q[0] = p[t];
+    return 1;
+  }
+  q[1] = q[k];
+  int m = 2; // 2 points in the stack
+  // Step 5:
+  // Finally we can start the scanning process.
+  // When a non-convex relationship between the 3 points is found
+  // (either concave shape or duplicated points),
+  // we pop the previous point from the stack
+  // until the 3-point relationship is convex again, or
+  // until the stack only contains two points
+  for (int i = k + 1; i < num_in; i++) {
+    while (m > 1) {
+      auto q1 = q[i] - q[m - 2], q2 = q[m - 1] - q[m - 2];
+      // cross_2d() uses FMA and therefore computes round(round(q1.x*q2.y) -
+      // q2.x*q1.y) So it may not return 0 even when q1==q2. Therefore we
+      // compare round(q1.x*q2.y) and round(q2.x*q1.y) directly. (round means
+      // round to nearest floating point).
+      if (q1.x * q2.y >= q2.x * q1.y)
+        m--;
+      else
+        break;
+    }
+    // Using double also helps, but float can solve the issue for now.
+    // while (m > 1 && cross_2d<T, double>(q[i] - q[m - 2], q[m - 1] - q[m - 2])
+    // >= 0) {
+    //     m--;
+    // }
+    q[m++] = q[i];
+  }
+
+  // Step 6 (Optional):
+  // In general sense we need the original coordinates, so we
+  // need to shift the points back (reverting Step 2)
+  // But if we're only interested in getting the area/perimeter of the shape
+  // We can simply return.
+  if (!shift_to_zero) {
+    for (int i = 0; i < m; i++) {
+      q[i] += start;
+    }
+  }
+
+  return m;
+}
+
+template <typename T>
+HOST_DEVICE_INLINE T polygon_area(const Point<T> (&q)[24], const int& m) {
+  if (m <= 2) {
+    return 0;
+  }
+
+  T area = 0;
+  for (int i = 1; i < m - 1; i++) {
+    area += fabs(cross_2d<T>(q[i] - q[0], q[i + 1] - q[0]));
+  }
+
+  return area / 2.0;
+}
+
+template <typename T>
+HOST_DEVICE_INLINE T rotated_boxes_intersection(
+    const RotatedBox<T>& box1,
+    const RotatedBox<T>& box2) {
+  // There are up to 4 x 4 + 4 + 4 = 24 intersections (including dups) returned
+  // from rotated_rect_intersection_pts
+  Point<T> intersectPts[24], orderedPts[24];
+
+  Point<T> pts1[4];
+  Point<T> pts2[4];
+  get_rotated_vertices<T>(box1, pts1);
+  get_rotated_vertices<T>(box2, pts2);
+
+  int num = get_intersection_points<T>(pts1, pts2, intersectPts);
+
+  if (num <= 2) {
+    return 0.0;
+  }
+
+  // Convex Hull to order the intersection points in clockwise order and find
+  // the contour area.
+  int num_convex = convex_hull_graham<T>(intersectPts, num, orderedPts, true);
+  return polygon_area<T>(orderedPts, num_convex);
+}
+
+} // namespace
+
+template <typename T>
+HOST_DEVICE_INLINE T
+single_box_iou_rotated(T const* const box1_raw, T const* const box2_raw) {
+  // shift center to the middle point to achieve higher precision in result
+  RotatedBox<T> box1, box2;
+  auto center_shift_x = (box1_raw[0] + box2_raw[0]) / 2.0;
+  auto center_shift_y = (box1_raw[1] + box2_raw[1]) / 2.0;
+  box1.x_ctr = box1_raw[0] - center_shift_x;
+  box1.y_ctr = box1_raw[1] - center_shift_y;
+  box1.w = box1_raw[2];
+  box1.h = box1_raw[3];
+  box1.a = box1_raw[4];
+  box2.x_ctr = box2_raw[0] - center_shift_x;
+  box2.y_ctr = box2_raw[1] - center_shift_y;
+  box2.w = box2_raw[2];
+  box2.h = box2_raw[3];
+  box2.a = box2_raw[4];
+
+  T area1 = box1.w * box1.h;
+  T area2 = box2.w * box2.h;
+  if (area1 < 1e-14 || area2 < 1e-14) {
+    return 0.f;
+  }
+
+  T intersection = rotated_boxes_intersection<T>(box1, box2);
+  T iou = intersection / (area1 + area2 - intersection);
+  return iou;
+}
+
+} // namespace detectron2