ootdiffusion

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

+# 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",
+]

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

+# 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)

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

+# 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.")

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

+# 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("_")]

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

+# -*- 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)

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

+# -*- 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)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/engine/__init__.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from .launch import *
+from .train_loop import *
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]
+
+
+# prefer to let hooks and defaults live in separate namespaces (therefore not in __all__)
+# but still make them available here
+from .hooks import *
+from .defaults import *

preprocess/humanparsing/mhp_extension/detectron2/detectron2/engine/defaults.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+"""
+This file contains components with some default boilerplate logic user may need
+in training / testing. They will not work for everyone, but many users may find them useful.
+
+The behavior of functions/classes in this file is subject to change,
+since they are meant to represent the "common default behavior" people need in their projects.
+"""
+
+import argparse
+import logging
+import os
+import sys
+from collections import OrderedDict
+import torch
+from fvcore.common.file_io import PathManager
+from fvcore.nn.precise_bn import get_bn_modules
+from torch.nn.parallel import DistributedDataParallel
+
+import detectron2.data.transforms as T
+from detectron2.checkpoint import DetectionCheckpointer
+from detectron2.data import (
+    MetadataCatalog,
+    build_detection_test_loader,
+    build_detection_train_loader,
+)
+from detectron2.evaluation import (
+    DatasetEvaluator,
+    inference_on_dataset,
+    print_csv_format,
+    verify_results,
+)
+from detectron2.modeling import build_model
+from detectron2.solver import build_lr_scheduler, build_optimizer
+from detectron2.utils import comm
+from detectron2.utils.collect_env import collect_env_info
+from detectron2.utils.env import seed_all_rng
+from detectron2.utils.events import CommonMetricPrinter, JSONWriter, TensorboardXWriter
+from detectron2.utils.logger import setup_logger
+
+from . import hooks
+from .train_loop import SimpleTrainer
+
+__all__ = ["default_argument_parser", "default_setup", "DefaultPredictor", "DefaultTrainer"]
+
+
+def default_argument_parser(epilog=None):
+    """
+    Create a parser with some common arguments used by detectron2 users.
+
+    Args:
+        epilog (str): epilog passed to ArgumentParser describing the usage.
+
+    Returns:
+        argparse.ArgumentParser:
+    """
+    parser = argparse.ArgumentParser(
+        epilog=epilog
+        or f"""
+Examples:
+
+Run on single machine:
+    $ {sys.argv[0]} --num-gpus 8 --config-file cfg.yaml MODEL.WEIGHTS /path/to/weight.pth
+
+Run on multiple machines:
+    (machine0)$ {sys.argv[0]} --machine-rank 0 --num-machines 2 --dist-url <URL> [--other-flags]
+    (machine1)$ {sys.argv[0]} --machine-rank 1 --num-machines 2 --dist-url <URL> [--other-flags]
+""",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument("--config-file", default="", metavar="FILE", help="path to config file")
+    parser.add_argument(
+        "--resume",
+        action="store_true",
+        help="whether to attempt to resume from the checkpoint directory",
+    )
+    parser.add_argument("--eval-only", action="store_true", help="perform evaluation only")
+    parser.add_argument("--num-gpus", type=int, default=1, help="number of gpus *per machine*")
+    parser.add_argument("--num-machines", type=int, default=1, help="total number of machines")
+    parser.add_argument(
+        "--machine-rank", type=int, default=0, help="the rank of this machine (unique per machine)"
+    )
+
+    # PyTorch still may leave orphan processes in multi-gpu training.
+    # Therefore we use a deterministic way to obtain port,
+    # so that users are aware of orphan processes by seeing the port occupied.
+    port = 2 ** 15 + 2 ** 14 + hash(os.getuid() if sys.platform != "win32" else 1) % 2 ** 14
+    parser.add_argument(
+        "--dist-url",
+        default="tcp://127.0.0.1:{}".format(port),
+        help="initialization URL for pytorch distributed backend. See "
+        "https://pytorch.org/docs/stable/distributed.html for details.",
+    )
+    parser.add_argument(
+        "opts",
+        help="Modify config options using the command-line",
+        default=None,
+        nargs=argparse.REMAINDER,
+    )
+    return parser
+
+
+def default_setup(cfg, args):
+    """
+    Perform some basic common setups at the beginning of a job, including:
+
+    1. Set up the detectron2 logger
+    2. Log basic information about environment, cmdline arguments, and config
+    3. Backup the config to the output directory
+
+    Args:
+        cfg (CfgNode): the full config to be used
+        args (argparse.NameSpace): the command line arguments to be logged
+    """
+    output_dir = cfg.OUTPUT_DIR
+    if comm.is_main_process() and output_dir:
+        PathManager.mkdirs(output_dir)
+
+    rank = comm.get_rank()
+    setup_logger(output_dir, distributed_rank=rank, name="fvcore")
+    logger = setup_logger(output_dir, distributed_rank=rank)
+
+    logger.info("Rank of current process: {}. World size: {}".format(rank, comm.get_world_size()))
+    logger.info("Environment info:\n" + collect_env_info())
+
+    logger.info("Command line arguments: " + str(args))
+    if hasattr(args, "config_file") and args.config_file != "":
+        logger.info(
+            "Contents of args.config_file={}:\n{}".format(
+                args.config_file, PathManager.open(args.config_file, "r").read()
+            )
+        )
+
+    logger.info("Running with full config:\n{}".format(cfg))
+    if comm.is_main_process() and output_dir:
+        # Note: some of our scripts may expect the existence of
+        # config.yaml in output directory
+        path = os.path.join(output_dir, "config.yaml")
+        with PathManager.open(path, "w") as f:
+            f.write(cfg.dump())
+        logger.info("Full config saved to {}".format(path))
+
+    # make sure each worker has a different, yet deterministic seed if specified
+    seed_all_rng(None if cfg.SEED < 0 else cfg.SEED + rank)
+
+    # cudnn benchmark has large overhead. It shouldn't be used considering the small size of
+    # typical validation set.
+    if not (hasattr(args, "eval_only") and args.eval_only):
+        torch.backends.cudnn.benchmark = cfg.CUDNN_BENCHMARK
+
+
+class DefaultPredictor:
+    """
+    Create a simple end-to-end predictor with the given config that runs on
+    single device for a single input image.
+
+    Compared to using the model directly, this class does the following additions:
+
+    1. Load checkpoint from `cfg.MODEL.WEIGHTS`.
+    2. Always take BGR image as the input and apply conversion defined by `cfg.INPUT.FORMAT`.
+    3. Apply resizing defined by `cfg.INPUT.{MIN,MAX}_SIZE_TEST`.
+    4. Take one input image and produce a single output, instead of a batch.
+
+    If you'd like to do anything more fancy, please refer to its source code
+    as examples to build and use the model manually.
+
+    Attributes:
+        metadata (Metadata): the metadata of the underlying dataset, obtained from
+            cfg.DATASETS.TEST.
+
+    Examples:
+
+    .. code-block:: python
+
+        pred = DefaultPredictor(cfg)
+        inputs = cv2.imread("input.jpg")
+        outputs = pred(inputs)
+    """
+
+    def __init__(self, cfg):
+        self.cfg = cfg.clone()  # cfg can be modified by model
+        self.model = build_model(self.cfg)
+        self.model.eval()
+        self.metadata = MetadataCatalog.get(cfg.DATASETS.TEST[0])
+
+        checkpointer = DetectionCheckpointer(self.model)
+        checkpointer.load(cfg.MODEL.WEIGHTS)
+
+        self.transform_gen = T.ResizeShortestEdge(
+            [cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST
+        )
+
+        self.input_format = cfg.INPUT.FORMAT
+        assert self.input_format in ["RGB", "BGR"], self.input_format
+
+    def __call__(self, original_image):
+        """
+        Args:
+            original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).
+
+        Returns:
+            predictions (dict):
+                the output of the model for one image only.
+                See :doc:`/tutorials/models` for details about the format.
+        """
+        with torch.no_grad():  # https://github.com/sphinx-doc/sphinx/issues/4258
+            # Apply pre-processing to image.
+            if self.input_format == "RGB":
+                # whether the model expects BGR inputs or RGB
+                original_image = original_image[:, :, ::-1]
+            height, width = original_image.shape[:2]
+            image = self.transform_gen.get_transform(original_image).apply_image(original_image)
+            image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
+
+            inputs = {"image": image, "height": height, "width": width}
+            predictions = self.model([inputs])[0]
+            return predictions
+
+
+class DefaultTrainer(SimpleTrainer):
+    """
+    A trainer with default training logic. Compared to `SimpleTrainer`, it
+    contains the following logic in addition:
+
+    1. Create model, optimizer, scheduler, dataloader from the given config.
+    2. Load a checkpoint or `cfg.MODEL.WEIGHTS`, if exists, when
+       `resume_or_load` is called.
+    3. Register a few common hooks.
+
+    It is created to simplify the **standard model training workflow** and reduce code boilerplate
+    for users who only need the standard training workflow, with standard features.
+    It means this class makes *many assumptions* about your training logic that
+    may easily become invalid in a new research. In fact, any assumptions beyond those made in the
+    :class:`SimpleTrainer` are too much for research.
+
+    The code of this class has been annotated about restrictive assumptions it mades.
+    When they do not work for you, you're encouraged to:
+
+    1. Overwrite methods of this class, OR:
+    2. Use :class:`SimpleTrainer`, which only does minimal SGD training and
+       nothing else. You can then add your own hooks if needed. OR:
+    3. Write your own training loop similar to `tools/plain_train_net.py`.
+
+    Also note that the behavior of this class, like other functions/classes in
+    this file, is not stable, since it is meant to represent the "common default behavior".
+    It is only guaranteed to work well with the standard models and training workflow in detectron2.
+    To obtain more stable behavior, write your own training logic with other public APIs.
+
+    Examples:
+
+    .. code-block:: python
+
+        trainer = DefaultTrainer(cfg)
+        trainer.resume_or_load()  # load last checkpoint or MODEL.WEIGHTS
+        trainer.train()
+
+    Attributes:
+        scheduler:
+        checkpointer (DetectionCheckpointer):
+        cfg (CfgNode):
+    """
+
+    def __init__(self, cfg):
+        """
+        Args:
+            cfg (CfgNode):
+        """
+        logger = logging.getLogger("detectron2")
+        if not logger.isEnabledFor(logging.INFO):  # setup_logger is not called for d2
+            setup_logger()
+        # Assume these objects must be constructed in this order.
+        model = self.build_model(cfg)
+        optimizer = self.build_optimizer(cfg, model)
+        data_loader = self.build_train_loader(cfg)
+
+        # For training, wrap with DDP. But don't need this for inference.
+        if comm.get_world_size() > 1:
+            model = DistributedDataParallel(
+                model, device_ids=[comm.get_local_rank()], broadcast_buffers=False
+            )
+        super().__init__(model, data_loader, optimizer)
+
+        self.scheduler = self.build_lr_scheduler(cfg, optimizer)
+        # Assume no other objects need to be checkpointed.
+        # We can later make it checkpoint the stateful hooks
+        self.checkpointer = DetectionCheckpointer(
+            # Assume you want to save checkpoints together with logs/statistics
+            model,
+            cfg.OUTPUT_DIR,
+            optimizer=optimizer,
+            scheduler=self.scheduler,
+        )
+        self.start_iter = 0
+        self.max_iter = cfg.SOLVER.MAX_ITER
+        self.cfg = cfg
+
+        self.register_hooks(self.build_hooks())
+
+    def resume_or_load(self, resume=True):
+        """
+        If `resume==True`, and last checkpoint exists, resume from it and load all
+        checkpointables (eg. optimizer and scheduler).
+
+        Otherwise, load the model specified by the config (skip all checkpointables).
+
+        Args:
+            resume (bool): whether to do resume or not
+        """
+        checkpoint = self.checkpointer.resume_or_load(self.cfg.MODEL.WEIGHTS, resume=resume)
+        self.start_iter = checkpoint.get("iteration", -1) if resume else -1
+        # The checkpoint stores the training iteration that just finished, thus we start
+        # at the next iteration (or iter zero if there's no checkpoint).
+        self.start_iter += 1
+
+    def build_hooks(self):
+        """
+        Build a list of default hooks, including timing, evaluation,
+        checkpointing, lr scheduling, precise BN, writing events.
+
+        Returns:
+            list[HookBase]:
+        """
+        cfg = self.cfg.clone()
+        cfg.defrost()
+        cfg.DATALOADER.NUM_WORKERS = 0  # save some memory and time for PreciseBN
+
+        ret = [
+            hooks.IterationTimer(),
+            hooks.LRScheduler(self.optimizer, self.scheduler),
+            hooks.PreciseBN(
+                # Run at the same freq as (but before) evaluation.
+                cfg.TEST.EVAL_PERIOD,
+                self.model,
+                # Build a new data loader to not affect training
+                self.build_train_loader(cfg),
+                cfg.TEST.PRECISE_BN.NUM_ITER,
+            )
+            if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model)
+            else None,
+        ]
+
+        # Do PreciseBN before checkpointer, because it updates the model and need to
+        # be saved by checkpointer.
+        # This is not always the best: if checkpointing has a different frequency,
+        # some checkpoints may have more precise statistics than others.
+        if comm.is_main_process():
+            ret.append(hooks.PeriodicCheckpointer(self.checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD))
+
+        def test_and_save_results():
+            self._last_eval_results = self.test(self.cfg, self.model)
+            return self._last_eval_results
+
+        # Do evaluation after checkpointer, because then if it fails,
+        # we can use the saved checkpoint to debug.
+        ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results))
+
+        if comm.is_main_process():
+            # run writers in the end, so that evaluation metrics are written
+            ret.append(hooks.PeriodicWriter(self.build_writers(), period=20))
+        return ret
+
+    def build_writers(self):
+        """
+        Build a list of writers to be used. By default it contains
+        writers that write metrics to the screen,
+        a json file, and a tensorboard event file respectively.
+        If you'd like a different list of writers, you can overwrite it in
+        your trainer.
+
+        Returns:
+            list[EventWriter]: a list of :class:`EventWriter` objects.
+
+        It is now implemented by:
+
+        .. code-block:: python
+
+            return [
+                CommonMetricPrinter(self.max_iter),
+                JSONWriter(os.path.join(self.cfg.OUTPUT_DIR, "metrics.json")),
+                TensorboardXWriter(self.cfg.OUTPUT_DIR),
+            ]
+
+        """
+        # Here the default print/log frequency of each writer is used.
+        return [
+            # It may not always print what you want to see, since it prints "common" metrics only.
+            CommonMetricPrinter(self.max_iter),
+            JSONWriter(os.path.join(self.cfg.OUTPUT_DIR, "metrics.json")),
+            TensorboardXWriter(self.cfg.OUTPUT_DIR),
+        ]
+
+    def train(self):
+        """
+        Run training.
+
+        Returns:
+            OrderedDict of results, if evaluation is enabled. Otherwise None.
+        """
+        super().train(self.start_iter, self.max_iter)
+        if len(self.cfg.TEST.EXPECTED_RESULTS) and comm.is_main_process():
+            assert hasattr(
+                self, "_last_eval_results"
+            ), "No evaluation results obtained during training!"
+            verify_results(self.cfg, self._last_eval_results)
+            return self._last_eval_results
+
+    @classmethod
+    def build_model(cls, cfg):
+        """
+        Returns:
+            torch.nn.Module:
+
+        It now calls :func:`detectron2.modeling.build_model`.
+        Overwrite it if you'd like a different model.
+        """
+        model = build_model(cfg)
+        logger = logging.getLogger(__name__)
+        logger.info("Model:\n{}".format(model))
+        return model
+
+    @classmethod
+    def build_optimizer(cls, cfg, model):
+        """
+        Returns:
+            torch.optim.Optimizer:
+
+        It now calls :func:`detectron2.solver.build_optimizer`.
+        Overwrite it if you'd like a different optimizer.
+        """
+        return build_optimizer(cfg, model)
+
+    @classmethod
+    def build_lr_scheduler(cls, cfg, optimizer):
+        """
+        It now calls :func:`detectron2.solver.build_lr_scheduler`.
+        Overwrite it if you'd like a different scheduler.
+        """
+        return build_lr_scheduler(cfg, optimizer)
+
+    @classmethod
+    def build_train_loader(cls, cfg):
+        """
+        Returns:
+            iterable
+
+        It now calls :func:`detectron2.data.build_detection_train_loader`.
+        Overwrite it if you'd like a different data loader.
+        """
+        return build_detection_train_loader(cfg)
+
+    @classmethod
+    def build_test_loader(cls, cfg, dataset_name):
+        """
+        Returns:
+            iterable
+
+        It now calls :func:`detectron2.data.build_detection_test_loader`.
+        Overwrite it if you'd like a different data loader.
+        """
+        return build_detection_test_loader(cfg, dataset_name)
+
+    @classmethod
+    def build_evaluator(cls, cfg, dataset_name):
+        """
+        Returns:
+            DatasetEvaluator or None
+
+        It is not implemented by default.
+        """
+        raise NotImplementedError(
+            """
+If you want DefaultTrainer to automatically run evaluation,
+please implement `build_evaluator()` in subclasses (see train_net.py for example).
+Alternatively, you can call evaluation functions yourself (see Colab balloon tutorial for example).
+"""
+        )
+
+    @classmethod
+    def test(cls, cfg, model, evaluators=None):
+        """
+        Args:
+            cfg (CfgNode):
+            model (nn.Module):
+            evaluators (list[DatasetEvaluator] or None): if None, will call
+                :meth:`build_evaluator`. Otherwise, must have the same length as
+                `cfg.DATASETS.TEST`.
+
+        Returns:
+            dict: a dict of result metrics
+        """
+        logger = logging.getLogger(__name__)
+        if isinstance(evaluators, DatasetEvaluator):
+            evaluators = [evaluators]
+        if evaluators is not None:
+            assert len(cfg.DATASETS.TEST) == len(evaluators), "{} != {}".format(
+                len(cfg.DATASETS.TEST), len(evaluators)
+            )
+
+        results = OrderedDict()
+        for idx, dataset_name in enumerate(cfg.DATASETS.TEST):
+            data_loader = cls.build_test_loader(cfg, dataset_name)
+            # When evaluators are passed in as arguments,
+            # implicitly assume that evaluators can be created before data_loader.
+            if evaluators is not None:
+                evaluator = evaluators[idx]
+            else:
+                try:
+                    evaluator = cls.build_evaluator(cfg, dataset_name)
+                except NotImplementedError:
+                    logger.warn(
+                        "No evaluator found. Use `DefaultTrainer.test(evaluators=)`, "
+                        "or implement its `build_evaluator` method."
+                    )
+                    results[dataset_name] = {}
+                    continue
+            results_i = inference_on_dataset(model, data_loader, evaluator)
+            results[dataset_name] = results_i
+            if comm.is_main_process():
+                assert isinstance(
+                    results_i, dict
+                ), "Evaluator must return a dict on the main process. Got {} instead.".format(
+                    results_i
+                )
+                logger.info("Evaluation results for {} in csv format:".format(dataset_name))
+                print_csv_format(results_i)
+
+        if len(results) == 1:
+            results = list(results.values())[0]
+        return results

preprocess/humanparsing/mhp_extension/detectron2/detectron2/engine/hooks.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import datetime
+import itertools
+import logging
+import os
+import tempfile
+import time
+from collections import Counter
+import torch
+from fvcore.common.checkpoint import PeriodicCheckpointer as _PeriodicCheckpointer
+from fvcore.common.file_io import PathManager
+from fvcore.common.timer import Timer
+from fvcore.nn.precise_bn import get_bn_modules, update_bn_stats
+
+import detectron2.utils.comm as comm
+from detectron2.evaluation.testing import flatten_results_dict
+from detectron2.utils.events import EventStorage, EventWriter
+
+from .train_loop import HookBase
+
+__all__ = [
+    "CallbackHook",
+    "IterationTimer",
+    "PeriodicWriter",
+    "PeriodicCheckpointer",
+    "LRScheduler",
+    "AutogradProfiler",
+    "EvalHook",
+    "PreciseBN",
+]
+
+
+"""
+Implement some common hooks.
+"""
+
+
+class CallbackHook(HookBase):
+    """
+    Create a hook using callback functions provided by the user.
+    """
+
+    def __init__(self, *, before_train=None, after_train=None, before_step=None, after_step=None):
+        """
+        Each argument is a function that takes one argument: the trainer.
+        """
+        self._before_train = before_train
+        self._before_step = before_step
+        self._after_step = after_step
+        self._after_train = after_train
+
+    def before_train(self):
+        if self._before_train:
+            self._before_train(self.trainer)
+
+    def after_train(self):
+        if self._after_train:
+            self._after_train(self.trainer)
+        # The functions may be closures that hold reference to the trainer
+        # Therefore, delete them to avoid circular reference.
+        del self._before_train, self._after_train
+        del self._before_step, self._after_step
+
+    def before_step(self):
+        if self._before_step:
+            self._before_step(self.trainer)
+
+    def after_step(self):
+        if self._after_step:
+            self._after_step(self.trainer)
+
+
+class IterationTimer(HookBase):
+    """
+    Track the time spent for each iteration (each run_step call in the trainer).
+    Print a summary in the end of training.
+
+    This hook uses the time between the call to its :meth:`before_step`
+    and :meth:`after_step` methods.
+    Under the convention that :meth:`before_step` of all hooks should only
+    take negligible amount of time, the :class:`IterationTimer` hook should be
+    placed at the beginning of the list of hooks to obtain accurate timing.
+    """
+
+    def __init__(self, warmup_iter=3):
+        """
+        Args:
+            warmup_iter (int): the number of iterations at the beginning to exclude
+                from timing.
+        """
+        self._warmup_iter = warmup_iter
+        self._step_timer = Timer()
+        self._start_time = time.perf_counter()
+        self._total_timer = Timer()
+
+    def before_train(self):
+        self._start_time = time.perf_counter()
+        self._total_timer.reset()
+        self._total_timer.pause()
+
+    def after_train(self):
+        logger = logging.getLogger(__name__)
+        total_time = time.perf_counter() - self._start_time
+        total_time_minus_hooks = self._total_timer.seconds()
+        hook_time = total_time - total_time_minus_hooks
+
+        num_iter = self.trainer.iter + 1 - self.trainer.start_iter - self._warmup_iter
+
+        if num_iter > 0 and total_time_minus_hooks > 0:
+            # Speed is meaningful only after warmup
+            # NOTE this format is parsed by grep in some scripts
+            logger.info(
+                "Overall training speed: {} iterations in {} ({:.4f} s / it)".format(
+                    num_iter,
+                    str(datetime.timedelta(seconds=int(total_time_minus_hooks))),
+                    total_time_minus_hooks / num_iter,
+                )
+            )
+
+        logger.info(
+            "Total training time: {} ({} on hooks)".format(
+                str(datetime.timedelta(seconds=int(total_time))),
+                str(datetime.timedelta(seconds=int(hook_time))),
+            )
+        )
+
+    def before_step(self):
+        self._step_timer.reset()
+        self._total_timer.resume()
+
+    def after_step(self):
+        # +1 because we're in after_step
+        iter_done = self.trainer.iter - self.trainer.start_iter + 1
+        if iter_done >= self._warmup_iter:
+            sec = self._step_timer.seconds()
+            self.trainer.storage.put_scalars(time=sec)
+        else:
+            self._start_time = time.perf_counter()
+            self._total_timer.reset()
+
+        self._total_timer.pause()
+
+
+class PeriodicWriter(HookBase):
+    """
+    Write events to EventStorage periodically.
+
+    It is executed every ``period`` iterations and after the last iteration.
+    """
+
+    def __init__(self, writers, period=20):
+        """
+        Args:
+            writers (list[EventWriter]): a list of EventWriter objects
+            period (int):
+        """
+        self._writers = writers
+        for w in writers:
+            assert isinstance(w, EventWriter), w
+        self._period = period
+
+    def after_step(self):
+        if (self.trainer.iter + 1) % self._period == 0 or (
+            self.trainer.iter == self.trainer.max_iter - 1
+        ):
+            for writer in self._writers:
+                writer.write()
+
+    def after_train(self):
+        for writer in self._writers:
+            writer.close()
+
+
+class PeriodicCheckpointer(_PeriodicCheckpointer, HookBase):
+    """
+    Same as :class:`detectron2.checkpoint.PeriodicCheckpointer`, but as a hook.
+
+    Note that when used as a hook,
+    it is unable to save additional data other than what's defined
+    by the given `checkpointer`.
+
+    It is executed every ``period`` iterations and after the last iteration.
+    """
+
+    def before_train(self):
+        self.max_iter = self.trainer.max_iter
+
+    def after_step(self):
+        # No way to use **kwargs
+        self.step(self.trainer.iter)
+
+
+class LRScheduler(HookBase):
+    """
+    A hook which executes a torch builtin LR scheduler and summarizes the LR.
+    It is executed after every iteration.
+    """
+
+    def __init__(self, optimizer, scheduler):
+        """
+        Args:
+            optimizer (torch.optim.Optimizer):
+            scheduler (torch.optim._LRScheduler)
+        """
+        self._optimizer = optimizer
+        self._scheduler = scheduler
+
+        # NOTE: some heuristics on what LR to summarize
+        # summarize the param group with most parameters
+        largest_group = max(len(g["params"]) for g in optimizer.param_groups)
+
+        if largest_group == 1:
+            # If all groups have one parameter,
+            # then find the most common initial LR, and use it for summary
+            lr_count = Counter([g["lr"] for g in optimizer.param_groups])
+            lr = lr_count.most_common()[0][0]
+            for i, g in enumerate(optimizer.param_groups):
+                if g["lr"] == lr:
+                    self._best_param_group_id = i
+                    break
+        else:
+            for i, g in enumerate(optimizer.param_groups):
+                if len(g["params"]) == largest_group:
+                    self._best_param_group_id = i
+                    break
+
+    def after_step(self):
+        lr = self._optimizer.param_groups[self._best_param_group_id]["lr"]
+        self.trainer.storage.put_scalar("lr", lr, smoothing_hint=False)
+        self._scheduler.step()
+
+
+class AutogradProfiler(HookBase):
+    """
+    A hook which runs `torch.autograd.profiler.profile`.
+
+    Examples:
+
+    .. code-block:: python
+
+        hooks.AutogradProfiler(
+             lambda trainer: trainer.iter > 10 and trainer.iter < 20, self.cfg.OUTPUT_DIR
+        )
+
+    The above example will run the profiler for iteration 10~20 and dump
+    results to ``OUTPUT_DIR``. We did not profile the first few iterations
+    because they are typically slower than the rest.
+    The result files can be loaded in the ``chrome://tracing`` page in chrome browser.
+
+    Note:
+        When used together with NCCL on older version of GPUs,
+        autograd profiler may cause deadlock because it unnecessarily allocates
+        memory on every device it sees. The memory management calls, if
+        interleaved with NCCL calls, lead to deadlock on GPUs that do not
+        support `cudaLaunchCooperativeKernelMultiDevice`.
+    """
+
+    def __init__(self, enable_predicate, output_dir, *, use_cuda=True):
+        """
+        Args:
+            enable_predicate (callable[trainer -> bool]): a function which takes a trainer,
+                and returns whether to enable the profiler.
+                It will be called once every step, and can be used to select which steps to profile.
+            output_dir (str): the output directory to dump tracing files.
+            use_cuda (bool): same as in `torch.autograd.profiler.profile`.
+        """
+        self._enable_predicate = enable_predicate
+        self._use_cuda = use_cuda
+        self._output_dir = output_dir
+
+    def before_step(self):
+        if self._enable_predicate(self.trainer):
+            self._profiler = torch.autograd.profiler.profile(use_cuda=self._use_cuda)
+            self._profiler.__enter__()
+        else:
+            self._profiler = None
+
+    def after_step(self):
+        if self._profiler is None:
+            return
+        self._profiler.__exit__(None, None, None)
+        PathManager.mkdirs(self._output_dir)
+        out_file = os.path.join(
+            self._output_dir, "profiler-trace-iter{}.json".format(self.trainer.iter)
+        )
+        if "://" not in out_file:
+            self._profiler.export_chrome_trace(out_file)
+        else:
+            # Support non-posix filesystems
+            with tempfile.TemporaryDirectory(prefix="detectron2_profiler") as d:
+                tmp_file = os.path.join(d, "tmp.json")
+                self._profiler.export_chrome_trace(tmp_file)
+                with open(tmp_file) as f:
+                    content = f.read()
+            with PathManager.open(out_file, "w") as f:
+                f.write(content)
+
+
+class EvalHook(HookBase):
+    """
+    Run an evaluation function periodically, and at the end of training.
+
+    It is executed every ``eval_period`` iterations and after the last iteration.
+    """
+
+    def __init__(self, eval_period, eval_function):
+        """
+        Args:
+            eval_period (int): the period to run `eval_function`.
+            eval_function (callable): a function which takes no arguments, and
+                returns a nested dict of evaluation metrics.
+
+        Note:
+            This hook must be enabled in all or none workers.
+            If you would like only certain workers to perform evaluation,
+            give other workers a no-op function (`eval_function=lambda: None`).
+        """
+        self._period = eval_period
+        self._func = eval_function
+
+    def _do_eval(self):
+        results = self._func()
+
+        if results:
+            assert isinstance(
+                results, dict
+            ), "Eval function must return a dict. Got {} instead.".format(results)
+
+            flattened_results = flatten_results_dict(results)
+            for k, v in flattened_results.items():
+                try:
+                    v = float(v)
+                except Exception:
+                    raise ValueError(
+                        "[EvalHook] eval_function should return a nested dict of float. "
+                        "Got '{}: {}' instead.".format(k, v)
+                    )
+            self.trainer.storage.put_scalars(**flattened_results, smoothing_hint=False)
+
+        # Evaluation may take different time among workers.
+        # A barrier make them start the next iteration together.
+        comm.synchronize()
+
+    def after_step(self):
+        next_iter = self.trainer.iter + 1
+        is_final = next_iter == self.trainer.max_iter
+        if is_final or (self._period > 0 and next_iter % self._period == 0):
+            self._do_eval()
+
+    def after_train(self):
+        # func is likely a closure that holds reference to the trainer
+        # therefore we clean it to avoid circular reference in the end
+        del self._func
+
+
+class PreciseBN(HookBase):
+    """
+    The standard implementation of BatchNorm uses EMA in inference, which is
+    sometimes suboptimal.
+    This class computes the true average of statistics rather than the moving average,
+    and put true averages to every BN layer in the given model.
+
+    It is executed every ``period`` iterations and after the last iteration.
+    """
+
+    def __init__(self, period, model, data_loader, num_iter):
+        """
+        Args:
+            period (int): the period this hook is run, or 0 to not run during training.
+                The hook will always run in the end of training.
+            model (nn.Module): a module whose all BN layers in training mode will be
+                updated by precise BN.
+                Note that user is responsible for ensuring the BN layers to be
+                updated are in training mode when this hook is triggered.
+            data_loader (iterable): it will produce data to be run by `model(data)`.
+            num_iter (int): number of iterations used to compute the precise
+                statistics.
+        """
+        self._logger = logging.getLogger(__name__)
+        if len(get_bn_modules(model)) == 0:
+            self._logger.info(
+                "PreciseBN is disabled because model does not contain BN layers in training mode."
+            )
+            self._disabled = True
+            return
+
+        self._model = model
+        self._data_loader = data_loader
+        self._num_iter = num_iter
+        self._period = period
+        self._disabled = False
+
+        self._data_iter = None
+
+    def after_step(self):
+        next_iter = self.trainer.iter + 1
+        is_final = next_iter == self.trainer.max_iter
+        if is_final or (self._period > 0 and next_iter % self._period == 0):
+            self.update_stats()
+
+    def update_stats(self):
+        """
+        Update the model with precise statistics. Users can manually call this method.
+        """
+        if self._disabled:
+            return
+
+        if self._data_iter is None:
+            self._data_iter = iter(self._data_loader)
+
+        def data_loader():
+            for num_iter in itertools.count(1):
+                if num_iter % 100 == 0:
+                    self._logger.info(
+                        "Running precise-BN ... {}/{} iterations.".format(num_iter, self._num_iter)
+                    )
+                # This way we can reuse the same iterator
+                yield next(self._data_iter)
+
+        with EventStorage():  # capture events in a new storage to discard them
+            self._logger.info(
+                "Running precise-BN for {} iterations...  ".format(self._num_iter)
+                + "Note that this could produce different statistics every time."
+            )
+            update_bn_stats(self._model, data_loader(), self._num_iter)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/engine/launch.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+
+from detectron2.utils import comm
+
+__all__ = ["launch"]
+
+
+def _find_free_port():
+    import socket
+
+    sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    # Binding to port 0 will cause the OS to find an available port for us
+    sock.bind(("", 0))
+    port = sock.getsockname()[1]
+    sock.close()
+    # NOTE: there is still a chance the port could be taken by other processes.
+    return port
+
+
+def launch(main_func, num_gpus_per_machine, num_machines=1, machine_rank=0, dist_url=None, args=()):
+    """
+    Args:
+        main_func: a function that will be called by `main_func(*args)`
+        num_machines (int): the total number of machines
+        machine_rank (int): the rank of this machine (one per machine)
+        dist_url (str): url to connect to for distributed jobs, including protocol
+                       e.g. "tcp://127.0.0.1:8686".
+                       Can be set to "auto" to automatically select a free port on localhost
+        args (tuple): arguments passed to main_func
+    """
+    world_size = num_machines * num_gpus_per_machine
+    if world_size > 1:
+        # https://github.com/pytorch/pytorch/pull/14391
+        # TODO prctl in spawned processes
+
+        if dist_url == "auto":
+            assert num_machines == 1, "dist_url=auto not supported in multi-machine jobs."
+            port = _find_free_port()
+            dist_url = f"tcp://127.0.0.1:{port}"
+        if num_machines > 1 and dist_url.startswith("file://"):
+            logger = logging.getLogger(__name__)
+            logger.warning(
+                "file:// is not a reliable init_method in multi-machine jobs. Prefer tcp://"
+            )
+
+        mp.spawn(
+            _distributed_worker,
+            nprocs=num_gpus_per_machine,
+            args=(main_func, world_size, num_gpus_per_machine, machine_rank, dist_url, args),
+            daemon=False,
+        )
+    else:
+        main_func(*args)
+
+
+def _distributed_worker(
+    local_rank, main_func, world_size, num_gpus_per_machine, machine_rank, dist_url, args
+):
+    assert torch.cuda.is_available(), "cuda is not available. Please check your installation."
+    global_rank = machine_rank * num_gpus_per_machine + local_rank
+    try:
+        dist.init_process_group(
+            backend="NCCL", init_method=dist_url, world_size=world_size, rank=global_rank
+        )
+    except Exception as e:
+        logger = logging.getLogger(__name__)
+        logger.error("Process group URL: {}".format(dist_url))
+        raise e
+    # synchronize is needed here to prevent a possible timeout after calling init_process_group
+    # See: https://github.com/facebookresearch/maskrcnn-benchmark/issues/172
+    comm.synchronize()
+
+    assert num_gpus_per_machine <= torch.cuda.device_count()
+    torch.cuda.set_device(local_rank)
+
+    # Setup the local process group (which contains ranks within the same machine)
+    assert comm._LOCAL_PROCESS_GROUP is None
+    num_machines = world_size // num_gpus_per_machine
+    for i in range(num_machines):
+        ranks_on_i = list(range(i * num_gpus_per_machine, (i + 1) * num_gpus_per_machine))
+        pg = dist.new_group(ranks_on_i)
+        if i == machine_rank:
+            comm._LOCAL_PROCESS_GROUP = pg
+
+    main_func(*args)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/engine/train_loop.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import logging
+import numpy as np
+import time
+import weakref
+import torch
+
+import detectron2.utils.comm as comm
+from detectron2.utils.events import EventStorage
+
+__all__ = ["HookBase", "TrainerBase", "SimpleTrainer"]
+
+
+class HookBase:
+    """
+    Base class for hooks that can be registered with :class:`TrainerBase`.
+
+    Each hook can implement 4 methods. The way they are called is demonstrated
+    in the following snippet:
+
+    .. code-block:: python
+
+        hook.before_train()
+        for iter in range(start_iter, max_iter):
+            hook.before_step()
+            trainer.run_step()
+            hook.after_step()
+        hook.after_train()
+
+    Notes:
+        1. In the hook method, users can access `self.trainer` to access more
+           properties about the context (e.g., current iteration).
+
+        2. A hook that does something in :meth:`before_step` can often be
+           implemented equivalently in :meth:`after_step`.
+           If the hook takes non-trivial time, it is strongly recommended to
+           implement the hook in :meth:`after_step` instead of :meth:`before_step`.
+           The convention is that :meth:`before_step` should only take negligible time.
+
+           Following this convention will allow hooks that do care about the difference
+           between :meth:`before_step` and :meth:`after_step` (e.g., timer) to
+           function properly.
+
+    Attributes:
+        trainer: A weak reference to the trainer object. Set by the trainer when the hook is
+            registered.
+    """
+
+    def before_train(self):
+        """
+        Called before the first iteration.
+        """
+        pass
+
+    def after_train(self):
+        """
+        Called after the last iteration.
+        """
+        pass
+
+    def before_step(self):
+        """
+        Called before each iteration.
+        """
+        pass
+
+    def after_step(self):
+        """
+        Called after each iteration.
+        """
+        pass
+
+
+class TrainerBase:
+    """
+    Base class for iterative trainer with hooks.
+
+    The only assumption we made here is: the training runs in a loop.
+    A subclass can implement what the loop is.
+    We made no assumptions about the existence of dataloader, optimizer, model, etc.
+
+    Attributes:
+        iter(int): the current iteration.
+
+        start_iter(int): The iteration to start with.
+            By convention the minimum possible value is 0.
+
+        max_iter(int): The iteration to end training.
+
+        storage(EventStorage): An EventStorage that's opened during the course of training.
+    """
+
+    def __init__(self):
+        self._hooks = []
+
+    def register_hooks(self, hooks):
+        """
+        Register hooks to the trainer. The hooks are executed in the order
+        they are registered.
+
+        Args:
+            hooks (list[Optional[HookBase]]): list of hooks
+        """
+        hooks = [h for h in hooks if h is not None]
+        for h in hooks:
+            assert isinstance(h, HookBase)
+            # To avoid circular reference, hooks and trainer cannot own each other.
+            # This normally does not matter, but will cause memory leak if the
+            # involved objects contain __del__:
+            # See http://engineering.hearsaysocial.com/2013/06/16/circular-references-in-python/
+            h.trainer = weakref.proxy(self)
+        self._hooks.extend(hooks)
+
+    def train(self, start_iter: int, max_iter: int):
+        """
+        Args:
+            start_iter, max_iter (int): See docs above
+        """
+        logger = logging.getLogger(__name__)
+        logger.info("Starting training from iteration {}".format(start_iter))
+
+        self.iter = self.start_iter = start_iter
+        self.max_iter = max_iter
+
+        with EventStorage(start_iter) as self.storage:
+            try:
+                self.before_train()
+                for self.iter in range(start_iter, max_iter):
+                    self.before_step()
+                    self.run_step()
+                    self.after_step()
+            except Exception:
+                logger.exception("Exception during training:")
+                raise
+            finally:
+                self.after_train()
+
+    def before_train(self):
+        for h in self._hooks:
+            h.before_train()
+
+    def after_train(self):
+        for h in self._hooks:
+            h.after_train()
+
+    def before_step(self):
+        for h in self._hooks:
+            h.before_step()
+
+    def after_step(self):
+        for h in self._hooks:
+            h.after_step()
+        # this guarantees, that in each hook's after_step, storage.iter == trainer.iter
+        self.storage.step()
+
+    def run_step(self):
+        raise NotImplementedError
+
+
+class SimpleTrainer(TrainerBase):
+    """
+    A simple trainer for the most common type of task:
+    single-cost single-optimizer single-data-source iterative optimization.
+    It assumes that every step, you:
+
+    1. Compute the loss with a data from the data_loader.
+    2. Compute the gradients with the above loss.
+    3. Update the model with the optimizer.
+
+    If you want to do anything fancier than this,
+    either subclass TrainerBase and implement your own `run_step`,
+    or write your own training loop.
+    """
+
+    def __init__(self, model, data_loader, optimizer):
+        """
+        Args:
+            model: a torch Module. Takes a data from data_loader and returns a
+                dict of losses.
+            data_loader: an iterable. Contains data to be used to call model.
+            optimizer: a torch optimizer.
+        """
+        super().__init__()
+
+        """
+        We set the model to training mode in the trainer.
+        However it's valid to train a model that's in eval mode.
+        If you want your model (or a submodule of it) to behave
+        like evaluation during training, you can overwrite its train() method.
+        """
+        model.train()
+
+        self.model = model
+        self.data_loader = data_loader
+        self._data_loader_iter = iter(data_loader)
+        self.optimizer = optimizer
+
+    def run_step(self):
+        """
+        Implement the standard training logic described above.
+        """
+        assert self.model.training, "[SimpleTrainer] model was changed to eval mode!"
+        start = time.perf_counter()
+        """
+        If you want to do something with the data, you can wrap the dataloader.
+        """
+        data = next(self._data_loader_iter)
+        data_time = time.perf_counter() - start
+
+        """
+        If you want to do something with the losses, you can wrap the model.
+        """
+        loss_dict = self.model(data)
+        losses = sum(loss_dict.values())
+        self._detect_anomaly(losses, loss_dict)
+
+        metrics_dict = loss_dict
+        metrics_dict["data_time"] = data_time
+        self._write_metrics(metrics_dict)
+
+        """
+        If you need to accumulate gradients or something similar, you can
+        wrap the optimizer with your custom `zero_grad()` method.
+        """
+        self.optimizer.zero_grad()
+        losses.backward()
+
+        """
+        If you need gradient clipping/scaling or other processing, you can
+        wrap the optimizer with your custom `step()` method.
+        """
+        self.optimizer.step()
+
+    def _detect_anomaly(self, losses, loss_dict):
+        if not torch.isfinite(losses).all():
+            raise FloatingPointError(
+                "Loss became infinite or NaN at iteration={}!\nloss_dict = {}".format(
+                    self.iter, loss_dict
+                )
+            )
+
+    def _write_metrics(self, metrics_dict: dict):
+        """
+        Args:
+            metrics_dict (dict): dict of scalar metrics
+        """
+        metrics_dict = {
+            k: v.detach().cpu().item() if isinstance(v, torch.Tensor) else float(v)
+            for k, v in metrics_dict.items()
+        }
+        # gather metrics among all workers for logging
+        # This assumes we do DDP-style training, which is currently the only
+        # supported method in detectron2.
+        all_metrics_dict = comm.gather(metrics_dict)
+
+        if comm.is_main_process():
+            if "data_time" in all_metrics_dict[0]:
+                # data_time among workers can have high variance. The actual latency
+                # caused by data_time is the maximum among workers.
+                data_time = np.max([x.pop("data_time") for x in all_metrics_dict])
+                self.storage.put_scalar("data_time", data_time)
+
+            # average the rest metrics
+            metrics_dict = {
+                k: np.mean([x[k] for x in all_metrics_dict]) for k in all_metrics_dict[0].keys()
+            }
+            total_losses_reduced = sum(loss for loss in metrics_dict.values())
+
+            self.storage.put_scalar("total_loss", total_losses_reduced)
+            if len(metrics_dict) > 1:
+                self.storage.put_scalars(**metrics_dict)

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

+# 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("_")]

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

+# 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

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

+# 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

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

+# 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)

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

+# 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

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

+# 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)

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

+# -*- 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

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

+# 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

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

+# 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