ootdiffusion

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/proposal_generator/build.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from detectron2.utils.registry import Registry
+
+PROPOSAL_GENERATOR_REGISTRY = Registry("PROPOSAL_GENERATOR")
+PROPOSAL_GENERATOR_REGISTRY.__doc__ = """
+Registry for proposal generator, which produces object proposals from feature maps.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+The call should return a `nn.Module` object.
+"""
+
+from . import rpn, rrpn  # noqa F401 isort:skip
+
+
+def build_proposal_generator(cfg, input_shape):
+    """
+    Build a proposal generator from `cfg.MODEL.PROPOSAL_GENERATOR.NAME`.
+    The name can be "PrecomputedProposals" to use no proposal generator.
+    """
+    name = cfg.MODEL.PROPOSAL_GENERATOR.NAME
+    if name == "PrecomputedProposals":
+        return None
+
+    return PROPOSAL_GENERATOR_REGISTRY.get(name)(cfg, input_shape)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/proposal_generator/proposal_utils.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import math
+import torch
+
+from detectron2.structures import Instances
+
+
+def add_ground_truth_to_proposals(gt_boxes, proposals):
+    """
+    Call `add_ground_truth_to_proposals_single_image` for all images.
+
+    Args:
+        gt_boxes(list[Boxes]): list of N elements. Element i is a Boxes
+            representing the gound-truth for image i.
+        proposals (list[Instances]): list of N elements. Element i is a Instances
+            representing the proposals for image i.
+
+    Returns:
+        list[Instances]: list of N Instances. Each is the proposals for the image,
+            with field "proposal_boxes" and "objectness_logits".
+    """
+    assert gt_boxes is not None
+
+    assert len(proposals) == len(gt_boxes)
+    if len(proposals) == 0:
+        return proposals
+
+    return [
+        add_ground_truth_to_proposals_single_image(gt_boxes_i, proposals_i)
+        for gt_boxes_i, proposals_i in zip(gt_boxes, proposals)
+    ]
+
+
+def add_ground_truth_to_proposals_single_image(gt_boxes, proposals):
+    """
+    Augment `proposals` with ground-truth boxes from `gt_boxes`.
+
+    Args:
+        Same as `add_ground_truth_to_proposals`, but with gt_boxes and proposals
+        per image.
+
+    Returns:
+        Same as `add_ground_truth_to_proposals`, but for only one image.
+    """
+    device = proposals.objectness_logits.device
+    # Concatenating gt_boxes with proposals requires them to have the same fields
+    # Assign all ground-truth boxes an objectness logit corresponding to P(object) \approx 1.
+    gt_logit_value = math.log((1.0 - 1e-10) / (1 - (1.0 - 1e-10)))
+
+    gt_logits = gt_logit_value * torch.ones(len(gt_boxes), device=device)
+    gt_proposal = Instances(proposals.image_size)
+
+    gt_proposal.proposal_boxes = gt_boxes
+    gt_proposal.objectness_logits = gt_logits
+    new_proposals = Instances.cat([proposals, gt_proposal])
+
+    return new_proposals

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/proposal_generator/rpn.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from typing import Dict, List
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from detectron2.config import configurable
+from detectron2.layers import ShapeSpec
+from detectron2.structures import Boxes, Instances, pairwise_iou
+from detectron2.utils.memory import retry_if_cuda_oom
+from detectron2.utils.registry import Registry
+
+from ..anchor_generator import build_anchor_generator
+from ..box_regression import Box2BoxTransform
+from ..matcher import Matcher
+from ..sampling import subsample_labels
+from .build import PROPOSAL_GENERATOR_REGISTRY
+from .rpn_outputs import RPNOutputs, find_top_rpn_proposals
+
+RPN_HEAD_REGISTRY = Registry("RPN_HEAD")
+RPN_HEAD_REGISTRY.__doc__ = """
+Registry for RPN heads, which take feature maps and perform
+objectness classification and bounding box regression for anchors.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+The call should return a `nn.Module` object.
+"""
+
+
+def build_rpn_head(cfg, input_shape):
+    """
+    Build an RPN head defined by `cfg.MODEL.RPN.HEAD_NAME`.
+    """
+    name = cfg.MODEL.RPN.HEAD_NAME
+    return RPN_HEAD_REGISTRY.get(name)(cfg, input_shape)
+
+
+@RPN_HEAD_REGISTRY.register()
+class StandardRPNHead(nn.Module):
+    """
+    Standard RPN classification and regression heads described in :paper:`Faster R-CNN`.
+    Uses a 3x3 conv to produce a shared hidden state from which one 1x1 conv predicts
+    objectness logits for each anchor and a second 1x1 conv predicts bounding-box deltas
+    specifying how to deform each anchor into an object proposal.
+    """
+
+    @configurable
+    def __init__(self, *, in_channels: int, num_anchors: int, box_dim: int = 4):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            in_channels (int): number of input feature channels. When using multiple
+                input features, they must have the same number of channels.
+            num_anchors (int): number of anchors to predict for *each spatial position*
+                on the feature map. The total number of anchors for each
+                feature map will be `num_anchors * H * W`.
+            box_dim (int): dimension of a box, which is also the number of box regression
+                predictions to make for each anchor. An axis aligned box has
+                box_dim=4, while a rotated box has box_dim=5.
+        """
+        super().__init__()
+        # 3x3 conv for the hidden representation
+        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+        # 1x1 conv for predicting objectness logits
+        self.objectness_logits = nn.Conv2d(in_channels, num_anchors, kernel_size=1, stride=1)
+        # 1x1 conv for predicting box2box transform deltas
+        self.anchor_deltas = nn.Conv2d(in_channels, num_anchors * box_dim, kernel_size=1, stride=1)
+
+        for l in [self.conv, self.objectness_logits, self.anchor_deltas]:
+            nn.init.normal_(l.weight, std=0.01)
+            nn.init.constant_(l.bias, 0)
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        # Standard RPN is shared across levels:
+        in_channels = [s.channels for s in input_shape]
+        assert len(set(in_channels)) == 1, "Each level must have the same channel!"
+        in_channels = in_channels[0]
+
+        # RPNHead should take the same input as anchor generator
+        # NOTE: it assumes that creating an anchor generator does not have unwanted side effect.
+        anchor_generator = build_anchor_generator(cfg, input_shape)
+        num_anchors = anchor_generator.num_anchors
+        box_dim = anchor_generator.box_dim
+        assert (
+            len(set(num_anchors)) == 1
+        ), "Each level must have the same number of anchors per spatial position"
+        return {"in_channels": in_channels, "num_anchors": num_anchors[0], "box_dim": box_dim}
+
+    def forward(self, features):
+        """
+        Args:
+            features (list[Tensor]): list of feature maps
+
+        Returns:
+            list[Tensor]: A list of L elements.
+                Element i is a tensor of shape (N, A, Hi, Wi) representing
+                the predicted objectness logits for all anchors. A is the number of cell anchors.
+            list[Tensor]: A list of L elements. Element i is a tensor of shape
+                (N, A*box_dim, Hi, Wi) representing the predicted "deltas" used to transform anchors
+                to proposals.
+        """
+        pred_objectness_logits = []
+        pred_anchor_deltas = []
+        for x in features:
+            t = F.relu(self.conv(x))
+            pred_objectness_logits.append(self.objectness_logits(t))
+            pred_anchor_deltas.append(self.anchor_deltas(t))
+        return pred_objectness_logits, pred_anchor_deltas
+
+
+@PROPOSAL_GENERATOR_REGISTRY.register()
+class RPN(nn.Module):
+    """
+    Region Proposal Network, introduced by :paper:`Faster R-CNN`.
+    """
+
+    def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):
+        super().__init__()
+
+        # fmt: off
+        self.min_box_side_len     = cfg.MODEL.PROPOSAL_GENERATOR.MIN_SIZE
+        self.in_features          = cfg.MODEL.RPN.IN_FEATURES
+        self.nms_thresh           = cfg.MODEL.RPN.NMS_THRESH
+        self.batch_size_per_image = cfg.MODEL.RPN.BATCH_SIZE_PER_IMAGE
+        self.positive_fraction    = cfg.MODEL.RPN.POSITIVE_FRACTION
+        self.smooth_l1_beta       = cfg.MODEL.RPN.SMOOTH_L1_BETA
+        self.loss_weight          = cfg.MODEL.RPN.LOSS_WEIGHT
+        # fmt: on
+
+        # Map from self.training state to train/test settings
+        self.pre_nms_topk = {
+            True: cfg.MODEL.RPN.PRE_NMS_TOPK_TRAIN,
+            False: cfg.MODEL.RPN.PRE_NMS_TOPK_TEST,
+        }
+        self.post_nms_topk = {
+            True: cfg.MODEL.RPN.POST_NMS_TOPK_TRAIN,
+            False: cfg.MODEL.RPN.POST_NMS_TOPK_TEST,
+        }
+        self.boundary_threshold = cfg.MODEL.RPN.BOUNDARY_THRESH
+
+        self.anchor_generator = build_anchor_generator(
+            cfg, [input_shape[f] for f in self.in_features]
+        )
+        self.box2box_transform = Box2BoxTransform(weights=cfg.MODEL.RPN.BBOX_REG_WEIGHTS)
+        self.anchor_matcher = Matcher(
+            cfg.MODEL.RPN.IOU_THRESHOLDS, cfg.MODEL.RPN.IOU_LABELS, allow_low_quality_matches=True
+        )
+        self.rpn_head = build_rpn_head(cfg, [input_shape[f] for f in self.in_features])
+
+    def _subsample_labels(self, label):
+        """
+        Randomly sample a subset of positive and negative examples, and overwrite
+        the label vector to the ignore value (-1) for all elements that are not
+        included in the sample.
+
+        Args:
+            labels (Tensor): a vector of -1, 0, 1. Will be modified in-place and returned.
+        """
+        pos_idx, neg_idx = subsample_labels(
+            label, self.batch_size_per_image, self.positive_fraction, 0
+        )
+        # Fill with the ignore label (-1), then set positive and negative labels
+        label.fill_(-1)
+        label.scatter_(0, pos_idx, 1)
+        label.scatter_(0, neg_idx, 0)
+        return label
+
+    @torch.no_grad()
+    def label_and_sample_anchors(self, anchors: List[Boxes], gt_instances: List[Instances]):
+        """
+        Args:
+            anchors (list[Boxes]): anchors for each feature map.
+            gt_instances: the ground-truth instances for each image.
+
+        Returns:
+            list[Tensor]:
+                List of #demo tensors. i-th element is a vector of labels whose length is
+                the total number of anchors across feature maps. Label values are in {-1, 0, 1},
+                with meanings: -1 = ignore; 0 = negative class; 1 = positive class.
+            list[Tensor]:
+                i-th element is a Nx4 tensor, where N is the total number of anchors across
+                feature maps.  The values are the matched gt boxes for each anchor.
+                Values are undefined for those anchors not labeled as 1.
+        """
+        anchors = Boxes.cat(anchors)
+
+        gt_boxes = [x.gt_boxes for x in gt_instances]
+        image_sizes = [x.image_size for x in gt_instances]
+        del gt_instances
+
+        gt_labels = []
+        matched_gt_boxes = []
+        for image_size_i, gt_boxes_i in zip(image_sizes, gt_boxes):
+            """
+            image_size_i: (h, w) for the i-th image
+            gt_boxes_i: ground-truth boxes for i-th image
+            """
+
+            match_quality_matrix = retry_if_cuda_oom(pairwise_iou)(gt_boxes_i, anchors)
+            matched_idxs, gt_labels_i = retry_if_cuda_oom(self.anchor_matcher)(match_quality_matrix)
+            # Matching is memory-expensive and may result in CPU tensors. But the result is small
+            gt_labels_i = gt_labels_i.to(device=gt_boxes_i.device)
+            del match_quality_matrix
+
+            if self.boundary_threshold >= 0:
+                # Discard anchors that go out of the boundaries of the image
+                # NOTE: This is legacy functionality that is turned off by default in Detectron2
+                anchors_inside_image = anchors.inside_box(image_size_i, self.boundary_threshold)
+                gt_labels_i[~anchors_inside_image] = -1
+
+            # A vector of labels (-1, 0, 1) for each anchor
+            gt_labels_i = self._subsample_labels(gt_labels_i)
+
+            if len(gt_boxes_i) == 0:
+                # These values won't be used anyway since the anchor is labeled as background
+                matched_gt_boxes_i = torch.zeros_like(anchors.tensor)
+            else:
+                # TODO wasted indexing computation for ignored boxes
+                matched_gt_boxes_i = gt_boxes_i[matched_idxs].tensor
+
+            gt_labels.append(gt_labels_i)  # N,AHW
+            matched_gt_boxes.append(matched_gt_boxes_i)
+        return gt_labels, matched_gt_boxes
+
+    def forward(self, images, features, gt_instances=None):
+        """
+        Args:
+            images (ImageList): input images of length `N`
+            features (dict[str: Tensor]): input data as a mapping from feature
+                map name to tensor. Axis 0 represents the number of images `N` in
+                the input data; axes 1-3 are channels, height, and width, which may
+                vary between feature maps (e.g., if a feature pyramid is used).
+            gt_instances (list[Instances], optional): a length `N` list of `Instances`s.
+                Each `Instances` stores ground-truth instances for the corresponding image.
+
+        Returns:
+            proposals: list[Instances]: contains fields "proposal_boxes", "objectness_logits"
+            loss: dict[Tensor] or None
+        """
+        features = [features[f] for f in self.in_features]
+        pred_objectness_logits, pred_anchor_deltas = self.rpn_head(features)
+        anchors = self.anchor_generator(features)
+
+        if self.training:
+            gt_labels, gt_boxes = self.label_and_sample_anchors(anchors, gt_instances)
+        else:
+            gt_labels, gt_boxes = None, None
+
+        outputs = RPNOutputs(
+            self.box2box_transform,
+            self.batch_size_per_image,
+            images,
+            pred_objectness_logits,
+            pred_anchor_deltas,
+            anchors,
+            gt_labels,
+            gt_boxes,
+            self.smooth_l1_beta,
+        )
+
+        if self.training:
+            losses = {k: v * self.loss_weight for k, v in outputs.losses().items()}
+        else:
+            losses = {}
+
+        with torch.no_grad():
+            # Find the top proposals by applying NMS and removing boxes that
+            # are too small. The proposals are treated as fixed for approximate
+            # joint training with roi heads. This approach ignores the derivative
+            # w.r.t. the proposal boxes’ coordinates that are also network
+            # responses, so is approximate.
+            proposals = find_top_rpn_proposals(
+                outputs.predict_proposals(),
+                outputs.predict_objectness_logits(),
+                images,
+                self.nms_thresh,
+                self.pre_nms_topk[self.training],
+                self.post_nms_topk[self.training],
+                self.min_box_side_len,
+                self.training,
+            )
+
+        return proposals, losses

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/proposal_generator/rpn_outputs.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import itertools
+import logging
+import torch
+import torch.nn.functional as F
+from fvcore.nn import smooth_l1_loss
+
+from detectron2.layers import batched_nms, cat
+from detectron2.structures import Boxes, Instances
+from detectron2.utils.events import get_event_storage
+
+logger = logging.getLogger(__name__)
+
+# TODO: comments for future refactoring of this module
+#
+# From @rbg:
+# This code involves a significant amount of tensor reshaping and permuting. Look for
+# ways to simplify this.
+
+"""
+Shape shorthand in this module:
+
+    N: number of images in the minibatch
+    L: number of feature maps per image on which RPN is run
+    A: number of cell anchors (must be the same for all feature maps)
+    Hi, Wi: height and width of the i-th feature map
+    4: size of the box parameterization
+
+Naming convention:
+
+    objectness: refers to the binary classification of an anchor as object vs. not
+    object.
+
+    deltas: refers to the 4-d (dx, dy, dw, dh) deltas that parameterize the box2box
+    transform (see :class:`box_regression.Box2BoxTransform`).
+
+    pred_objectness_logits: predicted objectness scores in [-inf, +inf]; use
+        sigmoid(pred_objectness_logits) to estimate P(object).
+
+    gt_labels: ground-truth binary classification labels for objectness
+
+    pred_anchor_deltas: predicted box2box transform deltas
+
+    gt_anchor_deltas: ground-truth box2box transform deltas
+"""
+
+
+def find_top_rpn_proposals(
+    proposals,
+    pred_objectness_logits,
+    images,
+    nms_thresh,
+    pre_nms_topk,
+    post_nms_topk,
+    min_box_side_len,
+    training,
+):
+    """
+    For each feature map, select the `pre_nms_topk` highest scoring proposals,
+    apply NMS, clip proposals, and remove small boxes. Return the `post_nms_topk`
+    highest scoring proposals among all the feature maps if `training` is True,
+    otherwise, returns the highest `post_nms_topk` scoring proposals for each
+    feature map.
+
+    Args:
+        proposals (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A, 4).
+            All proposal predictions on the feature maps.
+        pred_objectness_logits (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A).
+        images (ImageList): Input images as an :class:`ImageList`.
+        nms_thresh (float): IoU threshold to use for NMS
+        pre_nms_topk (int): number of top k scoring proposals to keep before applying NMS.
+            When RPN is run on multiple feature maps (as in FPN) this number is per
+            feature map.
+        post_nms_topk (int): number of top k scoring proposals to keep after applying NMS.
+            When RPN is run on multiple feature maps (as in FPN) this number is total,
+            over all feature maps.
+        min_box_side_len (float): minimum proposal box side length in pixels (absolute units
+            wrt input images).
+        training (bool): True if proposals are to be used in training, otherwise False.
+            This arg exists only to support a legacy bug; look for the "NB: Legacy bug ..."
+            comment.
+
+    Returns:
+        proposals (list[Instances]): list of N Instances. The i-th Instances
+            stores post_nms_topk object proposals for image i, sorted by their
+            objectness score in descending order.
+    """
+    image_sizes = images.image_sizes  # in (h, w) order
+    num_images = len(image_sizes)
+    device = proposals[0].device
+
+    # 1. Select top-k anchor for every level and every image
+    topk_scores = []  # #lvl Tensor, each of shape N x topk
+    topk_proposals = []
+    level_ids = []  # #lvl Tensor, each of shape (topk,)
+    batch_idx = torch.arange(num_images, device=device)
+    for level_id, proposals_i, logits_i in zip(
+        itertools.count(), proposals, pred_objectness_logits
+    ):
+        Hi_Wi_A = logits_i.shape[1]
+        num_proposals_i = min(pre_nms_topk, Hi_Wi_A)
+
+        # sort is faster than topk (https://github.com/pytorch/pytorch/issues/22812)
+        # topk_scores_i, topk_idx = logits_i.topk(num_proposals_i, dim=1)
+        logits_i, idx = logits_i.sort(descending=True, dim=1)
+        topk_scores_i = logits_i[batch_idx, :num_proposals_i]
+        topk_idx = idx[batch_idx, :num_proposals_i]
+
+        # each is N x topk
+        topk_proposals_i = proposals_i[batch_idx[:, None], topk_idx]  # N x topk x 4
+
+        topk_proposals.append(topk_proposals_i)
+        topk_scores.append(topk_scores_i)
+        level_ids.append(torch.full((num_proposals_i,), level_id, dtype=torch.int64, device=device))
+
+    # 2. Concat all levels together
+    topk_scores = cat(topk_scores, dim=1)
+    topk_proposals = cat(topk_proposals, dim=1)
+    level_ids = cat(level_ids, dim=0)
+
+    # 3. For each image, run a per-level NMS, and choose topk results.
+    results = []
+    for n, image_size in enumerate(image_sizes):
+        boxes = Boxes(topk_proposals[n])
+        scores_per_img = topk_scores[n]
+        lvl = level_ids
+
+        valid_mask = torch.isfinite(boxes.tensor).all(dim=1) & torch.isfinite(scores_per_img)
+        if not valid_mask.all():
+            if training:
+                raise FloatingPointError(
+                    "Predicted boxes or scores contain Inf/NaN. Training has diverged."
+                )
+            boxes = boxes[valid_mask]
+            scores_per_img = scores_per_img[valid_mask]
+            lvl = lvl[valid_mask]
+        boxes.clip(image_size)
+
+        # filter empty boxes
+        keep = boxes.nonempty(threshold=min_box_side_len)
+        if keep.sum().item() != len(boxes):
+            boxes, scores_per_img, lvl = boxes[keep], scores_per_img[keep], lvl[keep]
+
+        keep = batched_nms(boxes.tensor, scores_per_img, lvl, nms_thresh)
+        # In Detectron1, there was different behavior during training vs. testing.
+        # (https://github.com/facebookresearch/Detectron/issues/459)
+        # During training, topk is over the proposals from *all* images in the training batch.
+        # During testing, it is over the proposals for each image separately.
+        # As a result, the training behavior becomes batch-dependent,
+        # and the configuration "POST_NMS_TOPK_TRAIN" end up relying on the batch size.
+        # This bug is addressed in Detectron2 to make the behavior independent of batch size.
+        keep = keep[:post_nms_topk]  # keep is already sorted
+
+        res = Instances(image_size)
+        res.proposal_boxes = boxes[keep]
+        res.objectness_logits = scores_per_img[keep]
+        results.append(res)
+    return results
+
+
+def rpn_losses(
+    gt_labels, gt_anchor_deltas, pred_objectness_logits, pred_anchor_deltas, smooth_l1_beta
+):
+    """
+    Args:
+        gt_labels (Tensor): shape (N,), each element in {-1, 0, 1} representing
+            ground-truth objectness labels with: -1 = ignore; 0 = not object; 1 = object.
+        gt_anchor_deltas (Tensor): shape (N, box_dim), row i represents ground-truth
+            box2box transform targets (dx, dy, dw, dh) or (dx, dy, dw, dh, da) that map anchor i to
+            its matched ground-truth box.
+        pred_objectness_logits (Tensor): shape (N,), each element is a predicted objectness
+            logit.
+        pred_anchor_deltas (Tensor): shape (N, box_dim), each row is a predicted box2box
+            transform (dx, dy, dw, dh) or (dx, dy, dw, dh, da)
+        smooth_l1_beta (float): The transition point between L1 and L2 loss in
+            the smooth L1 loss function. When set to 0, the loss becomes L1. When
+            set to +inf, the loss becomes constant 0.
+
+    Returns:
+        objectness_loss, localization_loss, both unnormalized (summed over samples).
+    """
+    pos_masks = gt_labels == 1
+    localization_loss = smooth_l1_loss(
+        pred_anchor_deltas[pos_masks], gt_anchor_deltas[pos_masks], smooth_l1_beta, reduction="sum"
+    )
+
+    valid_masks = gt_labels >= 0
+    objectness_loss = F.binary_cross_entropy_with_logits(
+        pred_objectness_logits[valid_masks],
+        gt_labels[valid_masks].to(torch.float32),
+        reduction="sum",
+    )
+    return objectness_loss, localization_loss
+
+
+class RPNOutputs(object):
+    def __init__(
+        self,
+        box2box_transform,
+        batch_size_per_image,
+        images,
+        pred_objectness_logits,
+        pred_anchor_deltas,
+        anchors,
+        gt_labels=None,
+        gt_boxes=None,
+        smooth_l1_beta=0.0,
+    ):
+        """
+        Args:
+            box2box_transform (Box2BoxTransform): :class:`Box2BoxTransform` instance for
+                anchor-proposal transformations.
+            images (ImageList): :class:`ImageList` instance representing N input images
+            batch_size_per_image (int): number of proposals to sample when training
+            pred_objectness_logits (list[Tensor]): A list of L elements.
+                Element i is a tensor of shape (N, A, Hi, Wi) representing
+                the predicted objectness logits for anchors.
+            pred_anchor_deltas (list[Tensor]): A list of L elements. Element i is a tensor of shape
+                (N, A*4 or 5, Hi, Wi) representing the predicted "deltas" used to transform anchors
+                to proposals.
+            anchors (list[Boxes or RotatedBoxes]): A list of Boxes/RotatedBoxes storing the all
+                the anchors for each feature map. See :meth:`AnchorGenerator.forward`.
+            gt_labels (list[Tensor]): Available on in training.
+                See :meth:`RPN.label_and_sample_anchors`.
+            gt_boxes (list[Boxes or RotatedBoxes]): Available on in training.
+                See :meth:`RPN.label_and_sample_anchors`.
+            smooth_l1_beta (float): The transition point between L1 and L2 loss in
+                the smooth L1 loss function. When set to 0, the loss becomes L1. When
+                set to +inf, the loss becomes constant 0.
+        """
+        self.box2box_transform = box2box_transform
+        self.batch_size_per_image = batch_size_per_image
+
+        B = anchors[0].tensor.size(1)  # box dimension (4 or 5)
+        self.pred_objectness_logits = [
+            # Reshape: (N, A, Hi, Wi) -> (N, Hi, Wi, A) -> (N, Hi*Wi*A)
+            score.permute(0, 2, 3, 1).flatten(1)
+            for score in pred_objectness_logits
+        ]
+
+        self.pred_anchor_deltas = [
+            # Reshape: (N, A*B, Hi, Wi) -> (N, A, B, Hi, Wi) -> (N, Hi, Wi, A, B)
+            #          -> (N, Hi*Wi*A, B)
+            x.view(x.shape[0], -1, B, x.shape[-2], x.shape[-1])
+            .permute(0, 3, 4, 1, 2)
+            .flatten(1, -2)
+            for x in pred_anchor_deltas
+        ]
+
+        self.anchors = anchors
+
+        self.gt_boxes = gt_boxes
+        self.gt_labels = gt_labels
+
+        self.num_images = len(images)
+        self.smooth_l1_beta = smooth_l1_beta
+
+    def losses(self):
+        """
+        Return the losses from a set of RPN predictions and their associated ground-truth.
+
+        Returns:
+            dict[loss name -> loss value]: A dict mapping from loss name to loss value.
+                Loss names are: `loss_rpn_cls` for objectness classification and
+                `loss_rpn_loc` for proposal localization.
+        """
+        gt_labels = torch.stack(self.gt_labels)
+        anchors = self.anchors[0].cat(self.anchors).tensor  # Ax(4 or 5)
+        gt_anchor_deltas = [self.box2box_transform.get_deltas(anchors, k) for k in self.gt_boxes]
+        gt_anchor_deltas = torch.stack(gt_anchor_deltas)
+
+        # Log the number of positive/negative anchors per-image that's used in training
+        num_pos_anchors = (gt_labels == 1).sum().item()
+        num_neg_anchors = (gt_labels == 0).sum().item()
+        storage = get_event_storage()
+        storage.put_scalar("rpn/num_pos_anchors", num_pos_anchors / self.num_images)
+        storage.put_scalar("rpn/num_neg_anchors", num_neg_anchors / self.num_images)
+
+        objectness_loss, localization_loss = rpn_losses(
+            gt_labels,
+            gt_anchor_deltas,
+            # concat on the Hi*Wi*A dimension
+            cat(self.pred_objectness_logits, dim=1),
+            cat(self.pred_anchor_deltas, dim=1),
+            self.smooth_l1_beta,
+        )
+        normalizer = self.batch_size_per_image * self.num_images
+        return {
+            "loss_rpn_cls": objectness_loss / normalizer,
+            "loss_rpn_loc": localization_loss / normalizer,
+        }
+
+    def predict_proposals(self):
+        """
+        Transform anchors into proposals by applying the predicted anchor deltas.
+
+        Returns:
+            proposals (list[Tensor]): A list of L tensors. Tensor i has shape
+                (N, Hi*Wi*A, B), where B is box dimension (4 or 5).
+        """
+        proposals = []
+        # For each feature map
+        for anchors_i, pred_anchor_deltas_i in zip(self.anchors, self.pred_anchor_deltas):
+            B = anchors_i.tensor.size(1)
+            N = self.num_images
+            pred_anchor_deltas_i = pred_anchor_deltas_i.reshape(-1, B)
+            # Expand anchors to shape (N*Hi*Wi*A, B)
+            anchors_i = anchors_i.tensor.unsqueeze(0).expand(N, -1, -1).reshape(-1, B)
+            proposals_i = self.box2box_transform.apply_deltas(pred_anchor_deltas_i, anchors_i)
+            # Append feature map proposals with shape (N, Hi*Wi*A, B)
+            proposals.append(proposals_i.view(N, -1, B))
+        return proposals
+
+    def predict_objectness_logits(self):
+        """
+        Return objectness logits in the same format as the proposals returned by
+        :meth:`predict_proposals`.
+
+        Returns:
+            pred_objectness_logits (list[Tensor]): A list of L tensors. Tensor i has shape
+                (N, Hi*Wi*A).
+        """
+        return self.pred_objectness_logits

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/proposal_generator/rrpn.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import itertools
+import logging
+from typing import Dict, List
+import torch
+
+from detectron2.layers import ShapeSpec, batched_nms_rotated, cat
+from detectron2.structures import Instances, RotatedBoxes, pairwise_iou_rotated
+from detectron2.utils.memory import retry_if_cuda_oom
+
+from ..box_regression import Box2BoxTransformRotated
+from .build import PROPOSAL_GENERATOR_REGISTRY
+from .rpn import RPN
+from .rpn_outputs import RPNOutputs
+
+logger = logging.getLogger(__name__)
+
+
+def find_top_rrpn_proposals(
+    proposals,
+    pred_objectness_logits,
+    images,
+    nms_thresh,
+    pre_nms_topk,
+    post_nms_topk,
+    min_box_side_len,
+    training,
+):
+    """
+    For each feature map, select the `pre_nms_topk` highest scoring proposals,
+    apply NMS, clip proposals, and remove small boxes. Return the `post_nms_topk`
+    highest scoring proposals among all the feature maps if `training` is True,
+    otherwise, returns the highest `post_nms_topk` scoring proposals for each
+    feature map.
+
+    Args:
+        proposals (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A, 5).
+            All proposal predictions on the feature maps.
+        pred_objectness_logits (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A).
+        images (ImageList): Input images as an :class:`ImageList`.
+        nms_thresh (float): IoU threshold to use for NMS
+        pre_nms_topk (int): number of top k scoring proposals to keep before applying NMS.
+            When RRPN is run on multiple feature maps (as in FPN) this number is per
+            feature map.
+        post_nms_topk (int): number of top k scoring proposals to keep after applying NMS.
+            When RRPN is run on multiple feature maps (as in FPN) this number is total,
+            over all feature maps.
+        min_box_side_len (float): minimum proposal box side length in pixels (absolute units
+            wrt input images).
+        training (bool): True if proposals are to be used in training, otherwise False.
+            This arg exists only to support a legacy bug; look for the "NB: Legacy bug ..."
+            comment.
+
+    Returns:
+        proposals (list[Instances]): list of N Instances. The i-th Instances
+            stores post_nms_topk object proposals for image i.
+    """
+    image_sizes = images.image_sizes  # in (h, w) order
+    num_images = len(image_sizes)
+    device = proposals[0].device
+
+    # 1. Select top-k anchor for every level and every image
+    topk_scores = []  # #lvl Tensor, each of shape N x topk
+    topk_proposals = []
+    level_ids = []  # #lvl Tensor, each of shape (topk,)
+    batch_idx = torch.arange(num_images, device=device)
+    for level_id, proposals_i, logits_i in zip(
+        itertools.count(), proposals, pred_objectness_logits
+    ):
+        Hi_Wi_A = logits_i.shape[1]
+        num_proposals_i = min(pre_nms_topk, Hi_Wi_A)
+
+        # sort is faster than topk (https://github.com/pytorch/pytorch/issues/22812)
+        # topk_scores_i, topk_idx = logits_i.topk(num_proposals_i, dim=1)
+        logits_i, idx = logits_i.sort(descending=True, dim=1)
+        topk_scores_i = logits_i[batch_idx, :num_proposals_i]
+        topk_idx = idx[batch_idx, :num_proposals_i]
+
+        # each is N x topk
+        topk_proposals_i = proposals_i[batch_idx[:, None], topk_idx]  # N x topk x 5
+
+        topk_proposals.append(topk_proposals_i)
+        topk_scores.append(topk_scores_i)
+        level_ids.append(torch.full((num_proposals_i,), level_id, dtype=torch.int64, device=device))
+
+    # 2. Concat all levels together
+    topk_scores = cat(topk_scores, dim=1)
+    topk_proposals = cat(topk_proposals, dim=1)
+    level_ids = cat(level_ids, dim=0)
+
+    # 3. For each image, run a per-level NMS, and choose topk results.
+    results = []
+    for n, image_size in enumerate(image_sizes):
+        boxes = RotatedBoxes(topk_proposals[n])
+        scores_per_img = topk_scores[n]
+        valid_mask = torch.isfinite(boxes.tensor).all(dim=1) & torch.isfinite(scores_per_img)
+        if not valid_mask.all():
+            boxes = boxes[valid_mask]
+            scores_per_img = scores_per_img[valid_mask]
+        boxes.clip(image_size)
+
+        # filter empty boxes
+        keep = boxes.nonempty(threshold=min_box_side_len)
+        lvl = level_ids
+        if keep.sum().item() != len(boxes):
+            boxes, scores_per_img, lvl = (boxes[keep], scores_per_img[keep], level_ids[keep])
+
+        keep = batched_nms_rotated(boxes.tensor, scores_per_img, lvl, nms_thresh)
+        # In Detectron1, there was different behavior during training vs. testing.
+        # (https://github.com/facebookresearch/Detectron/issues/459)
+        # During training, topk is over the proposals from *all* images in the training batch.
+        # During testing, it is over the proposals for each image separately.
+        # As a result, the training behavior becomes batch-dependent,
+        # and the configuration "POST_NMS_TOPK_TRAIN" end up relying on the batch size.
+        # This bug is addressed in Detectron2 to make the behavior independent of batch size.
+        keep = keep[:post_nms_topk]
+
+        res = Instances(image_size)
+        res.proposal_boxes = boxes[keep]
+        res.objectness_logits = scores_per_img[keep]
+        results.append(res)
+    return results
+
+
+@PROPOSAL_GENERATOR_REGISTRY.register()
+class RRPN(RPN):
+    """
+    Rotated Region Proposal Network described in :paper:`RRPN`.
+    """
+
+    def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):
+        super().__init__(cfg, input_shape)
+        self.box2box_transform = Box2BoxTransformRotated(weights=cfg.MODEL.RPN.BBOX_REG_WEIGHTS)
+        if self.boundary_threshold >= 0:
+            raise NotImplementedError(
+                "boundary_threshold is a legacy option not implemented for RRPN."
+            )
+
+    @torch.no_grad()
+    def label_and_sample_anchors(self, anchors: List[RotatedBoxes], gt_instances: List[Instances]):
+        """
+        Args:
+            anchors (list[RotatedBoxes]): anchors for each feature map.
+            gt_instances: the ground-truth instances for each image.
+
+        Returns:
+            list[Tensor]:
+                List of #demo tensors. i-th element is a vector of labels whose length is
+                the total number of anchors across feature maps. Label values are in {-1, 0, 1},
+                with meanings: -1 = ignore; 0 = negative class; 1 = positive class.
+            list[Tensor]:
+                i-th element is a Nx5 tensor, where N is the total number of anchors across
+                feature maps.  The values are the matched gt boxes for each anchor.
+                Values are undefined for those anchors not labeled as 1.
+        """
+        anchors = RotatedBoxes.cat(anchors)
+
+        gt_boxes = [x.gt_boxes for x in gt_instances]
+        del gt_instances
+
+        gt_labels = []
+        matched_gt_boxes = []
+        for gt_boxes_i in gt_boxes:
+            """
+            gt_boxes_i: ground-truth boxes for i-th image
+            """
+            match_quality_matrix = retry_if_cuda_oom(pairwise_iou_rotated)(gt_boxes_i, anchors)
+            matched_idxs, gt_labels_i = retry_if_cuda_oom(self.anchor_matcher)(match_quality_matrix)
+            # Matching is memory-expensive and may result in CPU tensors. But the result is small
+            gt_labels_i = gt_labels_i.to(device=gt_boxes_i.device)
+
+            # A vector of labels (-1, 0, 1) for each anchor
+            gt_labels_i = self._subsample_labels(gt_labels_i)
+
+            if len(gt_boxes_i) == 0:
+                # These values won't be used anyway since the anchor is labeled as background
+                matched_gt_boxes_i = torch.zeros_like(anchors.tensor)
+            else:
+                # TODO wasted indexing computation for ignored boxes
+                matched_gt_boxes_i = gt_boxes_i[matched_idxs].tensor
+
+            gt_labels.append(gt_labels_i)  # N,AHW
+            matched_gt_boxes.append(matched_gt_boxes_i)
+        return gt_labels, matched_gt_boxes
+
+    def forward(self, images, features, gt_instances=None):
+        # same signature as RPN.forward
+        features = [features[f] for f in self.in_features]
+        pred_objectness_logits, pred_anchor_deltas = self.rpn_head(features)
+        anchors = self.anchor_generator(features)
+
+        if self.training:
+            gt_labels, gt_boxes = self.label_and_sample_anchors(anchors, gt_instances)
+        else:
+            gt_labels, gt_boxes = None, None
+
+        outputs = RPNOutputs(
+            self.box2box_transform,
+            self.batch_size_per_image,
+            images,
+            pred_objectness_logits,
+            pred_anchor_deltas,
+            anchors,
+            gt_labels,
+            gt_boxes,
+            self.smooth_l1_beta,
+        )
+
+        if self.training:
+            losses = {k: v * self.loss_weight for k, v in outputs.losses().items()}
+        else:
+            losses = {}
+
+        with torch.no_grad():
+            # Find the top proposals by applying NMS and removing boxes that
+            # are too small. The proposals are treated as fixed for approximate
+            # joint training with roi heads. This approach ignores the derivative
+            # w.r.t. the proposal boxes’ coordinates that are also network
+            # responses, so is approximate.
+
+            # Note: this line is the only difference v.s. RPN.forward
+            proposals = find_top_rrpn_proposals(
+                outputs.predict_proposals(),
+                outputs.predict_objectness_logits(),
+                images,
+                self.nms_thresh,
+                self.pre_nms_topk[self.training],
+                self.post_nms_topk[self.training],
+                self.min_box_side_len,
+                self.training,
+            )
+
+        return proposals, losses

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/__init__.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .box_head import ROI_BOX_HEAD_REGISTRY, build_box_head
+from .keypoint_head import ROI_KEYPOINT_HEAD_REGISTRY, build_keypoint_head, BaseKeypointRCNNHead
+from .mask_head import ROI_MASK_HEAD_REGISTRY, build_mask_head, BaseMaskRCNNHead
+from .roi_heads import (
+    ROI_HEADS_REGISTRY,
+    ROIHeads,
+    Res5ROIHeads,
+    StandardROIHeads,
+    build_roi_heads,
+    select_foreground_proposals,
+)
+from .rotated_fast_rcnn import RROIHeads
+from .fast_rcnn import FastRCNNOutputLayers
+
+from . import cascade_rcnn  # isort:skip

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/box_head.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import numpy as np
+from typing import List
+import fvcore.nn.weight_init as weight_init
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from detectron2.config import configurable
+from detectron2.layers import Conv2d, Linear, ShapeSpec, get_norm
+from detectron2.utils.registry import Registry
+
+ROI_BOX_HEAD_REGISTRY = Registry("ROI_BOX_HEAD")
+ROI_BOX_HEAD_REGISTRY.__doc__ = """
+Registry for box heads, which make box predictions from per-region features.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+"""
+
+
+@ROI_BOX_HEAD_REGISTRY.register()
+class FastRCNNConvFCHead(nn.Module):
+    """
+    A head with several 3x3 conv layers (each followed by norm & relu) and then
+    several fc layers (each followed by relu).
+    """
+
+    @configurable
+    def __init__(
+        self, input_shape: ShapeSpec, *, conv_dims: List[int], fc_dims: List[int], conv_norm=""
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            input_shape (ShapeSpec): shape of the input feature.
+            conv_dims (list[int]): the output dimensions of the conv layers
+            fc_dims (list[int]): the output dimensions of the fc layers
+            conv_norm (str or callable): normalization for the conv layers.
+                See :func:`detectron2.layers.get_norm` for supported types.
+        """
+        super().__init__()
+        assert len(conv_dims) + len(fc_dims) > 0
+
+        self._output_size = (input_shape.channels, input_shape.height, input_shape.width)
+
+        self.conv_norm_relus = []
+        for k, conv_dim in enumerate(conv_dims):
+            conv = Conv2d(
+                self._output_size[0],
+                conv_dim,
+                kernel_size=3,
+                padding=1,
+                bias=not conv_norm,
+                norm=get_norm(conv_norm, conv_dim),
+                activation=F.relu,
+            )
+            self.add_module("conv{}".format(k + 1), conv)
+            self.conv_norm_relus.append(conv)
+            self._output_size = (conv_dim, self._output_size[1], self._output_size[2])
+
+        self.fcs = []
+        for k, fc_dim in enumerate(fc_dims):
+            fc = Linear(np.prod(self._output_size), fc_dim)
+            self.add_module("fc{}".format(k + 1), fc)
+            self.fcs.append(fc)
+            self._output_size = fc_dim
+
+        for layer in self.conv_norm_relus:
+            weight_init.c2_msra_fill(layer)
+        for layer in self.fcs:
+            weight_init.c2_xavier_fill(layer)
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        num_conv = cfg.MODEL.ROI_BOX_HEAD.NUM_CONV
+        conv_dim = cfg.MODEL.ROI_BOX_HEAD.CONV_DIM
+        num_fc = cfg.MODEL.ROI_BOX_HEAD.NUM_FC
+        fc_dim = cfg.MODEL.ROI_BOX_HEAD.FC_DIM
+        return {
+            "input_shape": input_shape,
+            "conv_dims": [conv_dim] * num_conv,
+            "fc_dims": [fc_dim] * num_fc,
+            "conv_norm": cfg.MODEL.ROI_BOX_HEAD.NORM,
+        }
+
+    def forward(self, x):
+        for layer in self.conv_norm_relus:
+            x = layer(x)
+        if len(self.fcs):
+            if x.dim() > 2:
+                x = torch.flatten(x, start_dim=1)
+            for layer in self.fcs:
+                x = F.relu(layer(x))
+        return x
+
+    @property
+    def output_shape(self):
+        """
+        Returns:
+            ShapeSpec: the output feature shape
+        """
+        o = self._output_size
+        if isinstance(o, int):
+            return ShapeSpec(channels=o)
+        else:
+            return ShapeSpec(channels=o[0], height=o[1], width=o[2])
+
+
+def build_box_head(cfg, input_shape):
+    """
+    Build a box head defined by `cfg.MODEL.ROI_BOX_HEAD.NAME`.
+    """
+    name = cfg.MODEL.ROI_BOX_HEAD.NAME
+    return ROI_BOX_HEAD_REGISTRY.get(name)(cfg, input_shape)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/cascade_rcnn.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from typing import List
+import torch
+from torch import nn
+from torch.autograd.function import Function
+
+from detectron2.config import configurable
+from detectron2.layers import ShapeSpec
+from detectron2.structures import Boxes, Instances, pairwise_iou
+from detectron2.utils.events import get_event_storage
+
+from ..box_regression import Box2BoxTransform
+from ..matcher import Matcher
+from ..poolers import ROIPooler
+from .box_head import build_box_head
+from .fast_rcnn import FastRCNNOutputLayers, fast_rcnn_inference
+from .roi_heads import ROI_HEADS_REGISTRY, StandardROIHeads
+
+
+class _ScaleGradient(Function):
+    @staticmethod
+    def forward(ctx, input, scale):
+        ctx.scale = scale
+        return input
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        return grad_output * ctx.scale, None
+
+
+@ROI_HEADS_REGISTRY.register()
+class CascadeROIHeads(StandardROIHeads):
+    """
+    Implement :paper:`Cascade R-CNN`.
+    """
+
+    @configurable
+    def __init__(
+        self,
+        *,
+        box_in_features: List[str],
+        box_pooler: ROIPooler,
+        box_heads: List[nn.Module],
+        box_predictors: List[nn.Module],
+        proposal_matchers: List[Matcher],
+        **kwargs,
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            box_pooler (ROIPooler): pooler that extracts region features from given boxes
+            box_heads (list[nn.Module]): box head for each cascade stage
+            box_predictors (list[nn.Module]): box predictor for each cascade stage
+            proposal_matchers (list[Matcher]): matcher with different IoU thresholds to
+                match boxes with ground truth for each stage. The first matcher matches
+                RPN proposals with ground truth, the other matchers use boxes predicted
+                by the previous stage as proposals and match them with ground truth.
+        """
+        assert "proposal_matcher" not in kwargs, (
+            "CascadeROIHeads takes 'proposal_matchers=' for each stage instead "
+            "of one 'proposal_matcher='."
+        )
+        # The first matcher matches RPN proposals with ground truth, done in the base class
+        kwargs["proposal_matcher"] = proposal_matchers[0]
+        num_stages = self.num_cascade_stages = len(box_heads)
+        box_heads = nn.ModuleList(box_heads)
+        box_predictors = nn.ModuleList(box_predictors)
+        assert len(box_predictors) == num_stages, f"{len(box_predictors)} != {num_stages}!"
+        assert len(proposal_matchers) == num_stages, f"{len(proposal_matchers)} != {num_stages}!"
+        super().__init__(
+            box_in_features=box_in_features,
+            box_pooler=box_pooler,
+            box_head=box_heads,
+            box_predictor=box_predictors,
+            **kwargs,
+        )
+        self.proposal_matchers = proposal_matchers
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        ret = super().from_config(cfg, input_shape)
+        ret.pop("proposal_matcher")
+        return ret
+
+    @classmethod
+    def _init_box_head(cls, cfg, input_shape):
+        # fmt: off
+        in_features              = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution        = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
+        pooler_scales            = tuple(1.0 / input_shape[k].stride for k in in_features)
+        sampling_ratio           = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
+        pooler_type              = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE
+        cascade_bbox_reg_weights = cfg.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS
+        cascade_ious             = cfg.MODEL.ROI_BOX_CASCADE_HEAD.IOUS
+        assert len(cascade_bbox_reg_weights) == len(cascade_ious)
+        assert cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG,  \
+            "CascadeROIHeads only support class-agnostic regression now!"
+        assert cascade_ious[0] == cfg.MODEL.ROI_HEADS.IOU_THRESHOLDS[0]
+        # fmt: on
+
+        in_channels = [input_shape[f].channels for f in in_features]
+        # Check all channel counts are equal
+        assert len(set(in_channels)) == 1, in_channels
+        in_channels = in_channels[0]
+
+        box_pooler = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+        pooled_shape = ShapeSpec(
+            channels=in_channels, width=pooler_resolution, height=pooler_resolution
+        )
+
+        box_heads, box_predictors, proposal_matchers = [], [], []
+        for match_iou, bbox_reg_weights in zip(cascade_ious, cascade_bbox_reg_weights):
+            box_head = build_box_head(cfg, pooled_shape)
+            box_heads.append(box_head)
+            box_predictors.append(
+                FastRCNNOutputLayers(
+                    cfg,
+                    box_head.output_shape,
+                    box2box_transform=Box2BoxTransform(weights=bbox_reg_weights),
+                )
+            )
+            proposal_matchers.append(Matcher([match_iou], [0, 1], allow_low_quality_matches=False))
+        return {
+            "box_in_features": in_features,
+            "box_pooler": box_pooler,
+            "box_heads": box_heads,
+            "box_predictors": box_predictors,
+            "proposal_matchers": proposal_matchers,
+        }
+
+    def forward(self, images, features, proposals, targets=None):
+        del images
+        if self.training:
+            proposals = self.label_and_sample_proposals(proposals, targets)
+
+        if self.training:
+            # Need targets to box head
+            losses = self._forward_box(features, proposals, targets)
+            losses.update(self._forward_mask(features, proposals))
+            losses.update(self._forward_keypoint(features, proposals))
+            return proposals, losses
+        else:
+            pred_instances = self._forward_box(features, proposals)
+            pred_instances = self.forward_with_given_boxes(features, pred_instances)
+            return pred_instances, {}
+
+    def _forward_box(self, features, proposals, targets=None):
+        """
+        Args:
+            features, targets: the same as in
+                Same as in :meth:`ROIHeads.forward`.
+            proposals (list[Instances]): the per-image object proposals with
+                their matching ground truth.
+                Each has fields "proposal_boxes", and "objectness_logits",
+                "gt_classes", "gt_boxes".
+        """
+        features = [features[f] for f in self.box_in_features]
+        head_outputs = []  # (predictor, predictions, proposals)
+        prev_pred_boxes = None
+        image_sizes = [x.image_size for x in proposals]
+        for k in range(self.num_cascade_stages):
+            if k > 0:
+                # The output boxes of the previous stage are used to create the input
+                # proposals of the next stage.
+                proposals = self._create_proposals_from_boxes(prev_pred_boxes, image_sizes)
+                if self.training:
+                    proposals = self._match_and_label_boxes(proposals, k, targets)
+            predictions = self._run_stage(features, proposals, k)
+            prev_pred_boxes = self.box_predictor[k].predict_boxes(predictions, proposals)
+            head_outputs.append((self.box_predictor[k], predictions, proposals))
+
+        if self.training:
+            losses = {}
+            storage = get_event_storage()
+            for stage, (predictor, predictions, proposals) in enumerate(head_outputs):
+                with storage.name_scope("stage{}".format(stage)):
+                    stage_losses = predictor.losses(predictions, proposals)
+                losses.update({k + "_stage{}".format(stage): v for k, v in stage_losses.items()})
+            return losses
+        else:
+            # Each is a list[Tensor] of length #image. Each tensor is Ri x (K+1)
+            scores_per_stage = [h[0].predict_probs(h[1], h[2]) for h in head_outputs]
+
+            # Average the scores across heads
+            scores = [
+                sum(list(scores_per_image)) * (1.0 / self.num_cascade_stages)
+                for scores_per_image in zip(*scores_per_stage)
+            ]
+            # Use the boxes of the last head
+            predictor, predictions, proposals = head_outputs[-1]
+            boxes = predictor.predict_boxes(predictions, proposals)
+            pred_instances, _ = fast_rcnn_inference(
+                boxes,
+                scores,
+                image_sizes,
+                predictor.test_score_thresh,
+                predictor.test_nms_thresh,
+                predictor.test_topk_per_image,
+            )
+            return pred_instances
+
+    @torch.no_grad()
+    def _match_and_label_boxes(self, proposals, stage, targets):
+        """
+        Match proposals with groundtruth using the matcher at the given stage.
+        Label the proposals as foreground or background based on the match.
+
+        Args:
+            proposals (list[Instances]): One Instances for each image, with
+                the field "proposal_boxes".
+            stage (int): the current stage
+            targets (list[Instances]): the ground truth instances
+
+        Returns:
+            list[Instances]: the same proposals, but with fields "gt_classes" and "gt_boxes"
+        """
+        num_fg_samples, num_bg_samples = [], []
+        for proposals_per_image, targets_per_image in zip(proposals, targets):
+            match_quality_matrix = pairwise_iou(
+                targets_per_image.gt_boxes, proposals_per_image.proposal_boxes
+            )
+            # proposal_labels are 0 or 1
+            matched_idxs, proposal_labels = self.proposal_matchers[stage](match_quality_matrix)
+            if len(targets_per_image) > 0:
+                gt_classes = targets_per_image.gt_classes[matched_idxs]
+                # Label unmatched proposals (0 label from matcher) as background (label=num_classes)
+                gt_classes[proposal_labels == 0] = self.num_classes
+                gt_boxes = targets_per_image.gt_boxes[matched_idxs]
+            else:
+                gt_classes = torch.zeros_like(matched_idxs) + self.num_classes
+                gt_boxes = Boxes(
+                    targets_per_image.gt_boxes.tensor.new_zeros((len(proposals_per_image), 4))
+                )
+            proposals_per_image.gt_classes = gt_classes
+            proposals_per_image.gt_boxes = gt_boxes
+
+            num_fg_samples.append((proposal_labels == 1).sum().item())
+            num_bg_samples.append(proposal_labels.numel() - num_fg_samples[-1])
+
+        # Log the number of fg/bg samples in each stage
+        storage = get_event_storage()
+        storage.put_scalar(
+            "stage{}/roi_head/num_fg_samples".format(stage),
+            sum(num_fg_samples) / len(num_fg_samples),
+        )
+        storage.put_scalar(
+            "stage{}/roi_head/num_bg_samples".format(stage),
+            sum(num_bg_samples) / len(num_bg_samples),
+        )
+        return proposals
+
+    def _run_stage(self, features, proposals, stage):
+        """
+        Args:
+            features (list[Tensor]): #lvl input features to ROIHeads
+            proposals (list[Instances]): #image Instances, with the field "proposal_boxes"
+            stage (int): the current stage
+
+        Returns:
+            Same output as `FastRCNNOutputLayers.forward()`.
+        """
+        box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals])
+        # The original implementation averages the losses among heads,
+        # but scale up the parameter gradients of the heads.
+        # This is equivalent to adding the losses among heads,
+        # but scale down the gradients on features.
+        box_features = _ScaleGradient.apply(box_features, 1.0 / self.num_cascade_stages)
+        box_features = self.box_head[stage](box_features)
+        return self.box_predictor[stage](box_features)
+
+    def _create_proposals_from_boxes(self, boxes, image_sizes):
+        """
+        Args:
+            boxes (list[Tensor]): per-image predicted boxes, each of shape Ri x 4
+            image_sizes (list[tuple]): list of image shapes in (h, w)
+
+        Returns:
+            list[Instances]: per-image proposals with the given boxes.
+        """
+        # Just like RPN, the proposals should not have gradients
+        boxes = [Boxes(b.detach()) for b in boxes]
+        proposals = []
+        for boxes_per_image, image_size in zip(boxes, image_sizes):
+            boxes_per_image.clip(image_size)
+            if self.training:
+                # do not filter empty boxes at inference time,
+                # because the scores from each stage need to be aligned and added later
+                boxes_per_image = boxes_per_image[boxes_per_image.nonempty()]
+            prop = Instances(image_size)
+            prop.proposal_boxes = boxes_per_image
+            proposals.append(prop)
+        return proposals

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/fast_rcnn.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
+import torch
+from fvcore.nn import smooth_l1_loss
+from torch import nn
+from torch.nn import functional as F
+
+from detectron2.config import configurable
+from detectron2.layers import Linear, ShapeSpec, batched_nms, cat
+from detectron2.modeling.box_regression import Box2BoxTransform, apply_deltas_broadcast
+from detectron2.structures import Boxes, Instances
+from detectron2.utils.events import get_event_storage
+
+__all__ = ["fast_rcnn_inference", "FastRCNNOutputLayers"]
+
+
+logger = logging.getLogger(__name__)
+
+"""
+Shape shorthand in this module:
+
+    N: number of images in the minibatch
+    R: number of ROIs, combined over all images, in the minibatch
+    Ri: number of ROIs in image i
+    K: number of foreground classes. E.g.,there are 80 foreground classes in COCO.
+
+Naming convention:
+
+    deltas: refers to the 4-d (dx, dy, dw, dh) deltas that parameterize the box2box
+    transform (see :class:`box_regression.Box2BoxTransform`).
+
+    pred_class_logits: predicted class scores in [-inf, +inf]; use
+        softmax(pred_class_logits) to estimate P(class).
+
+    gt_classes: ground-truth classification labels in [0, K], where [0, K) represent
+        foreground object classes and K represents the background class.
+
+    pred_proposal_deltas: predicted box2box transform deltas for transforming proposals
+        to detection box predictions.
+
+    gt_proposal_deltas: ground-truth box2box transform deltas
+"""
+
+
+def fast_rcnn_inference(boxes, scores, image_shapes, score_thresh, nms_thresh, topk_per_image):
+    """
+    Call `fast_rcnn_inference_single_image` for all images.
+
+    Args:
+        boxes (list[Tensor]): A list of Tensors of predicted class-specific or class-agnostic
+            boxes for each image. Element i has shape (Ri, K * 4) if doing
+            class-specific regression, or (Ri, 4) if doing class-agnostic
+            regression, where Ri is the number of predicted objects for image i.
+            This is compatible with the output of :meth:`FastRCNNOutputLayers.predict_boxes`.
+        scores (list[Tensor]): A list of Tensors of predicted class scores for each image.
+            Element i has shape (Ri, K + 1), where Ri is the number of predicted objects
+            for image i. Compatible with the output of :meth:`FastRCNNOutputLayers.predict_probs`.
+        image_shapes (list[tuple]): A list of (width, height) tuples for each image in the batch.
+        score_thresh (float): Only return detections with a confidence score exceeding this
+            threshold.
+        nms_thresh (float):  The threshold to use for box non-maximum suppression. Value in [0, 1].
+        topk_per_image (int): The number of top scoring detections to return. Set < 0 to return
+            all detections.
+
+    Returns:
+        instances: (list[Instances]): A list of N instances, one for each image in the batch,
+            that stores the topk most confidence detections.
+        kept_indices: (list[Tensor]): A list of 1D tensor of length of N, each element indicates
+            the corresponding boxes/scores index in [0, Ri) from the input, for image i.
+    """
+    result_per_image = [
+        fast_rcnn_inference_single_image(
+            boxes_per_image, scores_per_image, image_shape, score_thresh, nms_thresh, topk_per_image
+        )
+        for scores_per_image, boxes_per_image, image_shape in zip(scores, boxes, image_shapes)
+    ]
+    return [x[0] for x in result_per_image], [x[1] for x in result_per_image]
+
+
+def fast_rcnn_inference_single_image(
+    boxes, scores, image_shape, score_thresh, nms_thresh, topk_per_image
+):
+    """
+    Single-image inference. Return bounding-box detection results by thresholding
+    on scores and applying non-maximum suppression (NMS).
+
+    Args:
+        Same as `fast_rcnn_inference`, but with boxes, scores, and image shapes
+        per image.
+
+    Returns:
+        Same as `fast_rcnn_inference`, but for only one image.
+    """
+    valid_mask = torch.isfinite(boxes).all(dim=1) & torch.isfinite(scores).all(dim=1)
+    if not valid_mask.all():
+        boxes = boxes[valid_mask]
+        scores = scores[valid_mask]
+
+    scores = scores[:, :-1]
+    num_bbox_reg_classes = boxes.shape[1] // 4
+    # Convert to Boxes to use the `clip` function ...
+    boxes = Boxes(boxes.reshape(-1, 4))
+    boxes.clip(image_shape)
+    boxes = boxes.tensor.view(-1, num_bbox_reg_classes, 4)  # R x C x 4
+
+    # Filter results based on detection scores
+    filter_mask = scores > score_thresh  # R x K
+    # R' x 2. First column contains indices of the R predictions;
+    # Second column contains indices of classes.
+    filter_inds = filter_mask.nonzero()
+    if num_bbox_reg_classes == 1:
+        boxes = boxes[filter_inds[:, 0], 0]
+    else:
+        boxes = boxes[filter_mask]
+    scores = scores[filter_mask]
+
+    # Apply per-class NMS
+    keep = batched_nms(boxes, scores, filter_inds[:, 1], nms_thresh)
+    if topk_per_image >= 0:
+        keep = keep[:topk_per_image]
+    boxes, scores, filter_inds = boxes[keep], scores[keep], filter_inds[keep]
+
+    result = Instances(image_shape)
+    result.pred_boxes = Boxes(boxes)
+    result.scores = scores
+    result.pred_classes = filter_inds[:, 1]
+    return result, filter_inds[:, 0]
+
+
+class FastRCNNOutputs(object):
+    """
+    A class that stores information about outputs of a Fast R-CNN head.
+    It provides methods that are used to decode the outputs of a Fast R-CNN head.
+    """
+
+    def __init__(
+        self,
+        box2box_transform,
+        pred_class_logits,
+        pred_proposal_deltas,
+        proposals,
+        smooth_l1_beta=0,
+    ):
+        """
+        Args:
+            box2box_transform (Box2BoxTransform/Box2BoxTransformRotated):
+                box2box transform instance for proposal-to-detection transformations.
+            pred_class_logits (Tensor): A tensor of shape (R, K + 1) storing the predicted class
+                logits for all R predicted object instances.
+                Each row corresponds to a predicted object instance.
+            pred_proposal_deltas (Tensor): A tensor of shape (R, K * B) or (R, B) for
+                class-specific or class-agnostic regression. It stores the predicted deltas that
+                transform proposals into final box detections.
+                B is the box dimension (4 or 5).
+                When B is 4, each row is [dx, dy, dw, dh (, ....)].
+                When B is 5, each row is [dx, dy, dw, dh, da (, ....)].
+            proposals (list[Instances]): A list of N Instances, where Instances i stores the
+                proposals for image i, in the field "proposal_boxes".
+                When training, each Instances must have ground-truth labels
+                stored in the field "gt_classes" and "gt_boxes".
+                The total number of all instances must be equal to R.
+            smooth_l1_beta (float): The transition point between L1 and L2 loss in
+                the smooth L1 loss function. When set to 0, the loss becomes L1. When
+                set to +inf, the loss becomes constant 0.
+        """
+        self.box2box_transform = box2box_transform
+        self.num_preds_per_image = [len(p) for p in proposals]
+        self.pred_class_logits = pred_class_logits
+        self.pred_proposal_deltas = pred_proposal_deltas
+        self.smooth_l1_beta = smooth_l1_beta
+        self.image_shapes = [x.image_size for x in proposals]
+
+        if len(proposals):
+            box_type = type(proposals[0].proposal_boxes)
+            # cat(..., dim=0) concatenates over all images in the batch
+            self.proposals = box_type.cat([p.proposal_boxes for p in proposals])
+            assert (
+                not self.proposals.tensor.requires_grad
+            ), "Proposals should not require gradients!"
+
+            # The following fields should exist only when training.
+            if proposals[0].has("gt_boxes"):
+                self.gt_boxes = box_type.cat([p.gt_boxes for p in proposals])
+                assert proposals[0].has("gt_classes")
+                self.gt_classes = cat([p.gt_classes for p in proposals], dim=0)
+        else:
+            self.proposals = Boxes(torch.zeros(0, 4, device=self.pred_proposal_deltas.device))
+        self._no_instances = len(proposals) == 0  # no instances found
+
+    def _log_accuracy(self):
+        """
+        Log the accuracy metrics to EventStorage.
+        """
+        num_instances = self.gt_classes.numel()
+        pred_classes = self.pred_class_logits.argmax(dim=1)
+        bg_class_ind = self.pred_class_logits.shape[1] - 1
+
+        fg_inds = (self.gt_classes >= 0) & (self.gt_classes < bg_class_ind)
+        num_fg = fg_inds.nonzero().numel()
+        fg_gt_classes = self.gt_classes[fg_inds]
+        fg_pred_classes = pred_classes[fg_inds]
+
+        num_false_negative = (fg_pred_classes == bg_class_ind).nonzero().numel()
+        num_accurate = (pred_classes == self.gt_classes).nonzero().numel()
+        fg_num_accurate = (fg_pred_classes == fg_gt_classes).nonzero().numel()
+
+        storage = get_event_storage()
+        if num_instances > 0:
+            storage.put_scalar("fast_rcnn/cls_accuracy", num_accurate / num_instances)
+            if num_fg > 0:
+                storage.put_scalar("fast_rcnn/fg_cls_accuracy", fg_num_accurate / num_fg)
+                storage.put_scalar("fast_rcnn/false_negative", num_false_negative / num_fg)
+
+    def softmax_cross_entropy_loss(self):
+        """
+        Compute the softmax cross entropy loss for box classification.
+
+        Returns:
+            scalar Tensor
+        """
+        if self._no_instances:
+            return 0.0 * self.pred_class_logits.sum()
+        else:
+            self._log_accuracy()
+            return F.cross_entropy(self.pred_class_logits, self.gt_classes, reduction="mean")
+
+    def smooth_l1_loss(self):
+        """
+        Compute the smooth L1 loss for box regression.
+
+        Returns:
+            scalar Tensor
+        """
+        if self._no_instances:
+            return 0.0 * self.pred_proposal_deltas.sum()
+        gt_proposal_deltas = self.box2box_transform.get_deltas(
+            self.proposals.tensor, self.gt_boxes.tensor
+        )
+        box_dim = gt_proposal_deltas.size(1)  # 4 or 5
+        cls_agnostic_bbox_reg = self.pred_proposal_deltas.size(1) == box_dim
+        device = self.pred_proposal_deltas.device
+
+        bg_class_ind = self.pred_class_logits.shape[1] - 1
+
+        # Box delta loss is only computed between the prediction for the gt class k
+        # (if 0 <= k < bg_class_ind) and the target; there is no loss defined on predictions
+        # for non-gt classes and background.
+        # Empty fg_inds produces a valid loss of zero as long as the size_average
+        # arg to smooth_l1_loss is False (otherwise it uses torch.mean internally
+        # and would produce a nan loss).
+        fg_inds = torch.nonzero(
+            (self.gt_classes >= 0) & (self.gt_classes < bg_class_ind), as_tuple=True
+        )[0]
+        if cls_agnostic_bbox_reg:
+            # pred_proposal_deltas only corresponds to foreground class for agnostic
+            gt_class_cols = torch.arange(box_dim, device=device)
+        else:
+            fg_gt_classes = self.gt_classes[fg_inds]
+            # pred_proposal_deltas for class k are located in columns [b * k : b * k + b],
+            # where b is the dimension of box representation (4 or 5)
+            # Note that compared to Detectron1,
+            # we do not perform bounding box regression for background classes.
+            gt_class_cols = box_dim * fg_gt_classes[:, None] + torch.arange(box_dim, device=device)
+
+        loss_box_reg = smooth_l1_loss(
+            self.pred_proposal_deltas[fg_inds[:, None], gt_class_cols],
+            gt_proposal_deltas[fg_inds],
+            self.smooth_l1_beta,
+            reduction="sum",
+        )
+        # The loss is normalized using the total number of regions (R), not the number
+        # of foreground regions even though the box regression loss is only defined on
+        # foreground regions. Why? Because doing so gives equal training influence to
+        # each foreground example. To see how, consider two different minibatches:
+        #  (1) Contains a single foreground region
+        #  (2) Contains 100 foreground regions
+        # If we normalize by the number of foreground regions, the single example in
+        # minibatch (1) will be given 100 times as much influence as each foreground
+        # example in minibatch (2). Normalizing by the total number of regions, R,
+        # means that the single example in minibatch (1) and each of the 100 examples
+        # in minibatch (2) are given equal influence.
+        loss_box_reg = loss_box_reg / self.gt_classes.numel()
+        return loss_box_reg
+
+    def _predict_boxes(self):
+        """
+        Returns:
+            Tensor: A Tensors of predicted class-specific or class-agnostic boxes
+                for all images in a batch. Element i has shape (Ri, K * B) or (Ri, B), where Ri is
+                the number of predicted objects for image i and B is the box dimension (4 or 5)
+        """
+        return apply_deltas_broadcast(
+            self.box2box_transform, self.pred_proposal_deltas, self.proposals.tensor
+        )
+
+    """
+    A subclass is expected to have the following methods because
+    they are used to query information about the head predictions.
+    """
+
+    def losses(self):
+        """
+        Compute the default losses for box head in Fast(er) R-CNN,
+        with softmax cross entropy loss and smooth L1 loss.
+
+        Returns:
+            A dict of losses (scalar tensors) containing keys "loss_cls" and "loss_box_reg".
+        """
+        return {
+            "loss_cls": self.softmax_cross_entropy_loss(),
+            "loss_box_reg": self.smooth_l1_loss(),
+        }
+
+    def predict_boxes(self):
+        """
+        Deprecated
+        """
+        return self._predict_boxes().split(self.num_preds_per_image, dim=0)
+
+    def predict_probs(self):
+        """
+        Deprecated
+        """
+        probs = F.softmax(self.pred_class_logits, dim=-1)
+        return probs.split(self.num_preds_per_image, dim=0)
+
+    def inference(self, score_thresh, nms_thresh, topk_per_image):
+        """
+        Deprecated
+        """
+        boxes = self.predict_boxes()
+        scores = self.predict_probs()
+        image_shapes = self.image_shapes
+        return fast_rcnn_inference(
+            boxes, scores, image_shapes, score_thresh, nms_thresh, topk_per_image
+        )
+
+
+class FastRCNNOutputLayers(nn.Module):
+    """
+    Two linear layers for predicting Fast R-CNN outputs:
+      (1) proposal-to-detection box regression deltas
+      (2) classification scores
+    """
+
+    @configurable
+    def __init__(
+        self,
+        input_shape,
+        *,
+        box2box_transform,
+        num_classes,
+        cls_agnostic_bbox_reg=False,
+        smooth_l1_beta=0.0,
+        test_score_thresh=0.0,
+        test_nms_thresh=0.5,
+        test_topk_per_image=100,
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            input_shape (ShapeSpec): shape of the input feature to this module
+            box2box_transform (Box2BoxTransform or Box2BoxTransformRotated):
+            num_classes (int): number of foreground classes
+            cls_agnostic_bbox_reg (bool): whether to use class agnostic for bbox regression
+            smooth_l1_beta (float): transition point from L1 to L2 loss.
+            test_score_thresh (float): threshold to filter predictions results.
+            test_nms_thresh (float): NMS threshold for prediction results.
+            test_topk_per_image (int): number of top predictions to produce per image.
+        """
+        super().__init__()
+        if isinstance(input_shape, int):  # some backward compatibility
+            input_shape = ShapeSpec(channels=input_shape)
+        input_size = input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1)
+        # The prediction layer for num_classes foreground classes and one background class
+        # (hence + 1)
+        self.cls_score = Linear(input_size, num_classes + 1)
+        num_bbox_reg_classes = 1 if cls_agnostic_bbox_reg else num_classes
+        box_dim = len(box2box_transform.weights)
+        self.bbox_pred = Linear(input_size, num_bbox_reg_classes * box_dim)
+
+        nn.init.normal_(self.cls_score.weight, std=0.01)
+        nn.init.normal_(self.bbox_pred.weight, std=0.001)
+        for l in [self.cls_score, self.bbox_pred]:
+            nn.init.constant_(l.bias, 0)
+
+        self.box2box_transform = box2box_transform
+        self.smooth_l1_beta = smooth_l1_beta
+        self.test_score_thresh = test_score_thresh
+        self.test_nms_thresh = test_nms_thresh
+        self.test_topk_per_image = test_topk_per_image
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        return {
+            "input_shape": input_shape,
+            "box2box_transform": Box2BoxTransform(weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS),
+            # fmt: off
+            "num_classes"           : cfg.MODEL.ROI_HEADS.NUM_CLASSES,
+            "cls_agnostic_bbox_reg" : cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG,
+            "smooth_l1_beta"        : cfg.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA,
+            "test_score_thresh"     : cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST,
+            "test_nms_thresh"       : cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST,
+            "test_topk_per_image"   : cfg.TEST.DETECTIONS_PER_IMAGE
+            # fmt: on
+        }
+
+    def forward(self, x):
+        """
+        Returns:
+            Tensor: Nx(K+1) scores for each box
+            Tensor: Nx4 or Nx(Kx4) bounding box regression deltas.
+        """
+        if x.dim() > 2:
+            x = torch.flatten(x, start_dim=1)
+        scores = self.cls_score(x)
+        proposal_deltas = self.bbox_pred(x)
+        return scores, proposal_deltas
+
+    # TODO: move the implementation to this class.
+    def losses(self, predictions, proposals):
+        """
+        Args:
+            predictions: return values of :meth:`forward()`.
+            proposals (list[Instances]): proposals that match the features
+                that were used to compute predictions.
+        """
+        scores, proposal_deltas = predictions
+        return FastRCNNOutputs(
+            self.box2box_transform, scores, proposal_deltas, proposals, self.smooth_l1_beta
+        ).losses()
+
+    def inference(self, predictions, proposals):
+        """
+        Returns:
+            list[Instances]: same as `fast_rcnn_inference`.
+            list[Tensor]: same as `fast_rcnn_inference`.
+        """
+        boxes = self.predict_boxes(predictions, proposals)
+        scores = self.predict_probs(predictions, proposals)
+        image_shapes = [x.image_size for x in proposals]
+        return fast_rcnn_inference(
+            boxes,
+            scores,
+            image_shapes,
+            self.test_score_thresh,
+            self.test_nms_thresh,
+            self.test_topk_per_image,
+        )
+
+    def predict_boxes_for_gt_classes(self, predictions, proposals):
+        """
+        Returns:
+            list[Tensor]: A list of Tensors of predicted boxes for GT classes in case of
+                class-specific box head. Element i of the list has shape (Ri, B), where Ri is
+                the number of predicted objects for image i and B is the box dimension (4 or 5)
+        """
+        if not len(proposals):
+            return []
+        scores, proposal_deltas = predictions
+        proposal_boxes = [p.proposal_boxes for p in proposals]
+        proposal_boxes = proposal_boxes[0].cat(proposal_boxes).tensor
+        N, B = proposal_boxes.shape
+        predict_boxes = apply_deltas_broadcast(
+            self.box2box_transform, proposal_deltas, proposal_boxes
+        )  # Nx(KxB)
+
+        K = predict_boxes.shape[1] // B
+        if K > 1:
+            gt_classes = torch.cat([p.gt_classes for p in proposals], dim=0)
+            # Some proposals are ignored or have a background class. Their gt_classes
+            # cannot be used as index.
+            gt_classes = gt_classes.clamp_(0, K - 1)
+
+            predict_boxes = predict_boxes.view(N, K, B)[
+                torch.arange(N, dtype=torch.long, device=predict_boxes.device), gt_classes
+            ]
+        num_prop_per_image = [len(p) for p in proposals]
+        return predict_boxes.split(num_prop_per_image)
+
+    def predict_boxes(self, predictions, proposals):
+        """
+        Returns:
+            list[Tensor]: A list of Tensors of predicted class-specific or class-agnostic boxes
+                for each image. Element i has shape (Ri, K * B) or (Ri, B), where Ri is
+                the number of predicted objects for image i and B is the box dimension (4 or 5)
+        """
+        if not len(proposals):
+            return []
+        _, proposal_deltas = predictions
+        num_prop_per_image = [len(p) for p in proposals]
+        proposal_boxes = [p.proposal_boxes for p in proposals]
+        proposal_boxes = proposal_boxes[0].cat(proposal_boxes).tensor
+        predict_boxes = apply_deltas_broadcast(
+            self.box2box_transform, proposal_deltas, proposal_boxes
+        )  # Nx(KxB)
+        return predict_boxes.split(num_prop_per_image)
+
+    def predict_probs(self, predictions, proposals):
+        """
+        Returns:
+            list[Tensor]: A list of Tensors of predicted class probabilities for each image.
+                Element i has shape (Ri, K + 1), where Ri is the number of predicted objects
+                for image i.
+        """
+        scores, _ = predictions
+        num_inst_per_image = [len(p) for p in proposals]
+        probs = F.softmax(scores, dim=-1)
+        return probs.split(num_inst_per_image, dim=0)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/keypoint_head.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from typing import List
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from detectron2.config import configurable
+from detectron2.layers import Conv2d, ConvTranspose2d, cat, interpolate
+from detectron2.structures import Instances, heatmaps_to_keypoints
+from detectron2.utils.events import get_event_storage
+from detectron2.utils.registry import Registry
+
+_TOTAL_SKIPPED = 0
+
+ROI_KEYPOINT_HEAD_REGISTRY = Registry("ROI_KEYPOINT_HEAD")
+ROI_KEYPOINT_HEAD_REGISTRY.__doc__ = """
+Registry for keypoint heads, which make keypoint predictions from per-region features.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+"""
+
+
+def build_keypoint_head(cfg, input_shape):
+    """
+    Build a keypoint head from `cfg.MODEL.ROI_KEYPOINT_HEAD.NAME`.
+    """
+    name = cfg.MODEL.ROI_KEYPOINT_HEAD.NAME
+    return ROI_KEYPOINT_HEAD_REGISTRY.get(name)(cfg, input_shape)
+
+
+def keypoint_rcnn_loss(pred_keypoint_logits, instances, normalizer):
+    """
+    Arguments:
+        pred_keypoint_logits (Tensor): A tensor of shape (N, K, S, S) where N is the total number
+            of instances in the batch, K is the number of keypoints, and S is the side length
+            of the keypoint heatmap. The values are spatial logits.
+        instances (list[Instances]): A list of M Instances, where M is the batch size.
+            These instances are predictions from the model
+            that are in 1:1 correspondence with pred_keypoint_logits.
+            Each Instances should contain a `gt_keypoints` field containing a `structures.Keypoint`
+            instance.
+        normalizer (float): Normalize the loss by this amount.
+            If not specified, we normalize by the number of visible keypoints in the minibatch.
+
+    Returns a scalar tensor containing the loss.
+    """
+    heatmaps = []
+    valid = []
+
+    keypoint_side_len = pred_keypoint_logits.shape[2]
+    for instances_per_image in instances:
+        if len(instances_per_image) == 0:
+            continue
+        keypoints = instances_per_image.gt_keypoints
+        heatmaps_per_image, valid_per_image = keypoints.to_heatmap(
+            instances_per_image.proposal_boxes.tensor, keypoint_side_len
+        )
+        heatmaps.append(heatmaps_per_image.view(-1))
+        valid.append(valid_per_image.view(-1))
+
+    if len(heatmaps):
+        keypoint_targets = cat(heatmaps, dim=0)
+        valid = cat(valid, dim=0).to(dtype=torch.uint8)
+        valid = torch.nonzero(valid).squeeze(1)
+
+    # torch.mean (in binary_cross_entropy_with_logits) doesn't
+    # accept empty tensors, so handle it separately
+    if len(heatmaps) == 0 or valid.numel() == 0:
+        global _TOTAL_SKIPPED
+        _TOTAL_SKIPPED += 1
+        storage = get_event_storage()
+        storage.put_scalar("kpts_num_skipped_batches", _TOTAL_SKIPPED, smoothing_hint=False)
+        return pred_keypoint_logits.sum() * 0
+
+    N, K, H, W = pred_keypoint_logits.shape
+    pred_keypoint_logits = pred_keypoint_logits.view(N * K, H * W)
+
+    keypoint_loss = F.cross_entropy(
+        pred_keypoint_logits[valid], keypoint_targets[valid], reduction="sum"
+    )
+
+    # If a normalizer isn't specified, normalize by the number of visible keypoints in the minibatch
+    if normalizer is None:
+        normalizer = valid.numel()
+    keypoint_loss /= normalizer
+
+    return keypoint_loss
+
+
+def keypoint_rcnn_inference(pred_keypoint_logits, pred_instances):
+    """
+    Post process each predicted keypoint heatmap in `pred_keypoint_logits` into (x, y, score)
+        and add it to the `pred_instances` as a `pred_keypoints` field.
+
+    Args:
+        pred_keypoint_logits (Tensor): A tensor of shape (R, K, S, S) where R is the total number
+           of instances in the batch, K is the number of keypoints, and S is the side length of
+           the keypoint heatmap. The values are spatial logits.
+        pred_instances (list[Instances]): A list of N Instances, where N is the number of images.
+
+    Returns:
+        None. Each element in pred_instances will contain an extra "pred_keypoints" field.
+            The field is a tensor of shape (#instance, K, 3) where the last
+            dimension corresponds to (x, y, score).
+            The scores are larger than 0.
+    """
+    # flatten all bboxes from all images together (list[Boxes] -> Rx4 tensor)
+    bboxes_flat = cat([b.pred_boxes.tensor for b in pred_instances], dim=0)
+
+    keypoint_results = heatmaps_to_keypoints(pred_keypoint_logits.detach(), bboxes_flat.detach())
+    num_instances_per_image = [len(i) for i in pred_instances]
+    keypoint_results = keypoint_results[:, :, [0, 1, 3]].split(num_instances_per_image, dim=0)
+
+    for keypoint_results_per_image, instances_per_image in zip(keypoint_results, pred_instances):
+        # keypoint_results_per_image is (num instances)x(num keypoints)x(x, y, score)
+        instances_per_image.pred_keypoints = keypoint_results_per_image
+
+
+class BaseKeypointRCNNHead(nn.Module):
+    """
+    Implement the basic Keypoint R-CNN losses and inference logic described in :paper:`Mask R-CNN`.
+    """
+
+    @configurable
+    def __init__(self, *, num_keypoints, loss_weight=1.0, loss_normalizer=1.0):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            num_keypoints (int): number of keypoints to predict
+            loss_weight (float): weight to multiple on the keypoint loss
+            loss_normalizer (float or str):
+                If float, divide the loss by `loss_normalizer * #images`.
+                If 'visible', the loss is normalized by the total number of
+                    visible keypoints across images.
+        """
+        super().__init__()
+        self.num_keypoints = num_keypoints
+        self.loss_weight = loss_weight
+        assert loss_normalizer == "visible" or isinstance(loss_normalizer, float), loss_normalizer
+        self.loss_normalizer = loss_normalizer
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        ret = {
+            "loss_weight": cfg.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT,
+            "num_keypoints": cfg.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS,
+        }
+        normalize_by_visible = (
+            cfg.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS
+        )  # noqa
+        if not normalize_by_visible:
+            batch_size_per_image = cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE
+            positive_sample_fraction = cfg.MODEL.ROI_HEADS.POSITIVE_FRACTION
+            ret["loss_normalizer"] = (
+                ret["num_keypoints"] * batch_size_per_image * positive_sample_fraction
+            )
+        else:
+            ret["loss_normalizer"] = "visible"
+        return ret
+
+    def forward(self, x, instances: List[Instances]):
+        """
+        Args:
+            x: input region feature(s) provided by :class:`ROIHeads`.
+            instances (list[Instances]): contains the boxes & labels corresponding
+                to the input features.
+                Exact format is up to its caller to decide.
+                Typically, this is the foreground instances in training, with
+                "proposal_boxes" field and other gt annotations.
+                In inference, it contains boxes that are already predicted.
+
+        Returns:
+            A dict of losses if in training. The predicted "instances" if in inference.
+        """
+        x = self.layers(x)
+        if self.training:
+            num_images = len(instances)
+            normalizer = (
+                None if self.loss_normalizer == "visible" else num_images * self.loss_normalizer
+            )
+            return {
+                "loss_keypoint": keypoint_rcnn_loss(x, instances, normalizer=normalizer)
+                * self.loss_weight
+            }
+        else:
+            keypoint_rcnn_inference(x, instances)
+            return instances
+
+    def layers(self, x):
+        """
+        Neural network layers that makes predictions from regional input features.
+        """
+        raise NotImplementedError
+
+
+@ROI_KEYPOINT_HEAD_REGISTRY.register()
+class KRCNNConvDeconvUpsampleHead(BaseKeypointRCNNHead):
+    """
+    A standard keypoint head containing a series of 3x3 convs, followed by
+    a transpose convolution and bilinear interpolation for upsampling.
+    """
+
+    @configurable
+    def __init__(self, input_shape, *, num_keypoints, conv_dims, **kwargs):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            input_shape (ShapeSpec): shape of the input feature
+            conv_dims: an iterable of output channel counts for each conv in the head
+                         e.g. (512, 512, 512) for three convs outputting 512 channels.
+        """
+        super().__init__(num_keypoints=num_keypoints, **kwargs)
+
+        # default up_scale to 2 (this can be made an option)
+        up_scale = 2
+        in_channels = input_shape.channels
+
+        self.blocks = []
+        for idx, layer_channels in enumerate(conv_dims, 1):
+            module = Conv2d(in_channels, layer_channels, 3, stride=1, padding=1)
+            self.add_module("conv_fcn{}".format(idx), module)
+            self.blocks.append(module)
+            in_channels = layer_channels
+
+        deconv_kernel = 4
+        self.score_lowres = ConvTranspose2d(
+            in_channels, num_keypoints, deconv_kernel, stride=2, padding=deconv_kernel // 2 - 1
+        )
+        self.up_scale = up_scale
+
+        for name, param in self.named_parameters():
+            if "bias" in name:
+                nn.init.constant_(param, 0)
+            elif "weight" in name:
+                # Caffe2 implementation uses MSRAFill, which in fact
+                # corresponds to kaiming_normal_ in PyTorch
+                nn.init.kaiming_normal_(param, mode="fan_out", nonlinearity="relu")
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        ret = super().from_config(cfg, input_shape)
+        ret["input_shape"] = input_shape
+        ret["conv_dims"] = cfg.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS
+        return ret
+
+    def layers(self, x):
+        for layer in self.blocks:
+            x = F.relu(layer(x))
+        x = self.score_lowres(x)
+        x = interpolate(x, scale_factor=self.up_scale, mode="bilinear", align_corners=False)
+        return x

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/mask_head.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from typing import List
+import fvcore.nn.weight_init as weight_init
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from detectron2.config import configurable
+from detectron2.layers import Conv2d, ConvTranspose2d, ShapeSpec, cat, get_norm
+from detectron2.structures import Instances
+from detectron2.utils.events import get_event_storage
+from detectron2.utils.registry import Registry
+
+ROI_MASK_HEAD_REGISTRY = Registry("ROI_MASK_HEAD")
+ROI_MASK_HEAD_REGISTRY.__doc__ = """
+Registry for mask heads, which predicts instance masks given
+per-region features.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+"""
+
+
+def mask_rcnn_loss(pred_mask_logits, instances, vis_period=0):
+    """
+    Compute the mask prediction loss defined in the Mask R-CNN paper.
+
+    Args:
+        pred_mask_logits (Tensor): A tensor of shape (B, C, Hmask, Wmask) or (B, 1, Hmask, Wmask)
+            for class-specific or class-agnostic, where B is the total number of predicted masks
+            in all images, C is the number of foreground classes, and Hmask, Wmask are the height
+            and width of the mask predictions. The values are logits.
+        instances (list[Instances]): A list of N Instances, where N is the number of images
+            in the batch. These instances are in 1:1
+            correspondence with the pred_mask_logits. The ground-truth labels (class, box, mask,
+            ...) associated with each instance are stored in fields.
+        vis_period (int): the period (in steps) to dump visualization.
+
+    Returns:
+        mask_loss (Tensor): A scalar tensor containing the loss.
+    """
+    cls_agnostic_mask = pred_mask_logits.size(1) == 1
+    total_num_masks = pred_mask_logits.size(0)
+    mask_side_len = pred_mask_logits.size(2)
+    assert pred_mask_logits.size(2) == pred_mask_logits.size(3), "Mask prediction must be square!"
+
+    gt_classes = []
+    gt_masks = []
+    for instances_per_image in instances:
+        if len(instances_per_image) == 0:
+            continue
+        if not cls_agnostic_mask:
+            gt_classes_per_image = instances_per_image.gt_classes.to(dtype=torch.int64)
+            gt_classes.append(gt_classes_per_image)
+
+        gt_masks_per_image = instances_per_image.gt_masks.crop_and_resize(
+            instances_per_image.proposal_boxes.tensor, mask_side_len
+        ).to(device=pred_mask_logits.device)
+        # A tensor of shape (N, M, M), N=#instances in the image; M=mask_side_len
+        gt_masks.append(gt_masks_per_image)
+
+    if len(gt_masks) == 0:
+        return pred_mask_logits.sum() * 0
+
+    gt_masks = cat(gt_masks, dim=0)
+
+    if cls_agnostic_mask:
+        pred_mask_logits = pred_mask_logits[:, 0]
+    else:
+        indices = torch.arange(total_num_masks)
+        gt_classes = cat(gt_classes, dim=0)
+        pred_mask_logits = pred_mask_logits[indices, gt_classes]
+
+    if gt_masks.dtype == torch.bool:
+        gt_masks_bool = gt_masks
+    else:
+        # Here we allow gt_masks to be float as well (depend on the implementation of rasterize())
+        gt_masks_bool = gt_masks > 0.5
+    gt_masks = gt_masks.to(dtype=torch.float32)
+
+    # Log the training accuracy (using gt classes and 0.5 threshold)
+    mask_incorrect = (pred_mask_logits > 0.0) != gt_masks_bool
+    mask_accuracy = 1 - (mask_incorrect.sum().item() / max(mask_incorrect.numel(), 1.0))
+    num_positive = gt_masks_bool.sum().item()
+    false_positive = (mask_incorrect & ~gt_masks_bool).sum().item() / max(
+        gt_masks_bool.numel() - num_positive, 1.0
+    )
+    false_negative = (mask_incorrect & gt_masks_bool).sum().item() / max(num_positive, 1.0)
+
+    storage = get_event_storage()
+    storage.put_scalar("mask_rcnn/accuracy", mask_accuracy)
+    storage.put_scalar("mask_rcnn/false_positive", false_positive)
+    storage.put_scalar("mask_rcnn/false_negative", false_negative)
+    if vis_period > 0 and storage.iter % vis_period == 0:
+        pred_masks = pred_mask_logits.sigmoid()
+        vis_masks = torch.cat([pred_masks, gt_masks], axis=2)
+        name = "Left: mask prediction;   Right: mask GT"
+        for idx, vis_mask in enumerate(vis_masks):
+            vis_mask = torch.stack([vis_mask] * 3, axis=0)
+            storage.put_image(name + f" ({idx})", vis_mask)
+
+    mask_loss = F.binary_cross_entropy_with_logits(pred_mask_logits, gt_masks, reduction="mean")
+    return mask_loss
+
+
+def mask_rcnn_inference(pred_mask_logits, pred_instances):
+    """
+    Convert pred_mask_logits to estimated foreground probability masks while also
+    extracting only the masks for the predicted classes in pred_instances. For each
+    predicted box, the mask of the same class is attached to the instance by adding a
+    new "pred_masks" field to pred_instances.
+
+    Args:
+        pred_mask_logits (Tensor): A tensor of shape (B, C, Hmask, Wmask) or (B, 1, Hmask, Wmask)
+            for class-specific or class-agnostic, where B is the total number of predicted masks
+            in all images, C is the number of foreground classes, and Hmask, Wmask are the height
+            and width of the mask predictions. The values are logits.
+        pred_instances (list[Instances]): A list of N Instances, where N is the number of images
+            in the batch. Each Instances must have field "pred_classes".
+
+    Returns:
+        None. pred_instances will contain an extra "pred_masks" field storing a mask of size (Hmask,
+            Wmask) for predicted class. Note that the masks are returned as a soft (non-quantized)
+            masks the resolution predicted by the network; post-processing steps, such as resizing
+            the predicted masks to the original image resolution and/or binarizing them, is left
+            to the caller.
+    """
+    cls_agnostic_mask = pred_mask_logits.size(1) == 1
+
+    if cls_agnostic_mask:
+        mask_probs_pred = pred_mask_logits.sigmoid()
+    else:
+        # Select masks corresponding to the predicted classes
+        num_masks = pred_mask_logits.shape[0]
+        class_pred = cat([i.pred_classes for i in pred_instances])
+        indices = torch.arange(num_masks, device=class_pred.device)
+        mask_probs_pred = pred_mask_logits[indices, class_pred][:, None].sigmoid()
+    # mask_probs_pred.shape: (B, 1, Hmask, Wmask)
+
+    num_boxes_per_image = [len(i) for i in pred_instances]
+    mask_probs_pred = mask_probs_pred.split(num_boxes_per_image, dim=0)
+
+    for prob, instances in zip(mask_probs_pred, pred_instances):
+        instances.pred_masks = prob  # (1, Hmask, Wmask)
+
+
+class BaseMaskRCNNHead(nn.Module):
+    """
+    Implement the basic Mask R-CNN losses and inference logic described in :paper:`Mask R-CNN`
+    """
+
+    @configurable
+    def __init__(self, *, vis_period=0):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            vis_period (int): visualization period
+        """
+        super().__init__()
+        self.vis_period = vis_period
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        return {"vis_period": cfg.VIS_PERIOD}
+
+    def forward(self, x, instances: List[Instances]):
+        """
+        Args:
+            x: input region feature(s) provided by :class:`ROIHeads`.
+            instances (list[Instances]): contains the boxes & labels corresponding
+                to the input features.
+                Exact format is up to its caller to decide.
+                Typically, this is the foreground instances in training, with
+                "proposal_boxes" field and other gt annotations.
+                In inference, it contains boxes that are already predicted.
+
+        Returns:
+            A dict of losses in training. The predicted "instances" in inference.
+        """
+        x = self.layers(x)
+        if self.training:
+            return {"loss_mask": mask_rcnn_loss(x, instances, self.vis_period)}
+        else:
+            mask_rcnn_inference(x, instances)
+            return instances
+
+    def layers(self, x):
+        """
+        Neural network layers that makes predictions from input features.
+        """
+        raise NotImplementedError
+
+
+@ROI_MASK_HEAD_REGISTRY.register()
+class MaskRCNNConvUpsampleHead(BaseMaskRCNNHead):
+    """
+    A mask head with several conv layers, plus an upsample layer (with `ConvTranspose2d`).
+    Predictions are made with a final 1x1 conv layer.
+    """
+
+    @configurable
+    def __init__(self, input_shape: ShapeSpec, *, num_classes, conv_dims, conv_norm="", **kwargs):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            input_shape (ShapeSpec): shape of the input feature
+            num_classes (int): the number of classes. 1 if using class agnostic prediction.
+            conv_dims (list[int]): a list of N>0 integers representing the output dimensions
+                of N-1 conv layers and the last upsample layer.
+            conv_norm (str or callable): normalization for the conv layers.
+                See :func:`detectron2.layers.get_norm` for supported types.
+        """
+        super().__init__(**kwargs)
+        assert len(conv_dims) >= 1, "conv_dims have to be non-empty!"
+
+        self.conv_norm_relus = []
+
+        cur_channels = input_shape.channels
+        for k, conv_dim in enumerate(conv_dims[:-1]):
+            conv = Conv2d(
+                cur_channels,
+                conv_dim,
+                kernel_size=3,
+                stride=1,
+                padding=1,
+                bias=not conv_norm,
+                norm=get_norm(conv_norm, conv_dim),
+                activation=F.relu,
+            )
+            self.add_module("mask_fcn{}".format(k + 1), conv)
+            self.conv_norm_relus.append(conv)
+            cur_channels = conv_dim
+
+        self.deconv = ConvTranspose2d(
+            cur_channels, conv_dims[-1], kernel_size=2, stride=2, padding=0
+        )
+        cur_channels = conv_dims[-1]
+
+        self.predictor = Conv2d(cur_channels, num_classes, kernel_size=1, stride=1, padding=0)
+
+        for layer in self.conv_norm_relus + [self.deconv]:
+            weight_init.c2_msra_fill(layer)
+        # use normal distribution initialization for mask prediction layer
+        nn.init.normal_(self.predictor.weight, std=0.001)
+        if self.predictor.bias is not None:
+            nn.init.constant_(self.predictor.bias, 0)
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        ret = super().from_config(cfg, input_shape)
+        conv_dim = cfg.MODEL.ROI_MASK_HEAD.CONV_DIM
+        num_conv = cfg.MODEL.ROI_MASK_HEAD.NUM_CONV
+        ret.update(
+            conv_dims=[conv_dim] * (num_conv + 1),  # +1 for ConvTranspose
+            conv_norm=cfg.MODEL.ROI_MASK_HEAD.NORM,
+            input_shape=input_shape,
+        )
+        if cfg.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK:
+            ret["num_classes"] = 1
+        else:
+            ret["num_classes"] = cfg.MODEL.ROI_HEADS.NUM_CLASSES
+        return ret
+
+    def layers(self, x):
+        for layer in self.conv_norm_relus:
+            x = layer(x)
+        x = F.relu(self.deconv(x))
+        return self.predictor(x)
+
+
+def build_mask_head(cfg, input_shape):
+    """
+    Build a mask head defined by `cfg.MODEL.ROI_MASK_HEAD.NAME`.
+    """
+    name = cfg.MODEL.ROI_MASK_HEAD.NAME
+    return ROI_MASK_HEAD_REGISTRY.get(name)(cfg, input_shape)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/roi_heads.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import inspect
+import logging
+import numpy as np
+from typing import Dict, List, Optional, Tuple, Union
+import torch
+from torch import nn
+
+from detectron2.config import configurable
+from detectron2.layers import ShapeSpec
+from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou
+from detectron2.utils.events import get_event_storage
+from detectron2.utils.registry import Registry
+
+from ..backbone.resnet import BottleneckBlock, make_stage
+from ..matcher import Matcher
+from ..poolers import ROIPooler
+from ..proposal_generator.proposal_utils import add_ground_truth_to_proposals
+from ..sampling import subsample_labels
+from .box_head import build_box_head
+from .fast_rcnn import FastRCNNOutputLayers
+from .keypoint_head import build_keypoint_head
+from .mask_head import build_mask_head
+
+ROI_HEADS_REGISTRY = Registry("ROI_HEADS")
+ROI_HEADS_REGISTRY.__doc__ = """
+Registry for ROI heads in a generalized R-CNN model.
+ROIHeads take feature maps and region proposals, and
+perform per-region computation.
+
+The registered object will be called with `obj(cfg, input_shape)`.
+The call is expected to return an :class:`ROIHeads`.
+"""
+
+logger = logging.getLogger(__name__)
+
+
+def build_roi_heads(cfg, input_shape):
+    """
+    Build ROIHeads defined by `cfg.MODEL.ROI_HEADS.NAME`.
+    """
+    name = cfg.MODEL.ROI_HEADS.NAME
+    return ROI_HEADS_REGISTRY.get(name)(cfg, input_shape)
+
+
+def select_foreground_proposals(
+    proposals: List[Instances], bg_label: int
+) -> Tuple[List[Instances], List[torch.Tensor]]:
+    """
+    Given a list of N Instances (for N images), each containing a `gt_classes` field,
+    return a list of Instances that contain only instances with `gt_classes != -1 &&
+    gt_classes != bg_label`.
+
+    Args:
+        proposals (list[Instances]): A list of N Instances, where N is the number of
+            images in the batch.
+        bg_label: label index of background class.
+
+    Returns:
+        list[Instances]: N Instances, each contains only the selected foreground instances.
+        list[Tensor]: N boolean vector, correspond to the selection mask of
+            each Instances object. True for selected instances.
+    """
+    assert isinstance(proposals, (list, tuple))
+    assert isinstance(proposals[0], Instances)
+    assert proposals[0].has("gt_classes")
+    fg_proposals = []
+    fg_selection_masks = []
+    for proposals_per_image in proposals:
+        gt_classes = proposals_per_image.gt_classes
+        fg_selection_mask = (gt_classes != -1) & (gt_classes != bg_label)
+        fg_idxs = fg_selection_mask.nonzero().squeeze(1)
+        fg_proposals.append(proposals_per_image[fg_idxs])
+        fg_selection_masks.append(fg_selection_mask)
+    return fg_proposals, fg_selection_masks
+
+
+def select_proposals_with_visible_keypoints(proposals: List[Instances]) -> List[Instances]:
+    """
+    Args:
+        proposals (list[Instances]): a list of N Instances, where N is the
+            number of images.
+
+    Returns:
+        proposals: only contains proposals with at least one visible keypoint.
+
+    Note that this is still slightly different from Detectron.
+    In Detectron, proposals for training keypoint head are re-sampled from
+    all the proposals with IOU>threshold & >=1 visible keypoint.
+
+    Here, the proposals are first sampled from all proposals with
+    IOU>threshold, then proposals with no visible keypoint are filtered out.
+    This strategy seems to make no difference on Detectron and is easier to implement.
+    """
+    ret = []
+    all_num_fg = []
+    for proposals_per_image in proposals:
+        # If empty/unannotated image (hard negatives), skip filtering for train
+        if len(proposals_per_image) == 0:
+            ret.append(proposals_per_image)
+            continue
+        gt_keypoints = proposals_per_image.gt_keypoints.tensor
+        # #fg x K x 3
+        vis_mask = gt_keypoints[:, :, 2] >= 1
+        xs, ys = gt_keypoints[:, :, 0], gt_keypoints[:, :, 1]
+        proposal_boxes = proposals_per_image.proposal_boxes.tensor.unsqueeze(dim=1)  # #fg x 1 x 4
+        kp_in_box = (
+            (xs >= proposal_boxes[:, :, 0])
+            & (xs <= proposal_boxes[:, :, 2])
+            & (ys >= proposal_boxes[:, :, 1])
+            & (ys <= proposal_boxes[:, :, 3])
+        )
+        selection = (kp_in_box & vis_mask).any(dim=1)
+        selection_idxs = torch.nonzero(selection, as_tuple=True)[0]
+        all_num_fg.append(selection_idxs.numel())
+        ret.append(proposals_per_image[selection_idxs])
+
+    storage = get_event_storage()
+    storage.put_scalar("keypoint_head/num_fg_samples", np.mean(all_num_fg))
+    return ret
+
+
+class ROIHeads(torch.nn.Module):
+    """
+    ROIHeads perform all per-region computation in an R-CNN.
+
+    It typically contains logic to
+    1. (in training only) match proposals with ground truth and sample them
+    2. crop the regions and extract per-region features using proposals
+    3. make per-region predictions with different heads
+
+    It can have many variants, implemented as subclasses of this class.
+    This base class contains the logic to match/sample proposals.
+    But it is not necessary to inherit this class if the sampling logic is not needed.
+    """
+
+    @configurable
+    def __init__(
+        self,
+        *,
+        num_classes,
+        batch_size_per_image,
+        positive_sample_fraction,
+        proposal_matcher,
+        proposal_append_gt=True
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            num_classes (int): number of classes. Used to label background proposals.
+            batch_size_per_image (int): number of proposals to use for training
+            positive_sample_fraction (float): fraction of positive (foreground) proposals
+                to use for training.
+            proposal_matcher (Matcher): matcher that matches proposals and ground truth
+            proposal_append_gt (bool): whether to include ground truth as proposals as well
+        """
+        super().__init__()
+        self.batch_size_per_image = batch_size_per_image
+        self.positive_sample_fraction = positive_sample_fraction
+        self.num_classes = num_classes
+        self.proposal_matcher = proposal_matcher
+        self.proposal_append_gt = proposal_append_gt
+
+    @classmethod
+    def from_config(cls, cfg):
+        return {
+            "batch_size_per_image": cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE,
+            "positive_sample_fraction": cfg.MODEL.ROI_HEADS.POSITIVE_FRACTION,
+            "num_classes": cfg.MODEL.ROI_HEADS.NUM_CLASSES,
+            "proposal_append_gt": cfg.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT,
+            # Matcher to assign box proposals to gt boxes
+            "proposal_matcher": Matcher(
+                cfg.MODEL.ROI_HEADS.IOU_THRESHOLDS,
+                cfg.MODEL.ROI_HEADS.IOU_LABELS,
+                allow_low_quality_matches=False,
+            ),
+        }
+
+    def _sample_proposals(
+        self, matched_idxs: torch.Tensor, matched_labels: torch.Tensor, gt_classes: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Based on the matching between N proposals and M groundtruth,
+        sample the proposals and set their classification labels.
+
+        Args:
+            matched_idxs (Tensor): a vector of length N, each is the best-matched
+                gt index in [0, M) for each proposal.
+            matched_labels (Tensor): a vector of length N, the matcher's label
+                (one of cfg.MODEL.ROI_HEADS.IOU_LABELS) for each proposal.
+            gt_classes (Tensor): a vector of length M.
+
+        Returns:
+            Tensor: a vector of indices of sampled proposals. Each is in [0, N).
+            Tensor: a vector of the same length, the classification label for
+                each sampled proposal. Each sample is labeled as either a category in
+                [0, num_classes) or the background (num_classes).
+        """
+        has_gt = gt_classes.numel() > 0
+        # Get the corresponding GT for each proposal
+        if has_gt:
+            gt_classes = gt_classes[matched_idxs]
+            # Label unmatched proposals (0 label from matcher) as background (label=num_classes)
+            gt_classes[matched_labels == 0] = self.num_classes
+            # Label ignore proposals (-1 label)
+            gt_classes[matched_labels == -1] = -1
+        else:
+            gt_classes = torch.zeros_like(matched_idxs) + self.num_classes
+
+        sampled_fg_idxs, sampled_bg_idxs = subsample_labels(
+            gt_classes, self.batch_size_per_image, self.positive_sample_fraction, self.num_classes
+        )
+
+        sampled_idxs = torch.cat([sampled_fg_idxs, sampled_bg_idxs], dim=0)
+        return sampled_idxs, gt_classes[sampled_idxs]
+
+    @torch.no_grad()
+    def label_and_sample_proposals(
+        self, proposals: List[Instances], targets: List[Instances]
+    ) -> List[Instances]:
+        """
+        Prepare some proposals to be used to train the ROI heads.
+        It performs box matching between `proposals` and `targets`, and assigns
+        training labels to the proposals.
+        It returns ``self.batch_size_per_image`` random samples from proposals and groundtruth
+        boxes, with a fraction of positives that is no larger than
+        ``self.positive_sample_fraction``.
+
+        Args:
+            See :meth:`ROIHeads.forward`
+
+        Returns:
+            list[Instances]:
+                length `N` list of `Instances`s containing the proposals
+                sampled for training. Each `Instances` has the following fields:
+
+                - proposal_boxes: the proposal boxes
+                - gt_boxes: the ground-truth box that the proposal is assigned to
+                  (this is only meaningful if the proposal has a label > 0; if label = 0
+                  then the ground-truth box is random)
+
+                Other fields such as "gt_classes", "gt_masks", that's included in `targets`.
+        """
+        gt_boxes = [x.gt_boxes for x in targets]
+        # Augment proposals with ground-truth boxes.
+        # In the case of learned proposals (e.g., RPN), when training starts
+        # the proposals will be low quality due to random initialization.
+        # It's possible that none of these initial
+        # proposals have high enough overlap with the gt objects to be used
+        # as positive examples for the second stage components (box head,
+        # cls head, mask head). Adding the gt boxes to the set of proposals
+        # ensures that the second stage components will have some positive
+        # examples from the start of training. For RPN, this augmentation improves
+        # convergence and empirically improves box AP on COCO by about 0.5
+        # points (under one tested configuration).
+        if self.proposal_append_gt:
+            proposals = add_ground_truth_to_proposals(gt_boxes, proposals)
+
+        proposals_with_gt = []
+
+        num_fg_samples = []
+        num_bg_samples = []
+        for proposals_per_image, targets_per_image in zip(proposals, targets):
+            has_gt = len(targets_per_image) > 0
+            match_quality_matrix = pairwise_iou(
+                targets_per_image.gt_boxes, proposals_per_image.proposal_boxes
+            )
+            matched_idxs, matched_labels = self.proposal_matcher(match_quality_matrix)
+            sampled_idxs, gt_classes = self._sample_proposals(
+                matched_idxs, matched_labels, targets_per_image.gt_classes
+            )
+
+            # Set target attributes of the sampled proposals:
+            proposals_per_image = proposals_per_image[sampled_idxs]
+            proposals_per_image.gt_classes = gt_classes
+
+            # We index all the attributes of targets that start with "gt_"
+            # and have not been added to proposals yet (="gt_classes").
+            if has_gt:
+                sampled_targets = matched_idxs[sampled_idxs]
+                # NOTE: here the indexing waste some compute, because heads
+                # like masks, keypoints, etc, will filter the proposals again,
+                # (by foreground/background, or number of keypoints in the image, etc)
+                # so we essentially index the data twice.
+                for (trg_name, trg_value) in targets_per_image.get_fields().items():
+                    if trg_name.startswith("gt_") and not proposals_per_image.has(trg_name):
+                        proposals_per_image.set(trg_name, trg_value[sampled_targets])
+            else:
+                gt_boxes = Boxes(
+                    targets_per_image.gt_boxes.tensor.new_zeros((len(sampled_idxs), 4))
+                )
+                proposals_per_image.gt_boxes = gt_boxes
+
+            num_bg_samples.append((gt_classes == self.num_classes).sum().item())
+            num_fg_samples.append(gt_classes.numel() - num_bg_samples[-1])
+            proposals_with_gt.append(proposals_per_image)
+
+        # Log the number of fg/bg samples that are selected for training ROI heads
+        storage = get_event_storage()
+        storage.put_scalar("roi_head/num_fg_samples", np.mean(num_fg_samples))
+        storage.put_scalar("roi_head/num_bg_samples", np.mean(num_bg_samples))
+
+        return proposals_with_gt
+
+    def forward(
+        self,
+        images: ImageList,
+        features: Dict[str, torch.Tensor],
+        proposals: List[Instances],
+        targets: Optional[List[Instances]] = None,
+    ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]:
+        """
+        Args:
+            images (ImageList):
+            features (dict[str,Tensor]): input data as a mapping from feature
+                map name to tensor. Axis 0 represents the number of images `N` in
+                the input data; axes 1-3 are channels, height, and width, which may
+                vary between feature maps (e.g., if a feature pyramid is used).
+            proposals (list[Instances]): length `N` list of `Instances`. The i-th
+                `Instances` contains object proposals for the i-th input image,
+                with fields "proposal_boxes" and "objectness_logits".
+            targets (list[Instances], optional): length `N` list of `Instances`. The i-th
+                `Instances` contains the ground-truth per-instance annotations
+                for the i-th input image.  Specify `targets` during training only.
+                It may have the following fields:
+
+                - gt_boxes: the bounding box of each instance.
+                - gt_classes: the label for each instance with a category ranging in [0, #class].
+                - gt_masks: PolygonMasks or BitMasks, the ground-truth masks of each instance.
+                - gt_keypoints: NxKx3, the groud-truth keypoints for each instance.
+
+        Returns:
+            list[Instances]: length `N` list of `Instances` containing the
+            detected instances. Returned during inference only; may be [] during training.
+
+            dict[str->Tensor]:
+            mapping from a named loss to a tensor storing the loss. Used during training only.
+        """
+        raise NotImplementedError()
+
+
+@ROI_HEADS_REGISTRY.register()
+class Res5ROIHeads(ROIHeads):
+    """
+    The ROIHeads in a typical "C4" R-CNN model, where
+    the box and mask head share the cropping and
+    the per-region feature computation by a Res5 block.
+    """
+
+    def __init__(self, cfg, input_shape):
+        super().__init__(cfg)
+
+        # fmt: off
+        self.in_features  = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
+        pooler_type       = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE
+        pooler_scales     = (1.0 / input_shape[self.in_features[0]].stride, )
+        sampling_ratio    = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
+        self.mask_on      = cfg.MODEL.MASK_ON
+        # fmt: on
+        assert not cfg.MODEL.KEYPOINT_ON
+        assert len(self.in_features) == 1
+
+        self.pooler = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+
+        self.res5, out_channels = self._build_res5_block(cfg)
+        self.box_predictor = FastRCNNOutputLayers(
+            cfg, ShapeSpec(channels=out_channels, height=1, width=1)
+        )
+
+        if self.mask_on:
+            self.mask_head = build_mask_head(
+                cfg,
+                ShapeSpec(channels=out_channels, width=pooler_resolution, height=pooler_resolution),
+            )
+
+    def _build_res5_block(self, cfg):
+        # fmt: off
+        stage_channel_factor = 2 ** 3  # res5 is 8x res2
+        num_groups           = cfg.MODEL.RESNETS.NUM_GROUPS
+        width_per_group      = cfg.MODEL.RESNETS.WIDTH_PER_GROUP
+        bottleneck_channels  = num_groups * width_per_group * stage_channel_factor
+        out_channels         = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS * stage_channel_factor
+        stride_in_1x1        = cfg.MODEL.RESNETS.STRIDE_IN_1X1
+        norm                 = cfg.MODEL.RESNETS.NORM
+        assert not cfg.MODEL.RESNETS.DEFORM_ON_PER_STAGE[-1], \
+            "Deformable conv is not yet supported in res5 head."
+        # fmt: on
+
+        blocks = make_stage(
+            BottleneckBlock,
+            3,
+            first_stride=2,
+            in_channels=out_channels // 2,
+            bottleneck_channels=bottleneck_channels,
+            out_channels=out_channels,
+            num_groups=num_groups,
+            norm=norm,
+            stride_in_1x1=stride_in_1x1,
+        )
+        return nn.Sequential(*blocks), out_channels
+
+    def _shared_roi_transform(self, features, boxes):
+        x = self.pooler(features, boxes)
+        return self.res5(x)
+
+    def forward(self, images, features, proposals, targets=None):
+        """
+        See :meth:`ROIHeads.forward`.
+        """
+        del images
+
+        if self.training:
+            assert targets
+            proposals = self.label_and_sample_proposals(proposals, targets)
+        del targets
+
+        proposal_boxes = [x.proposal_boxes for x in proposals]
+        box_features = self._shared_roi_transform(
+            [features[f] for f in self.in_features], proposal_boxes
+        )
+        predictions = self.box_predictor(box_features.mean(dim=[2, 3]))
+
+        if self.training:
+            del features
+            losses = self.box_predictor.losses(predictions, proposals)
+            if self.mask_on:
+                proposals, fg_selection_masks = select_foreground_proposals(
+                    proposals, self.num_classes
+                )
+                # Since the ROI feature transform is shared between boxes and masks,
+                # we don't need to recompute features. The mask loss is only defined
+                # on foreground proposals, so we need to select out the foreground
+                # features.
+                mask_features = box_features[torch.cat(fg_selection_masks, dim=0)]
+                del box_features
+                losses.update(self.mask_head(mask_features, proposals))
+            return [], losses
+        else:
+            pred_instances, _ = self.box_predictor.inference(predictions, proposals)
+            pred_instances = self.forward_with_given_boxes(features, pred_instances)
+            return pred_instances, {}
+
+    def forward_with_given_boxes(self, features, instances):
+        """
+        Use the given boxes in `instances` to produce other (non-box) per-ROI outputs.
+
+        Args:
+            features: same as in `forward()`
+            instances (list[Instances]): instances to predict other outputs. Expect the keys
+                "pred_boxes" and "pred_classes" to exist.
+
+        Returns:
+            instances (Instances):
+                the same `Instances` object, with extra
+                fields such as `pred_masks` or `pred_keypoints`.
+        """
+        assert not self.training
+        assert instances[0].has("pred_boxes") and instances[0].has("pred_classes")
+
+        if self.mask_on:
+            features = [features[f] for f in self.in_features]
+            x = self._shared_roi_transform(features, [x.pred_boxes for x in instances])
+            return self.mask_head(x, instances)
+        else:
+            return instances
+
+
+@ROI_HEADS_REGISTRY.register()
+class StandardROIHeads(ROIHeads):
+    """
+    It's "standard" in a sense that there is no ROI transform sharing
+    or feature sharing between tasks.
+    Each head independently processes the input features by each head's
+    own pooler and head.
+
+    This class is used by most models, such as FPN and C5.
+    To implement more models, you can subclass it and implement a different
+    :meth:`forward()` or a head.
+    """
+
+    @configurable
+    def __init__(
+        self,
+        *,
+        box_in_features: List[str],
+        box_pooler: ROIPooler,
+        box_head: nn.Module,
+        box_predictor: nn.Module,
+        mask_in_features: Optional[List[str]] = None,
+        mask_pooler: Optional[ROIPooler] = None,
+        mask_head: Optional[nn.Module] = None,
+        keypoint_in_features: Optional[List[str]] = None,
+        keypoint_pooler: Optional[ROIPooler] = None,
+        keypoint_head: Optional[nn.Module] = None,
+        train_on_pred_boxes: bool = False,
+        **kwargs
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            box_in_features (list[str]): list of feature names to use for the box head.
+            box_pooler (ROIPooler): pooler to extra region features for box head
+            box_head (nn.Module): transform features to make box predictions
+            box_predictor (nn.Module): make box predictions from the feature.
+                Should have the same interface as :class:`FastRCNNOutputLayers`.
+            mask_in_features (list[str]): list of feature names to use for the mask head.
+                None if not using mask head.
+            mask_pooler (ROIPooler): pooler to extra region features for mask head
+            mask_head (nn.Module): transform features to make mask predictions
+            keypoint_in_features, keypoint_pooler, keypoint_head: similar to ``mask*``.
+            train_on_pred_boxes (bool): whether to use proposal boxes or
+                predicted boxes from the box head to train other heads.
+        """
+        super().__init__(**kwargs)
+        # keep self.in_features for backward compatibility
+        self.in_features = self.box_in_features = box_in_features
+        self.box_pooler = box_pooler
+        self.box_head = box_head
+        self.box_predictor = box_predictor
+
+        self.mask_on = mask_in_features is not None
+        if self.mask_on:
+            self.mask_in_features = mask_in_features
+            self.mask_pooler = mask_pooler
+            self.mask_head = mask_head
+        self.keypoint_on = keypoint_in_features is not None
+        if self.keypoint_on:
+            self.keypoint_in_features = keypoint_in_features
+            self.keypoint_pooler = keypoint_pooler
+            self.keypoint_head = keypoint_head
+
+        self.train_on_pred_boxes = train_on_pred_boxes
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        ret = super().from_config(cfg)
+        ret["train_on_pred_boxes"] = cfg.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES
+        # Subclasses that have not been updated to use from_config style construction
+        # may have overridden _init_*_head methods. In this case, those overridden methods
+        # will not be classmethods and we need to avoid trying to call them here.
+        # We test for this with ismethod which only returns True for bound methods of cls.
+        # Such subclasses will need to handle calling their overridden _init_*_head methods.
+        if inspect.ismethod(cls._init_box_head):
+            ret.update(cls._init_box_head(cfg, input_shape))
+        if inspect.ismethod(cls._init_mask_head):
+            ret.update(cls._init_mask_head(cfg, input_shape))
+        if inspect.ismethod(cls._init_keypoint_head):
+            ret.update(cls._init_keypoint_head(cfg, input_shape))
+        return ret
+
+    @classmethod
+    def _init_box_head(cls, cfg, input_shape):
+        # fmt: off
+        in_features       = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
+        pooler_scales     = tuple(1.0 / input_shape[k].stride for k in in_features)
+        sampling_ratio    = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
+        pooler_type       = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE
+        # fmt: on
+
+        # If StandardROIHeads is applied on multiple feature maps (as in FPN),
+        # then we share the same predictors and therefore the channel counts must be the same
+        in_channels = [input_shape[f].channels for f in in_features]
+        # Check all channel counts are equal
+        assert len(set(in_channels)) == 1, in_channels
+        in_channels = in_channels[0]
+
+        box_pooler = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+        # Here we split "box head" and "box predictor", which is mainly due to historical reasons.
+        # They are used together so the "box predictor" layers should be part of the "box head".
+        # New subclasses of ROIHeads do not need "box predictor"s.
+        box_head = build_box_head(
+            cfg, ShapeSpec(channels=in_channels, height=pooler_resolution, width=pooler_resolution)
+        )
+        box_predictor = FastRCNNOutputLayers(cfg, box_head.output_shape)
+        return {
+            "box_in_features": in_features,
+            "box_pooler": box_pooler,
+            "box_head": box_head,
+            "box_predictor": box_predictor,
+        }
+
+    @classmethod
+    def _init_mask_head(cls, cfg, input_shape):
+        if not cfg.MODEL.MASK_ON:
+            return {}
+        # fmt: off
+        in_features       = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution = cfg.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION
+        pooler_scales     = tuple(1.0 / input_shape[k].stride for k in in_features)
+        sampling_ratio    = cfg.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO
+        pooler_type       = cfg.MODEL.ROI_MASK_HEAD.POOLER_TYPE
+        # fmt: on
+
+        in_channels = [input_shape[f].channels for f in in_features][0]
+
+        ret = {"mask_in_features": in_features}
+        ret["mask_pooler"] = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+        ret["mask_head"] = build_mask_head(
+            cfg, ShapeSpec(channels=in_channels, width=pooler_resolution, height=pooler_resolution)
+        )
+        return ret
+
+    @classmethod
+    def _init_keypoint_head(cls, cfg, input_shape):
+        if not cfg.MODEL.KEYPOINT_ON:
+            return {}
+        # fmt: off
+        in_features       = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION
+        pooler_scales     = tuple(1.0 / input_shape[k].stride for k in in_features)  # noqa
+        sampling_ratio    = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO
+        pooler_type       = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE
+        # fmt: on
+
+        in_channels = [input_shape[f].channels for f in in_features][0]
+
+        ret = {"keypoint_in_features": in_features}
+        ret["keypoint_pooler"] = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+        ret["keypoint_head"] = build_keypoint_head(
+            cfg, ShapeSpec(channels=in_channels, width=pooler_resolution, height=pooler_resolution)
+        )
+        return ret
+
+    def forward(
+        self,
+        images: ImageList,
+        features: Dict[str, torch.Tensor],
+        proposals: List[Instances],
+        targets: Optional[List[Instances]] = None,
+    ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]:
+        """
+        See :class:`ROIHeads.forward`.
+        """
+        del images
+        if self.training:
+            assert targets
+            proposals = self.label_and_sample_proposals(proposals, targets)
+        del targets
+
+        if self.training:
+            losses = self._forward_box(features, proposals)
+            # Usually the original proposals used by the box head are used by the mask, keypoint
+            # heads. But when `self.train_on_pred_boxes is True`, proposals will contain boxes
+            # predicted by the box head.
+            losses.update(self._forward_mask(features, proposals))
+            losses.update(self._forward_keypoint(features, proposals))
+            return proposals, losses
+        else:
+            pred_instances = self._forward_box(features, proposals)
+            # During inference cascaded prediction is used: the mask and keypoints heads are only
+            # applied to the top scoring box detections.
+            pred_instances = self.forward_with_given_boxes(features, pred_instances)
+            return pred_instances, {}
+
+    def forward_with_given_boxes(
+        self, features: Dict[str, torch.Tensor], instances: List[Instances]
+    ) -> List[Instances]:
+        """
+        Use the given boxes in `instances` to produce other (non-box) per-ROI outputs.
+
+        This is useful for downstream tasks where a box is known, but need to obtain
+        other attributes (outputs of other heads).
+        Test-time augmentation also uses this.
+
+        Args:
+            features: same as in `forward()`
+            instances (list[Instances]): instances to predict other outputs. Expect the keys
+                "pred_boxes" and "pred_classes" to exist.
+
+        Returns:
+            instances (list[Instances]):
+                the same `Instances` objects, with extra
+                fields such as `pred_masks` or `pred_keypoints`.
+        """
+        assert not self.training
+        assert instances[0].has("pred_boxes") and instances[0].has("pred_classes")
+
+        instances = self._forward_mask(features, instances)
+        instances = self._forward_keypoint(features, instances)
+        return instances
+
+    def _forward_box(
+        self, features: Dict[str, torch.Tensor], proposals: List[Instances]
+    ) -> Union[Dict[str, torch.Tensor], List[Instances]]:
+        """
+        Forward logic of the box prediction branch. If `self.train_on_pred_boxes is True`,
+            the function puts predicted boxes in the `proposal_boxes` field of `proposals` argument.
+
+        Args:
+            features (dict[str, Tensor]): mapping from feature map names to tensor.
+                Same as in :meth:`ROIHeads.forward`.
+            proposals (list[Instances]): the per-image object proposals with
+                their matching ground truth.
+                Each has fields "proposal_boxes", and "objectness_logits",
+                "gt_classes", "gt_boxes".
+
+        Returns:
+            In training, a dict of losses.
+            In inference, a list of `Instances`, the predicted instances.
+        """
+        features = [features[f] for f in self.box_in_features]
+        box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals])
+        box_features = self.box_head(box_features)
+        predictions = self.box_predictor(box_features)
+        del box_features
+
+        if self.training:
+            losses = self.box_predictor.losses(predictions, proposals)
+            # proposals is modified in-place below, so losses must be computed first.
+            if self.train_on_pred_boxes:
+                with torch.no_grad():
+                    pred_boxes = self.box_predictor.predict_boxes_for_gt_classes(
+                        predictions, proposals
+                    )
+                    for proposals_per_image, pred_boxes_per_image in zip(proposals, pred_boxes):
+                        proposals_per_image.proposal_boxes = Boxes(pred_boxes_per_image)
+            return losses
+        else:
+            pred_instances, _ = self.box_predictor.inference(predictions, proposals)
+            return pred_instances
+
+    def _forward_mask(
+        self, features: Dict[str, torch.Tensor], instances: List[Instances]
+    ) -> Union[Dict[str, torch.Tensor], List[Instances]]:
+        """
+        Forward logic of the mask prediction branch.
+
+        Args:
+            features (dict[str, Tensor]): mapping from feature map names to tensor.
+                Same as in :meth:`ROIHeads.forward`.
+            instances (list[Instances]): the per-image instances to train/predict masks.
+                In training, they can be the proposals.
+                In inference, they can be the predicted boxes.
+
+        Returns:
+            In training, a dict of losses.
+            In inference, update `instances` with new fields "pred_masks" and return it.
+        """
+        if not self.mask_on:
+            return {} if self.training else instances
+
+        features = [features[f] for f in self.mask_in_features]
+
+        if self.training:
+            # The loss is only defined on positive proposals.
+            proposals, _ = select_foreground_proposals(instances, self.num_classes)
+            proposal_boxes = [x.proposal_boxes for x in proposals]
+            mask_features = self.mask_pooler(features, proposal_boxes)
+            return self.mask_head(mask_features, proposals)
+        else:
+            pred_boxes = [x.pred_boxes for x in instances]
+            mask_features = self.mask_pooler(features, pred_boxes)
+            return self.mask_head(mask_features, instances)
+
+    def _forward_keypoint(
+        self, features: Dict[str, torch.Tensor], instances: List[Instances]
+    ) -> Union[Dict[str, torch.Tensor], List[Instances]]:
+        """
+        Forward logic of the keypoint prediction branch.
+
+        Args:
+            features (dict[str, Tensor]): mapping from feature map names to tensor.
+                Same as in :meth:`ROIHeads.forward`.
+            instances (list[Instances]): the per-image instances to train/predict keypoints.
+                In training, they can be the proposals.
+                In inference, they can be the predicted boxes.
+
+        Returns:
+            In training, a dict of losses.
+            In inference, update `instances` with new fields "pred_keypoints" and return it.
+        """
+        if not self.keypoint_on:
+            return {} if self.training else instances
+
+        features = [features[f] for f in self.keypoint_in_features]
+
+        if self.training:
+            # The loss is defined on positive proposals with >=1 visible keypoints.
+            proposals, _ = select_foreground_proposals(instances, self.num_classes)
+            proposals = select_proposals_with_visible_keypoints(proposals)
+            proposal_boxes = [x.proposal_boxes for x in proposals]
+
+            keypoint_features = self.keypoint_pooler(features, proposal_boxes)
+            return self.keypoint_head(keypoint_features, proposals)
+        else:
+            pred_boxes = [x.pred_boxes for x in instances]
+            keypoint_features = self.keypoint_pooler(features, pred_boxes)
+            return self.keypoint_head(keypoint_features, instances)

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/roi_heads/rotated_fast_rcnn.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
+import numpy as np
+import torch
+
+from detectron2.config import configurable
+from detectron2.layers import ShapeSpec, batched_nms_rotated
+from detectron2.structures import Instances, RotatedBoxes, pairwise_iou_rotated
+from detectron2.utils.events import get_event_storage
+
+from ..box_regression import Box2BoxTransformRotated
+from ..poolers import ROIPooler
+from ..proposal_generator.proposal_utils import add_ground_truth_to_proposals
+from .box_head import build_box_head
+from .fast_rcnn import FastRCNNOutputLayers
+from .roi_heads import ROI_HEADS_REGISTRY, StandardROIHeads
+
+logger = logging.getLogger(__name__)
+
+"""
+Shape shorthand in this module:
+
+    N: number of images in the minibatch
+    R: number of ROIs, combined over all images, in the minibatch
+    Ri: number of ROIs in image i
+    K: number of foreground classes. E.g.,there are 80 foreground classes in COCO.
+
+Naming convention:
+
+    deltas: refers to the 5-d (dx, dy, dw, dh, da) deltas that parameterize the box2box
+    transform (see :class:`box_regression.Box2BoxTransformRotated`).
+
+    pred_class_logits: predicted class scores in [-inf, +inf]; use
+        softmax(pred_class_logits) to estimate P(class).
+
+    gt_classes: ground-truth classification labels in [0, K], where [0, K) represent
+        foreground object classes and K represents the background class.
+
+    pred_proposal_deltas: predicted rotated box2box transform deltas for transforming proposals
+        to detection box predictions.
+
+    gt_proposal_deltas: ground-truth rotated box2box transform deltas
+"""
+
+
+def fast_rcnn_inference_rotated(
+    boxes, scores, image_shapes, score_thresh, nms_thresh, topk_per_image
+):
+    """
+    Call `fast_rcnn_inference_single_image_rotated` for all images.
+
+    Args:
+        boxes (list[Tensor]): A list of Tensors of predicted class-specific or class-agnostic
+            boxes for each image. Element i has shape (Ri, K * 5) if doing
+            class-specific regression, or (Ri, 5) if doing class-agnostic
+            regression, where Ri is the number of predicted objects for image i.
+            This is compatible with the output of :meth:`FastRCNNOutputs.predict_boxes`.
+        scores (list[Tensor]): A list of Tensors of predicted class scores for each image.
+            Element i has shape (Ri, K + 1), where Ri is the number of predicted objects
+            for image i. Compatible with the output of :meth:`FastRCNNOutputs.predict_probs`.
+        image_shapes (list[tuple]): A list of (width, height) tuples for each image in the batch.
+        score_thresh (float): Only return detections with a confidence score exceeding this
+            threshold.
+        nms_thresh (float):  The threshold to use for box non-maximum suppression. Value in [0, 1].
+        topk_per_image (int): The number of top scoring detections to return. Set < 0 to return
+            all detections.
+
+    Returns:
+        instances: (list[Instances]): A list of N instances, one for each image in the batch,
+            that stores the topk most confidence detections.
+        kept_indices: (list[Tensor]): A list of 1D tensor of length of N, each element indicates
+            the corresponding boxes/scores index in [0, Ri) from the input, for image i.
+    """
+    result_per_image = [
+        fast_rcnn_inference_single_image_rotated(
+            boxes_per_image, scores_per_image, image_shape, score_thresh, nms_thresh, topk_per_image
+        )
+        for scores_per_image, boxes_per_image, image_shape in zip(scores, boxes, image_shapes)
+    ]
+    return [x[0] for x in result_per_image], [x[1] for x in result_per_image]
+
+
+def fast_rcnn_inference_single_image_rotated(
+    boxes, scores, image_shape, score_thresh, nms_thresh, topk_per_image
+):
+    """
+    Single-image inference. Return rotated bounding-box detection results by thresholding
+    on scores and applying rotated non-maximum suppression (Rotated NMS).
+
+    Args:
+        Same as `fast_rcnn_inference_rotated`, but with rotated boxes, scores, and image shapes
+        per image.
+
+    Returns:
+        Same as `fast_rcnn_inference_rotated`, but for only one image.
+    """
+    valid_mask = torch.isfinite(boxes).all(dim=1) & torch.isfinite(scores).all(dim=1)
+    if not valid_mask.all():
+        boxes = boxes[valid_mask]
+        scores = scores[valid_mask]
+
+    B = 5  # box dimension
+    scores = scores[:, :-1]
+    num_bbox_reg_classes = boxes.shape[1] // B
+    # Convert to Boxes to use the `clip` function ...
+    boxes = RotatedBoxes(boxes.reshape(-1, B))
+    boxes.clip(image_shape)
+    boxes = boxes.tensor.view(-1, num_bbox_reg_classes, B)  # R x C x B
+    # Filter results based on detection scores
+    filter_mask = scores > score_thresh  # R x K
+    # R' x 2. First column contains indices of the R predictions;
+    # Second column contains indices of classes.
+    filter_inds = filter_mask.nonzero()
+    if num_bbox_reg_classes == 1:
+        boxes = boxes[filter_inds[:, 0], 0]
+    else:
+        boxes = boxes[filter_mask]
+    scores = scores[filter_mask]
+
+    # Apply per-class Rotated NMS
+    keep = batched_nms_rotated(boxes, scores, filter_inds[:, 1], nms_thresh)
+    if topk_per_image >= 0:
+        keep = keep[:topk_per_image]
+    boxes, scores, filter_inds = boxes[keep], scores[keep], filter_inds[keep]
+
+    result = Instances(image_shape)
+    result.pred_boxes = RotatedBoxes(boxes)
+    result.scores = scores
+    result.pred_classes = filter_inds[:, 1]
+
+    return result, filter_inds[:, 0]
+
+
+class RotatedFastRCNNOutputLayers(FastRCNNOutputLayers):
+    """
+    Two linear layers for predicting Rotated Fast R-CNN outputs.
+    """
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        args = super().from_config(cfg, input_shape)
+        args["box2box_transform"] = Box2BoxTransformRotated(
+            weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS
+        )
+        return args
+
+    def inference(self, predictions, proposals):
+        """
+        Returns:
+            list[Instances]: same as `fast_rcnn_inference_rotated`.
+            list[Tensor]: same as `fast_rcnn_inference_rotated`.
+        """
+        boxes = self.predict_boxes(predictions, proposals)
+        scores = self.predict_probs(predictions, proposals)
+        image_shapes = [x.image_size for x in proposals]
+
+        return fast_rcnn_inference_rotated(
+            boxes,
+            scores,
+            image_shapes,
+            self.test_score_thresh,
+            self.test_nms_thresh,
+            self.test_topk_per_image,
+        )
+
+
+@ROI_HEADS_REGISTRY.register()
+class RROIHeads(StandardROIHeads):
+    """
+    This class is used by Rotated Fast R-CNN to detect rotated boxes.
+    For now, it only supports box predictions but not mask or keypoints.
+    """
+
+    @configurable
+    def __init__(self, **kwargs):
+        """
+        NOTE: this interface is experimental.
+        """
+        super().__init__(**kwargs)
+        assert (
+            not self.mask_on and not self.keypoint_on
+        ), "Mask/Keypoints not supported in Rotated ROIHeads."
+        assert not self.train_on_pred_boxes, "train_on_pred_boxes not implemented for RROIHeads!"
+
+    @classmethod
+    def _init_box_head(cls, cfg, input_shape):
+        # fmt: off
+        in_features       = cfg.MODEL.ROI_HEADS.IN_FEATURES
+        pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION
+        pooler_scales     = tuple(1.0 / input_shape[k].stride for k in in_features)
+        sampling_ratio    = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO
+        pooler_type       = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE
+        # fmt: on
+        assert pooler_type in ["ROIAlignRotated"], pooler_type
+        # assume all channel counts are equal
+        in_channels = [input_shape[f].channels for f in in_features][0]
+
+        box_pooler = ROIPooler(
+            output_size=pooler_resolution,
+            scales=pooler_scales,
+            sampling_ratio=sampling_ratio,
+            pooler_type=pooler_type,
+        )
+        box_head = build_box_head(
+            cfg, ShapeSpec(channels=in_channels, height=pooler_resolution, width=pooler_resolution)
+        )
+        # This line is the only difference v.s. StandardROIHeads
+        box_predictor = RotatedFastRCNNOutputLayers(cfg, box_head.output_shape)
+        return {
+            "box_in_features": in_features,
+            "box_pooler": box_pooler,
+            "box_head": box_head,
+            "box_predictor": box_predictor,
+        }
+
+    @torch.no_grad()
+    def label_and_sample_proposals(self, proposals, targets):
+        """
+        Prepare some proposals to be used to train the RROI heads.
+        It performs box matching between `proposals` and `targets`, and assigns
+        training labels to the proposals.
+        It returns `self.batch_size_per_image` random samples from proposals and groundtruth boxes,
+        with a fraction of positives that is no larger than `self.positive_sample_fraction.
+
+        Args:
+            See :meth:`StandardROIHeads.forward`
+
+        Returns:
+            list[Instances]: length `N` list of `Instances`s containing the proposals
+                sampled for training. Each `Instances` has the following fields:
+                - proposal_boxes: the rotated proposal boxes
+                - gt_boxes: the ground-truth rotated boxes that the proposal is assigned to
+                  (this is only meaningful if the proposal has a label > 0; if label = 0
+                   then the ground-truth box is random)
+                - gt_classes: the ground-truth classification lable for each proposal
+        """
+        gt_boxes = [x.gt_boxes for x in targets]
+        if self.proposal_append_gt:
+            proposals = add_ground_truth_to_proposals(gt_boxes, proposals)
+
+        proposals_with_gt = []
+
+        num_fg_samples = []
+        num_bg_samples = []
+        for proposals_per_image, targets_per_image in zip(proposals, targets):
+            has_gt = len(targets_per_image) > 0
+            match_quality_matrix = pairwise_iou_rotated(
+                targets_per_image.gt_boxes, proposals_per_image.proposal_boxes
+            )
+            matched_idxs, matched_labels = self.proposal_matcher(match_quality_matrix)
+            sampled_idxs, gt_classes = self._sample_proposals(
+                matched_idxs, matched_labels, targets_per_image.gt_classes
+            )
+
+            proposals_per_image = proposals_per_image[sampled_idxs]
+            proposals_per_image.gt_classes = gt_classes
+
+            if has_gt:
+                sampled_targets = matched_idxs[sampled_idxs]
+                proposals_per_image.gt_boxes = targets_per_image.gt_boxes[sampled_targets]
+            else:
+                gt_boxes = RotatedBoxes(
+                    targets_per_image.gt_boxes.tensor.new_zeros((len(sampled_idxs), 5))
+                )
+                proposals_per_image.gt_boxes = gt_boxes
+
+            num_bg_samples.append((gt_classes == self.num_classes).sum().item())
+            num_fg_samples.append(gt_classes.numel() - num_bg_samples[-1])
+            proposals_with_gt.append(proposals_per_image)
+
+        # Log the number of fg/bg samples that are selected for training ROI heads
+        storage = get_event_storage()
+        storage.put_scalar("roi_head/num_fg_samples", np.mean(num_fg_samples))
+        storage.put_scalar("roi_head/num_bg_samples", np.mean(num_bg_samples))
+
+        return proposals_with_gt

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/sampling.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import torch
+
+__all__ = ["subsample_labels"]
+
+
+def subsample_labels(labels, num_samples, positive_fraction, bg_label):
+    """
+    Return `num_samples` (or fewer, if not enough found)
+    random samples from `labels` which is a mixture of positives & negatives.
+    It will try to return as many positives as possible without
+    exceeding `positive_fraction * num_samples`, and then try to
+    fill the remaining slots with negatives.
+
+    Args:
+        labels (Tensor): (N, ) label vector with values:
+            * -1: ignore
+            * bg_label: background ("negative") class
+            * otherwise: one or more foreground ("positive") classes
+        num_samples (int): The total number of labels with value >= 0 to return.
+            Values that are not sampled will be filled with -1 (ignore).
+        positive_fraction (float): The number of subsampled labels with values > 0
+            is `min(num_positives, int(positive_fraction * num_samples))`. The number
+            of negatives sampled is `min(num_negatives, num_samples - num_positives_sampled)`.
+            In order words, if there are not enough positives, the sample is filled with
+            negatives. If there are also not enough negatives, then as many elements are
+            sampled as is possible.
+        bg_label (int): label index of background ("negative") class.
+
+    Returns:
+        pos_idx, neg_idx (Tensor):
+            1D vector of indices. The total length of both is `num_samples` or fewer.
+    """
+    positive = torch.nonzero((labels != -1) & (labels != bg_label), as_tuple=True)[0]
+    negative = torch.nonzero(labels == bg_label, as_tuple=True)[0]
+
+    num_pos = int(num_samples * positive_fraction)
+    # protect against not enough positive examples
+    num_pos = min(positive.numel(), num_pos)
+    num_neg = num_samples - num_pos
+    # protect against not enough negative examples
+    num_neg = min(negative.numel(), num_neg)
+
+    # randomly select positive and negative examples
+    perm1 = torch.randperm(positive.numel(), device=positive.device)[:num_pos]
+    perm2 = torch.randperm(negative.numel(), device=negative.device)[:num_neg]
+
+    pos_idx = positive[perm1]
+    neg_idx = negative[perm2]
+    return pos_idx, neg_idx

preprocess/humanparsing/mhp_extension/detectron2/detectron2/modeling/test_time_augmentation.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import copy
+import numpy as np
+from contextlib import contextmanager
+from itertools import count
+import torch
+from torch import nn
+from torch.nn.parallel import DistributedDataParallel
+
+from detectron2.data.detection_utils import read_image
+from detectron2.data.transforms import ResizeShortestEdge
+from detectron2.structures import Instances
+
+from .meta_arch import GeneralizedRCNN
+from .postprocessing import detector_postprocess
+from .roi_heads.fast_rcnn import fast_rcnn_inference_single_image
+
+__all__ = ["DatasetMapperTTA", "GeneralizedRCNNWithTTA"]
+
+
+class DatasetMapperTTA:
+    """
+    Implement test-time augmentation for detection data.
+    It is a callable which takes a dataset dict from a detection dataset,
+    and returns a list of dataset dicts where the images
+    are augmented from the input image by the transformations defined in the config.
+    This is used for test-time augmentation.
+    """
+
+    def __init__(self, cfg):
+        self.min_sizes = cfg.TEST.AUG.MIN_SIZES
+        self.max_size = cfg.TEST.AUG.MAX_SIZE
+        self.flip = cfg.TEST.AUG.FLIP
+        self.image_format = cfg.INPUT.FORMAT
+
+    def __call__(self, dataset_dict):
+        """
+        Args:
+            dict: a detection dataset dict
+
+        Returns:
+            list[dict]:
+                a list of dataset dicts, which contain augmented version of the input image.
+                The total number of dicts is ``len(min_sizes) * (2 if flip else 1)``.
+        """
+        ret = []
+        if "image" not in dataset_dict:
+            numpy_image = read_image(dataset_dict["file_name"], self.image_format)
+        else:
+            numpy_image = dataset_dict["image"].permute(1, 2, 0).numpy().astype("uint8")
+        for min_size in self.min_sizes:
+            image = np.copy(numpy_image)
+            tfm = ResizeShortestEdge(min_size, self.max_size).get_transform(image)
+            resized = tfm.apply_image(image)
+            resized = torch.as_tensor(resized.transpose(2, 0, 1).astype("float32"))
+
+            dic = copy.deepcopy(dataset_dict)
+            dic["horiz_flip"] = False
+            dic["image"] = resized
+            ret.append(dic)
+
+            if self.flip:
+                dic = copy.deepcopy(dataset_dict)
+                dic["horiz_flip"] = True
+                dic["image"] = torch.flip(resized, dims=[2])
+                ret.append(dic)
+        return ret
+
+
+class GeneralizedRCNNWithTTA(nn.Module):
+    """
+    A GeneralizedRCNN with test-time augmentation enabled.
+    Its :meth:`__call__` method has the same interface as :meth:`GeneralizedRCNN.forward`.
+    """
+
+    def __init__(self, cfg, model, tta_mapper=None, batch_size=3):
+        """
+        Args:
+            cfg (CfgNode):
+            model (GeneralizedRCNN): a GeneralizedRCNN to apply TTA on.
+            tta_mapper (callable): takes a dataset dict and returns a list of
+                augmented versions of the dataset dict. Defaults to
+                `DatasetMapperTTA(cfg)`.
+            batch_size (int): batch the augmented images into this batch size for inference.
+        """
+        super().__init__()
+        if isinstance(model, DistributedDataParallel):
+            model = model.module
+        assert isinstance(
+            model, GeneralizedRCNN
+        ), "TTA is only supported on GeneralizedRCNN. Got a model of type {}".format(type(model))
+        self.cfg = cfg.clone()
+        assert not self.cfg.MODEL.KEYPOINT_ON, "TTA for keypoint is not supported yet"
+        assert (
+            not self.cfg.MODEL.LOAD_PROPOSALS
+        ), "TTA for pre-computed proposals is not supported yet"
+
+        self.model = model
+
+        if tta_mapper is None:
+            tta_mapper = DatasetMapperTTA(cfg)
+        self.tta_mapper = tta_mapper
+        self.batch_size = batch_size
+
+    @contextmanager
+    def _turn_off_roi_heads(self, attrs):
+        """
+        Open a context where some heads in `model.roi_heads` are temporarily turned off.
+        Args:
+            attr (list[str]): the attribute in `model.roi_heads` which can be used
+                to turn off a specific head, e.g., "mask_on", "keypoint_on".
+        """
+        roi_heads = self.model.roi_heads
+        old = {}
+        for attr in attrs:
+            try:
+                old[attr] = getattr(roi_heads, attr)
+            except AttributeError:
+                # The head may not be implemented in certain ROIHeads
+                pass
+
+        if len(old.keys()) == 0:
+            yield
+        else:
+            for attr in old.keys():
+                setattr(roi_heads, attr, False)
+            yield
+            for attr in old.keys():
+                setattr(roi_heads, attr, old[attr])
+
+    def _batch_inference(self, batched_inputs, detected_instances=None, do_postprocess=True):
+        """
+        Execute inference on a list of inputs,
+        using batch size = self.batch_size, instead of the length of the list.
+
+        Inputs & outputs have the same format as :meth:`GeneralizedRCNN.inference`
+        """
+        if detected_instances is None:
+            detected_instances = [None] * len(batched_inputs)
+
+        outputs = []
+        inputs, instances = [], []
+        for idx, input, instance in zip(count(), batched_inputs, detected_instances):
+            inputs.append(input)
+            instances.append(instance)
+            if len(inputs) == self.batch_size or idx == len(batched_inputs) - 1:
+                outputs.extend(
+                    self.model.inference(
+                        inputs,
+                        instances if instances[0] is not None else None,
+                        do_postprocess=do_postprocess,
+                    )
+                )
+                inputs, instances = [], []
+        return outputs
+
+    def __call__(self, batched_inputs):
+        """
+        Same input/output format as :meth:`GeneralizedRCNN.forward`
+        """
+        return [self._inference_one_image(x) for x in batched_inputs]
+
+    def _detector_postprocess(self, outputs, aug_vars):
+        return detector_postprocess(outputs, aug_vars["height"], aug_vars["width"])
+
+    def _inference_one_image(self, input):
+        """
+        Args:
+            input (dict): one dataset dict
+
+        Returns:
+            dict: one output dict
+        """
+
+        augmented_inputs, aug_vars = self._get_augmented_inputs(input)
+        # Detect boxes from all augmented versions
+        with self._turn_off_roi_heads(["mask_on", "keypoint_on"]):
+            # temporarily disable roi heads
+            all_boxes, all_scores, all_classes = self._get_augmented_boxes(
+                augmented_inputs, aug_vars
+            )
+        merged_instances = self._merge_detections(
+            all_boxes, all_scores, all_classes, (aug_vars["height"], aug_vars["width"])
+        )
+
+        if self.cfg.MODEL.MASK_ON:
+            # Use the detected boxes to obtain new fields
+            augmented_instances = self._rescale_detected_boxes(
+                augmented_inputs, merged_instances, aug_vars
+            )
+            # run forward on the detected boxes
+            outputs = self._batch_inference(
+                augmented_inputs, augmented_instances, do_postprocess=False
+            )
+            # Delete now useless variables to avoid being out of memory
+            del augmented_inputs, augmented_instances, merged_instances
+            # average the predictions
+            outputs[0].pred_masks = self._reduce_pred_masks(outputs, aug_vars)
+            # postprocess
+            output = self._detector_postprocess(outputs[0], aug_vars)
+            return {"instances": output}
+        else:
+            return {"instances": merged_instances}
+
+    def _get_augmented_inputs(self, input):
+        augmented_inputs = self.tta_mapper(input)
+
+        do_hflip = [k.pop("horiz_flip", False) for k in augmented_inputs]
+        heights = [k["height"] for k in augmented_inputs]
+        widths = [k["width"] for k in augmented_inputs]
+        assert (
+            len(set(heights)) == 1 and len(set(widths)) == 1
+        ), "Augmented version of the inputs should have the same original resolution!"
+        height = heights[0]
+        width = widths[0]
+        aug_vars = {"height": height, "width": width, "do_hflip": do_hflip}
+
+        return augmented_inputs, aug_vars
+
+    def _get_augmented_boxes(self, augmented_inputs, aug_vars):
+        # 1: forward with all augmented images
+        outputs = self._batch_inference(augmented_inputs, do_postprocess=False)
+        # 2: union the results
+        all_boxes = []
+        all_scores = []
+        all_classes = []
+        for idx, output in enumerate(outputs):
+            rescaled_output = self._detector_postprocess(output, aug_vars)
+            pred_boxes = rescaled_output.pred_boxes.tensor
+            if aug_vars["do_hflip"][idx]:
+                pred_boxes[:, [0, 2]] = aug_vars["width"] - pred_boxes[:, [2, 0]]
+            all_boxes.append(pred_boxes)
+            all_scores.extend(rescaled_output.scores)
+            all_classes.extend(rescaled_output.pred_classes)
+        all_boxes = torch.cat(all_boxes, dim=0).cpu()
+        return all_boxes, all_scores, all_classes
+
+    def _merge_detections(self, all_boxes, all_scores, all_classes, shape_hw):
+        # select from the union of all results
+        num_boxes = len(all_boxes)
+        num_classes = self.cfg.MODEL.ROI_HEADS.NUM_CLASSES
+        # +1 because fast_rcnn_inference expects background scores as well
+        all_scores_2d = torch.zeros(num_boxes, num_classes + 1, device=all_boxes.device)
+        for idx, cls, score in zip(count(), all_classes, all_scores):
+            all_scores_2d[idx, cls] = score
+
+        merged_instances, _ = fast_rcnn_inference_single_image(
+            all_boxes,
+            all_scores_2d,
+            shape_hw,
+            1e-8,
+            self.cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST,
+            self.cfg.TEST.DETECTIONS_PER_IMAGE,
+        )
+
+        return merged_instances
+
+    def _rescale_detected_boxes(self, augmented_inputs, merged_instances, aug_vars):
+        augmented_instances = []
+        for idx, input in enumerate(augmented_inputs):
+            actual_height, actual_width = input["image"].shape[1:3]
+            scale_x = actual_width * 1.0 / aug_vars["width"]
+            scale_y = actual_height * 1.0 / aug_vars["height"]
+            pred_boxes = merged_instances.pred_boxes.clone()
+            pred_boxes.tensor[:, 0::2] *= scale_x
+            pred_boxes.tensor[:, 1::2] *= scale_y
+            if aug_vars["do_hflip"][idx]:
+                pred_boxes.tensor[:, [0, 2]] = actual_width - pred_boxes.tensor[:, [2, 0]]
+
+            aug_instances = Instances(
+                image_size=(actual_height, actual_width),
+                pred_boxes=pred_boxes,
+                pred_classes=merged_instances.pred_classes,
+                scores=merged_instances.scores,
+            )
+            augmented_instances.append(aug_instances)
+        return augmented_instances
+
+    def _reduce_pred_masks(self, outputs, aug_vars):
+        for idx, output in enumerate(outputs):
+            if aug_vars["do_hflip"][idx]:
+                output.pred_masks = output.pred_masks.flip(dims=[3])
+        all_pred_masks = torch.stack([o.pred_masks for o in outputs], dim=0)
+        avg_pred_masks = torch.mean(all_pred_masks, dim=0)
+        return avg_pred_masks

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

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .build import build_lr_scheduler, build_optimizer
+from .lr_scheduler import WarmupCosineLR, WarmupMultiStepLR
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

preprocess/humanparsing/mhp_extension/detectron2/detectron2/solver/build.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from enum import Enum
+from typing import Any, Callable, Dict, Iterable, List, Set, Type, Union
+import torch
+
+from detectron2.config import CfgNode
+
+from .lr_scheduler import WarmupCosineLR, WarmupMultiStepLR
+
+_GradientClipperInput = Union[torch.Tensor, Iterable[torch.Tensor]]
+_GradientClipper = Callable[[_GradientClipperInput], None]
+
+
+class GradientClipType(Enum):
+    VALUE = "value"
+    NORM = "norm"
+
+
+def _create_gradient_clipper(cfg: CfgNode) -> _GradientClipper:
+    """
+    Creates gradient clipping closure to clip by value or by norm,
+    according to the provided config.
+    """
+    cfg = cfg.clone()
+
+    def clip_grad_norm(p: _GradientClipperInput):
+        torch.nn.utils.clip_grad_norm_(p, cfg.CLIP_VALUE, cfg.NORM_TYPE)
+
+    def clip_grad_value(p: _GradientClipperInput):
+        torch.nn.utils.clip_grad_value_(p, cfg.CLIP_VALUE)
+
+    _GRADIENT_CLIP_TYPE_TO_CLIPPER = {
+        GradientClipType.VALUE: clip_grad_value,
+        GradientClipType.NORM: clip_grad_norm,
+    }
+    return _GRADIENT_CLIP_TYPE_TO_CLIPPER[GradientClipType(cfg.CLIP_TYPE)]
+
+
+def _generate_optimizer_class_with_gradient_clipping(
+    optimizer_type: Type[torch.optim.Optimizer], gradient_clipper: _GradientClipper
+) -> Type[torch.optim.Optimizer]:
+    """
+    Dynamically creates a new type that inherits the type of a given instance
+    and overrides the `step` method to add gradient clipping
+    """
+
+    def optimizer_wgc_step(self, closure=None):
+        for group in self.param_groups:
+            for p in group["params"]:
+                gradient_clipper(p)
+        super(type(self), self).step(closure)
+
+    OptimizerWithGradientClip = type(
+        optimizer_type.__name__ + "WithGradientClip",
+        (optimizer_type,),
+        {"step": optimizer_wgc_step},
+    )
+    return OptimizerWithGradientClip
+
+
+def maybe_add_gradient_clipping(
+    cfg: CfgNode, optimizer: torch.optim.Optimizer
+) -> torch.optim.Optimizer:
+    """
+    If gradient clipping is enabled through config options, wraps the existing
+    optimizer instance of some type OptimizerType to become an instance
+    of the new dynamically created class OptimizerTypeWithGradientClip
+    that inherits OptimizerType and overrides the `step` method to
+    include gradient clipping.
+
+    Args:
+        cfg: CfgNode
+            configuration options
+        optimizer: torch.optim.Optimizer
+            existing optimizer instance
+
+    Return:
+        optimizer: torch.optim.Optimizer
+            either the unmodified optimizer instance (if gradient clipping is
+            disabled), or the same instance with adjusted __class__ to override
+            the `step` method and include gradient clipping
+    """
+    if not cfg.SOLVER.CLIP_GRADIENTS.ENABLED:
+        return optimizer
+    grad_clipper = _create_gradient_clipper(cfg.SOLVER.CLIP_GRADIENTS)
+    OptimizerWithGradientClip = _generate_optimizer_class_with_gradient_clipping(
+        type(optimizer), grad_clipper
+    )
+    optimizer.__class__ = OptimizerWithGradientClip
+    return optimizer
+
+
+def build_optimizer(cfg: CfgNode, model: torch.nn.Module) -> torch.optim.Optimizer:
+    """
+    Build an optimizer from config.
+    """
+    norm_module_types = (
+        torch.nn.BatchNorm1d,
+        torch.nn.BatchNorm2d,
+        torch.nn.BatchNorm3d,
+        torch.nn.SyncBatchNorm,
+        # NaiveSyncBatchNorm inherits from BatchNorm2d
+        torch.nn.GroupNorm,
+        torch.nn.InstanceNorm1d,
+        torch.nn.InstanceNorm2d,
+        torch.nn.InstanceNorm3d,
+        torch.nn.LayerNorm,
+        torch.nn.LocalResponseNorm,
+    )
+    params: List[Dict[str, Any]] = []
+    memo: Set[torch.nn.parameter.Parameter] = set()
+    for module in model.modules():
+        for key, value in module.named_parameters(recurse=False):
+            if not value.requires_grad:
+                continue
+            # Avoid duplicating parameters
+            if value in memo:
+                continue
+            memo.add(value)
+            lr = cfg.SOLVER.BASE_LR
+            weight_decay = cfg.SOLVER.WEIGHT_DECAY
+            if isinstance(module, norm_module_types):
+                weight_decay = cfg.SOLVER.WEIGHT_DECAY_NORM
+            elif key == "bias":
+                # NOTE: unlike Detectron v1, we now default BIAS_LR_FACTOR to 1.0
+                # and WEIGHT_DECAY_BIAS to WEIGHT_DECAY so that bias optimizer
+                # hyperparameters are by default exactly the same as for regular
+                # weights.
+                lr = cfg.SOLVER.BASE_LR * cfg.SOLVER.BIAS_LR_FACTOR
+                weight_decay = cfg.SOLVER.WEIGHT_DECAY_BIAS
+            params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]
+
+    optimizer = torch.optim.SGD(
+        params, cfg.SOLVER.BASE_LR, momentum=cfg.SOLVER.MOMENTUM, nesterov=cfg.SOLVER.NESTEROV
+    )
+    optimizer = maybe_add_gradient_clipping(cfg, optimizer)
+    return optimizer
+
+
+def build_lr_scheduler(
+    cfg: CfgNode, optimizer: torch.optim.Optimizer
+) -> torch.optim.lr_scheduler._LRScheduler:
+    """
+    Build a LR scheduler from config.
+    """
+    name = cfg.SOLVER.LR_SCHEDULER_NAME
+    if name == "WarmupMultiStepLR":
+        return WarmupMultiStepLR(
+            optimizer,
+            cfg.SOLVER.STEPS,
+            cfg.SOLVER.GAMMA,
+            warmup_factor=cfg.SOLVER.WARMUP_FACTOR,
+            warmup_iters=cfg.SOLVER.WARMUP_ITERS,
+            warmup_method=cfg.SOLVER.WARMUP_METHOD,
+        )
+    elif name == "WarmupCosineLR":
+        return WarmupCosineLR(
+            optimizer,
+            cfg.SOLVER.MAX_ITER,
+            warmup_factor=cfg.SOLVER.WARMUP_FACTOR,
+            warmup_iters=cfg.SOLVER.WARMUP_ITERS,
+            warmup_method=cfg.SOLVER.WARMUP_METHOD,
+        )
+    else:
+        raise ValueError("Unknown LR scheduler: {}".format(name))

preprocess/humanparsing/mhp_extension/detectron2/detectron2/solver/lr_scheduler.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import math
+from bisect import bisect_right
+from typing import List
+import torch
+
+# NOTE: PyTorch's LR scheduler interface uses names that assume the LR changes
+# only on epoch boundaries. We typically use iteration based schedules instead.
+# As a result, "epoch" (e.g., as in self.last_epoch) should be understood to mean
+# "iteration" instead.
+
+# FIXME: ideally this would be achieved with a CombinedLRScheduler, separating
+# MultiStepLR with WarmupLR but the current LRScheduler design doesn't allow it.
+
+
+class WarmupMultiStepLR(torch.optim.lr_scheduler._LRScheduler):
+    def __init__(
+        self,
+        optimizer: torch.optim.Optimizer,
+        milestones: List[int],
+        gamma: float = 0.1,
+        warmup_factor: float = 0.001,
+        warmup_iters: int = 1000,
+        warmup_method: str = "linear",
+        last_epoch: int = -1,
+    ):
+        if not list(milestones) == sorted(milestones):
+            raise ValueError(
+                "Milestones should be a list of" " increasing integers. Got {}", milestones
+            )
+        self.milestones = milestones
+        self.gamma = gamma
+        self.warmup_factor = warmup_factor
+        self.warmup_iters = warmup_iters
+        self.warmup_method = warmup_method
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self) -> List[float]:
+        warmup_factor = _get_warmup_factor_at_iter(
+            self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor
+        )
+        return [
+            base_lr * warmup_factor * self.gamma ** bisect_right(self.milestones, self.last_epoch)
+            for base_lr in self.base_lrs
+        ]
+
+    def _compute_values(self) -> List[float]:
+        # The new interface
+        return self.get_lr()
+
+
+class WarmupCosineLR(torch.optim.lr_scheduler._LRScheduler):
+    def __init__(
+        self,
+        optimizer: torch.optim.Optimizer,
+        max_iters: int,
+        warmup_factor: float = 0.001,
+        warmup_iters: int = 1000,
+        warmup_method: str = "linear",
+        last_epoch: int = -1,
+    ):
+        self.max_iters = max_iters
+        self.warmup_factor = warmup_factor
+        self.warmup_iters = warmup_iters
+        self.warmup_method = warmup_method
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self) -> List[float]:
+        warmup_factor = _get_warmup_factor_at_iter(
+            self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor
+        )
+        # Different definitions of half-cosine with warmup are possible. For
+        # simplicity we multiply the standard half-cosine schedule by the warmup
+        # factor. An alternative is to start the period of the cosine at warmup_iters
+        # instead of at 0. In the case that warmup_iters << max_iters the two are
+        # very close to each other.
+        return [
+            base_lr
+            * warmup_factor
+            * 0.5
+            * (1.0 + math.cos(math.pi * self.last_epoch / self.max_iters))
+            for base_lr in self.base_lrs
+        ]
+
+    def _compute_values(self) -> List[float]:
+        # The new interface
+        return self.get_lr()
+
+
+def _get_warmup_factor_at_iter(
+    method: str, iter: int, warmup_iters: int, warmup_factor: float
+) -> float:
+    """
+    Return the learning rate warmup factor at a specific iteration.
+    See :paper:`in1k1h` for more details.
+
+    Args:
+        method (str): warmup method; either "constant" or "linear".
+        iter (int): iteration at which to calculate the warmup factor.
+        warmup_iters (int): the number of warmup iterations.
+        warmup_factor (float): the base warmup factor (the meaning changes according
+            to the method used).
+
+    Returns:
+        float: the effective warmup factor at the given iteration.
+    """
+    if iter >= warmup_iters:
+        return 1.0
+
+    if method == "constant":
+        return warmup_factor
+    elif method == "linear":
+        alpha = iter / warmup_iters
+        return warmup_factor * (1 - alpha) + alpha
+    else:
+        raise ValueError("Unknown warmup method: {}".format(method))

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

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+from .boxes import Boxes, BoxMode, pairwise_iou
+from .image_list import ImageList
+
+from .instances import Instances
+from .keypoints import Keypoints, heatmaps_to_keypoints
+from .masks import BitMasks, PolygonMasks, rasterize_polygons_within_box, polygons_to_bitmask
+from .rotated_boxes import RotatedBoxes
+from .rotated_boxes import pairwise_iou as pairwise_iou_rotated
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

preprocess/humanparsing/mhp_extension/detectron2/detectron2/structures/boxes.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import math
+import numpy as np
+from enum import IntEnum, unique
+from typing import Iterator, List, Tuple, Union
+import torch
+
+_RawBoxType = Union[List[float], Tuple[float, ...], torch.Tensor, np.ndarray]
+
+
+@unique
+class BoxMode(IntEnum):
+    """
+    Enum of different ways to represent a box.
+    """
+
+    XYXY_ABS = 0
+    """
+    (x0, y0, x1, y1) in absolute floating points coordinates.
+    The coordinates in range [0, width or height].
+    """
+    XYWH_ABS = 1
+    """
+    (x0, y0, w, h) in absolute floating points coordinates.
+    """
+    XYXY_REL = 2
+    """
+    Not yet supported!
+    (x0, y0, x1, y1) in range [0, 1]. They are relative to the size of the image.
+    """
+    XYWH_REL = 3
+    """
+    Not yet supported!
+    (x0, y0, w, h) in range [0, 1]. They are relative to the size of the image.
+    """
+    XYWHA_ABS = 4
+    """
+    (xc, yc, w, h, a) in absolute floating points coordinates.
+    (xc, yc) is the center of the rotated box, and the angle a is in degrees ccw.
+    """
+
+    @staticmethod
+    def convert(box: _RawBoxType, from_mode: "BoxMode", to_mode: "BoxMode") -> _RawBoxType:
+        """
+        Args:
+            box: can be a k-tuple, k-list or an Nxk array/tensor, where k = 4 or 5
+            from_mode, to_mode (BoxMode)
+
+        Returns:
+            The converted box of the same type.
+        """
+        if from_mode == to_mode:
+            return box
+
+        original_type = type(box)
+        is_numpy = isinstance(box, np.ndarray)
+        single_box = isinstance(box, (list, tuple))
+        if single_box:
+            assert len(box) == 4 or len(box) == 5, (
+                "BoxMode.convert takes either a k-tuple/list or an Nxk array/tensor,"
+                " where k == 4 or 5"
+            )
+            arr = torch.tensor(box)[None, :]
+        else:
+            # avoid modifying the input box
+            if is_numpy:
+                arr = torch.from_numpy(np.asarray(box)).clone()
+            else:
+                arr = box.clone()
+
+        assert to_mode.value not in [
+            BoxMode.XYXY_REL,
+            BoxMode.XYWH_REL,
+        ] and from_mode.value not in [
+            BoxMode.XYXY_REL,
+            BoxMode.XYWH_REL,
+        ], "Relative mode not yet supported!"
+
+        if from_mode == BoxMode.XYWHA_ABS and to_mode == BoxMode.XYXY_ABS:
+            assert (
+                arr.shape[-1] == 5
+            ), "The last dimension of input shape must be 5 for XYWHA format"
+            original_dtype = arr.dtype
+            arr = arr.double()
+
+            w = arr[:, 2]
+            h = arr[:, 3]
+            a = arr[:, 4]
+            c = torch.abs(torch.cos(a * math.pi / 180.0))
+            s = torch.abs(torch.sin(a * math.pi / 180.0))
+            # This basically computes the horizontal bounding rectangle of the rotated box
+            new_w = c * w + s * h
+            new_h = c * h + s * w
+
+            # convert center to top-left corner
+            arr[:, 0] -= new_w / 2.0
+            arr[:, 1] -= new_h / 2.0
+            # bottom-right corner
+            arr[:, 2] = arr[:, 0] + new_w
+            arr[:, 3] = arr[:, 1] + new_h
+
+            arr = arr[:, :4].to(dtype=original_dtype)
+        elif from_mode == BoxMode.XYWH_ABS and to_mode == BoxMode.XYWHA_ABS:
+            original_dtype = arr.dtype
+            arr = arr.double()
+            arr[:, 0] += arr[:, 2] / 2.0
+            arr[:, 1] += arr[:, 3] / 2.0
+            angles = torch.zeros((arr.shape[0], 1), dtype=arr.dtype)
+            arr = torch.cat((arr, angles), axis=1).to(dtype=original_dtype)
+        else:
+            if to_mode == BoxMode.XYXY_ABS and from_mode == BoxMode.XYWH_ABS:
+                arr[:, 2] += arr[:, 0]
+                arr[:, 3] += arr[:, 1]
+            elif from_mode == BoxMode.XYXY_ABS and to_mode == BoxMode.XYWH_ABS:
+                arr[:, 2] -= arr[:, 0]
+                arr[:, 3] -= arr[:, 1]
+            else:
+                raise NotImplementedError(
+                    "Conversion from BoxMode {} to {} is not supported yet".format(
+                        from_mode, to_mode
+                    )
+                )
+
+        if single_box:
+            return original_type(arr.flatten().tolist())
+        if is_numpy:
+            return arr.numpy()
+        else:
+            return arr
+
+
+class Boxes:
+    """
+    This structure stores a list of boxes as a Nx4 torch.Tensor.
+    It supports some common methods about boxes
+    (`area`, `clip`, `nonempty`, etc),
+    and also behaves like a Tensor
+    (support indexing, `to(device)`, `.device`, and iteration over all boxes)
+
+    Attributes:
+        tensor (torch.Tensor): float matrix of Nx4. Each row is (x1, y1, x2, y2).
+    """
+
+    BoxSizeType = Union[List[int], Tuple[int, int]]
+
+    def __init__(self, tensor: torch.Tensor):
+        """
+        Args:
+            tensor (Tensor[float]): a Nx4 matrix.  Each row is (x1, y1, x2, y2).
+        """
+        device = tensor.device if isinstance(tensor, torch.Tensor) else torch.device("cpu")
+        tensor = torch.as_tensor(tensor, dtype=torch.float32, device=device)
+        if tensor.numel() == 0:
+            # Use reshape, so we don't end up creating a new tensor that does not depend on
+            # the inputs (and consequently confuses jit)
+            tensor = tensor.reshape((0, 4)).to(dtype=torch.float32, device=device)
+        assert tensor.dim() == 2 and tensor.size(-1) == 4, tensor.size()
+
+        self.tensor = tensor
+
+    def clone(self) -> "Boxes":
+        """
+        Clone the Boxes.
+
+        Returns:
+            Boxes
+        """
+        return Boxes(self.tensor.clone())
+
+    def to(self, device: str) -> "Boxes":
+        return Boxes(self.tensor.to(device))
+
+    def area(self) -> torch.Tensor:
+        """
+        Computes the area of all the boxes.
+
+        Returns:
+            torch.Tensor: a vector with areas of each box.
+        """
+        box = self.tensor
+        area = (box[:, 2] - box[:, 0]) * (box[:, 3] - box[:, 1])
+        return area
+
+    def clip(self, box_size: BoxSizeType) -> None:
+        """
+        Clip (in place) the boxes by limiting x coordinates to the range [0, width]
+        and y coordinates to the range [0, height].
+
+        Args:
+            box_size (height, width): The clipping box's size.
+        """
+        assert torch.isfinite(self.tensor).all(), "Box tensor contains infinite or NaN!"
+        h, w = box_size
+        self.tensor[:, 0].clamp_(min=0, max=w)
+        self.tensor[:, 1].clamp_(min=0, max=h)
+        self.tensor[:, 2].clamp_(min=0, max=w)
+        self.tensor[:, 3].clamp_(min=0, max=h)
+
+    def nonempty(self, threshold: float = 0.0) -> torch.Tensor:
+        """
+        Find boxes that are non-empty.
+        A box is considered empty, if either of its side is no larger than threshold.
+
+        Returns:
+            Tensor:
+                a binary vector which represents whether each box is empty
+                (False) or non-empty (True).
+        """
+        box = self.tensor
+        widths = box[:, 2] - box[:, 0]
+        heights = box[:, 3] - box[:, 1]
+        keep = (widths > threshold) & (heights > threshold)
+        return keep
+
+    def __getitem__(self, item: Union[int, slice, torch.BoolTensor]) -> "Boxes":
+        """
+        Returns:
+            Boxes: Create a new :class:`Boxes` by indexing.
+
+        The following usage are allowed:
+
+        1. `new_boxes = boxes[3]`: return a `Boxes` which contains only one box.
+        2. `new_boxes = boxes[2:10]`: return a slice of boxes.
+        3. `new_boxes = boxes[vector]`, where vector is a torch.BoolTensor
+           with `length = len(boxes)`. Nonzero elements in the vector will be selected.
+
+        Note that the returned Boxes might share storage with this Boxes,
+        subject to Pytorch's indexing semantics.
+        """
+        if isinstance(item, int):
+            return Boxes(self.tensor[item].view(1, -1))
+        b = self.tensor[item]
+        assert b.dim() == 2, "Indexing on Boxes with {} failed to return a matrix!".format(item)
+        return Boxes(b)
+
+    def __len__(self) -> int:
+        return self.tensor.shape[0]
+
+    def __repr__(self) -> str:
+        return "Boxes(" + str(self.tensor) + ")"
+
+    def inside_box(self, box_size: BoxSizeType, boundary_threshold: int = 0) -> torch.Tensor:
+        """
+        Args:
+            box_size (height, width): Size of the reference box.
+            boundary_threshold (int): Boxes that extend beyond the reference box
+                boundary by more than boundary_threshold are considered "outside".
+
+        Returns:
+            a binary vector, indicating whether each box is inside the reference box.
+        """
+        height, width = box_size
+        inds_inside = (
+            (self.tensor[..., 0] >= -boundary_threshold)
+            & (self.tensor[..., 1] >= -boundary_threshold)
+            & (self.tensor[..., 2] < width + boundary_threshold)
+            & (self.tensor[..., 3] < height + boundary_threshold)
+        )
+        return inds_inside
+
+    def get_centers(self) -> torch.Tensor:
+        """
+        Returns:
+            The box centers in a Nx2 array of (x, y).
+        """
+        return (self.tensor[:, :2] + self.tensor[:, 2:]) / 2
+
+    def scale(self, scale_x: float, scale_y: float) -> None:
+        """
+        Scale the box with horizontal and vertical scaling factors
+        """
+        self.tensor[:, 0::2] *= scale_x
+        self.tensor[:, 1::2] *= scale_y
+
+    @classmethod
+    def cat(cls, boxes_list: List["Boxes"]) -> "Boxes":
+        """
+        Concatenates a list of Boxes into a single Boxes
+
+        Arguments:
+            boxes_list (list[Boxes])
+
+        Returns:
+            Boxes: the concatenated Boxes
+        """
+        assert isinstance(boxes_list, (list, tuple))
+        if len(boxes_list) == 0:
+            return cls(torch.empty(0))
+        assert all(isinstance(box, Boxes) for box in boxes_list)
+
+        # use torch.cat (v.s. layers.cat) so the returned boxes never share storage with input
+        cat_boxes = cls(torch.cat([b.tensor for b in boxes_list], dim=0))
+        return cat_boxes
+
+    @property
+    def device(self) -> torch.device:
+        return self.tensor.device
+
+    def __iter__(self) -> Iterator[torch.Tensor]:
+        """
+        Yield a box as a Tensor of shape (4,) at a time.
+        """
+        yield from self.tensor
+
+
+# implementation from https://github.com/kuangliu/torchcv/blob/master/torchcv/utils/box.py
+# with slight modifications
+def pairwise_iou(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor:
+    """
+    Given two lists of boxes of size N and M,
+    compute the IoU (intersection over union)
+    between __all__ N x M pairs of boxes.
+    The box order must be (xmin, ymin, xmax, ymax).
+
+    Args:
+        boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively.
+
+    Returns:
+        Tensor: IoU, sized [N,M].
+    """
+    area1 = boxes1.area()
+    area2 = boxes2.area()
+
+    boxes1, boxes2 = boxes1.tensor, boxes2.tensor
+
+    width_height = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) - torch.max(
+        boxes1[:, None, :2], boxes2[:, :2]
+    )  # [N,M,2]
+
+    width_height.clamp_(min=0)  # [N,M,2]
+    inter = width_height.prod(dim=2)  # [N,M]
+    del width_height
+
+    # handle empty boxes
+    iou = torch.where(
+        inter > 0,
+        inter / (area1[:, None] + area2 - inter),
+        torch.zeros(1, dtype=inter.dtype, device=inter.device),
+    )
+    return iou
+
+
+def matched_boxlist_iou(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor:
+    """
+    Compute pairwise intersection over union (IOU) of two sets of matched
+    boxes. The box order must be (xmin, ymin, xmax, ymax).
+    Similar to boxlist_iou, but computes only diagonal elements of the matrix
+    Arguments:
+        boxes1: (Boxes) bounding boxes, sized [N,4].
+        boxes2: (Boxes) bounding boxes, sized [N,4].
+    Returns:
+        (tensor) iou, sized [N].
+    """
+    assert len(boxes1) == len(
+        boxes2
+    ), "boxlists should have the same" "number of entries, got {}, {}".format(
+        len(boxes1), len(boxes2)
+    )
+    area1 = boxes1.area()  # [N]
+    area2 = boxes2.area()  # [N]
+    box1, box2 = boxes1.tensor, boxes2.tensor
+    lt = torch.max(box1[:, :2], box2[:, :2])  # [N,2]
+    rb = torch.min(box1[:, 2:], box2[:, 2:])  # [N,2]
+    wh = (rb - lt).clamp(min=0)  # [N,2]
+    inter = wh[:, 0] * wh[:, 1]  # [N]
+    iou = inter / (area1 + area2 - inter)  # [N]
+    return iou