YOLO-Tutorial-v2/yolo/models/rtdetr/loss_utils.py

import math
import torch
import torch.nn.functional as F
import torch.distributed as dist
from torchvision.ops.boxes import box_area


# ------------------------- For loss -------------------------
## FocalLoss
def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2):
    """
    Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
    Args:
        inputs: A float tensor of arbitrary shape.
                The predictions for each example.
        targets: A float tensor with the same shape as inputs. Stores the binary
                 classification label for each element in inputs
                (0 for the negative class and 1 for the positive class).
        alpha: (optional) Weighting factor in range (0,1) to balance
                positive vs negative examples. Default = -1 (no weighting).
        gamma: Exponent of the modulating factor (1 - p_t) to
               balance easy vs hard examples.
    Returns:
        Loss tensor
    """
    prob = inputs.sigmoid()
    ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
    p_t = prob * targets + (1 - prob) * (1 - targets)
    loss = ce_loss * ((1 - p_t) ** gamma)

    if alpha >= 0:
        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
        loss = alpha_t * loss

    return loss.mean(1).sum() / num_boxes

## Variable FocalLoss
def varifocal_loss_with_logits(pred_logits,
                               gt_score,
                               label,
                               normalizer=1.0,
                               alpha=0.75,
                               gamma=2.0):
    pred_score = F.sigmoid(pred_logits)
    weight = alpha * pred_score.pow(gamma) * (1 - label) + gt_score * label
    loss = F.binary_cross_entropy_with_logits(pred_logits, gt_score, reduction='none')
    loss = loss * weight

    return loss.mean(1).sum() / normalizer

## InverseSigmoid
def inverse_sigmoid(x, eps=1e-5):
    x = x.clamp(min=0, max=1)
    x1 = x.clamp(min=eps)
    x2 = (1 - x).clamp(min=eps)
    return torch.log(x1/x2)

## GIoU loss
class GIoULoss(object):
    """ Modified GIoULoss from Paddle-Paddle"""
    def __init__(self, eps=1e-10, reduction='none'):
        self.eps = eps
        self.reduction = reduction
        assert reduction in ('none', 'mean', 'sum')

    def bbox_overlap(self, box1, box2, eps=1e-10):
        """calculate the iou of box1 and box2
        Args:
            box1 (Tensor): box1 with the shape (..., 4)
            box2 (Tensor): box1 with the shape (..., 4)
            eps (float): epsilon to avoid divide by zero
        Return:
            iou (Tensor): iou of box1 and box2
            overlap (Tensor): overlap of box1 and box2
            union (Tensor): union of box1 and box2
        """
        x1, y1, x2, y2 = box1
        x1g, y1g, x2g, y2g = box2

        xkis1 = torch.max(x1, x1g)
        ykis1 = torch.max(y1, y1g)
        xkis2 = torch.min(x2, x2g)
        ykis2 = torch.min(y2, y2g)
        w_inter = (xkis2 - xkis1).clip(0)
        h_inter = (ykis2 - ykis1).clip(0)
        overlap = w_inter * h_inter

        area1 = (x2 - x1) * (y2 - y1)
        area2 = (x2g - x1g) * (y2g - y1g)
        union = area1 + area2 - overlap + eps
        iou = overlap / union

        return iou, overlap, union

    def __call__(self, pbox, gbox):
        # x1, y1, x2, y2 = torch.split(pbox, 4, dim=-1)
        # x1g, y1g, x2g, y2g = torch.split(gbox, 4, dim=-1)
        x1, y1, x2, y2 = torch.chunk(pbox, 4, dim=-1)
        x1g, y1g, x2g, y2g = torch.chunk(gbox, 4, dim=-1)
        box1 = [x1, y1, x2, y2]
        box2 = [x1g, y1g, x2g, y2g]
        iou, _, union = self.bbox_overlap(box1, box2, self.eps)
        xc1 = torch.min(x1, x1g)
        yc1 = torch.min(y1, y1g)
        xc2 = torch.max(x2, x2g)
        yc2 = torch.max(y2, y2g)

        area_c = (xc2 - xc1) * (yc2 - yc1) + self.eps
        miou = iou - ((area_c - union) / area_c)
        giou = 1 - miou

        if self.reduction == 'none':
            loss = giou
        elif self.reduction == 'sum':
            loss = giou.sum()
        elif self.reduction == 'mean':
            loss = giou.mean()

        return loss


# ------------------------- For box -------------------------
def box_cxcywh_to_xyxy(x):
    x_c, y_c, w, h = x.unbind(-1)
    b = [(x_c - 0.5 * w), (y_c - 0.5 * h),
         (x_c + 0.5 * w), (y_c + 0.5 * h)]
    return torch.stack(b, dim=-1)

def box_xyxy_to_cxcywh(x):
    x0, y0, x1, y1 = x.unbind(-1)
    b = [(x0 + x1) / 2, (y0 + y1) / 2,
         (x1 - x0), (y1 - y0)]
    return torch.stack(b, dim=-1)

def box_iou(boxes1, boxes2):
    area1 = box_area(boxes1)
    area2 = box_area(boxes2)

    lt = torch.max(boxes1[:, None, :2], boxes2[:, :2])  # [N,M,2]
    rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])  # [N,M,2]

    wh = (rb - lt).clamp(min=0)  # [N,M,2]
    inter = wh[:, :, 0] * wh[:, :, 1]  # [N,M]

    union = area1[:, None] + area2 - inter

    iou = inter / union
    return iou, union

def generalized_box_iou(boxes1, boxes2):
    """
    Generalized IoU from https://giou.stanford.edu/

    The boxes should be in [x0, y0, x1, y1] format

    Returns a [N, M] pairwise matrix, where N = len(boxes1)
    and M = len(boxes2)
    """
    # degenerate boxes gives inf / nan results
    # so do an early check
    assert (boxes1[:, 2:] >= boxes1[:, :2]).all()
    assert (boxes2[:, 2:] >= boxes2[:, :2]).all()
    iou, union = box_iou(boxes1, boxes2)

    lt = torch.min(boxes1[:, None, :2], boxes2[:, :2])
    rb = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])

    wh = (rb - lt).clamp(min=0)  # [N,M,2]
    area = wh[:, :, 0] * wh[:, :, 1]

    return iou - (area - union) / area

def bbox_iou(box1, box2, giou=False, diou=False, ciou=False, eps=1e-9):
    """Modified from Paddle-paddle
    Args:
        box1 (list): [x, y, w, h], all have the shape [b, na, h, w, 1]
        box2 (list): [x, y, w, h], all have the shape [b, na, h, w, 1]
        giou (bool): whether use giou or not, default False
        diou (bool): whether use diou or not, default False
        ciou (bool): whether use ciou or not, default False
        eps (float): epsilon to avoid divide by zero
    Return:
        iou (Tensor): iou of box1 and box1, with the shape [b, na, h, w, 1]
    """
    px1, py1, px2, py2 = torch.chunk(box1, 4, -1)
    gx1, gy1, gx2, gy2 = torch.chunk(box2, 4, -1)
    x1 = torch.max(px1, gx1)
    y1 = torch.max(py1, gy1)
    x2 = torch.min(px2, gx2)
    y2 = torch.min(py2, gy2)

    overlap = ((x2 - x1).clamp(0)) * ((y2 - y1).clamp(0))

    area1 = (px2 - px1) * (py2 - py1)
    area1 = area1.clamp(0)

    area2 = (gx2 - gx1) * (gy2 - gy1)
    area2 = area2.clamp(0)

    union = area1 + area2 - overlap + eps
    iou = overlap / union

    if giou or ciou or diou:
        # convex w, h
        cw = torch.max(px2, gx2) - torch.min(px1, gx1)
        ch = torch.max(py2, gy2) - torch.min(py1, gy1)
        if giou:
            c_area = cw * ch + eps
            return iou - (c_area - union) / c_area
        else:
            # convex diagonal squared
            c2 = cw**2 + ch**2 + eps
            # center distance
            rho2 = ((px1 + px2 - gx1 - gx2)**2 + (py1 + py2 - gy1 - gy2)**2) / 4
            if diou:
                return iou - rho2 / c2
            else:
                w1, h1 = px2 - px1, py2 - py1 + eps
                w2, h2 = gx2 - gx1, gy2 - gy1 + eps
                delta = torch.atan(w1 / h1) - torch.atan(w2 / h2)
                v = (4 / math.pi**2) * torch.pow(delta, 2)
                alpha = v / (1 + eps - iou + v)
                alpha.requires_grad_ = False
                return iou - (rho2 / c2 + v * alpha)
    else:
        return iou


# ------------------------- For distributed -------------------------
def is_dist_avail_and_initialized():
    if not dist.is_available():
        return False
    if not dist.is_initialized():
        return False
    return True

def get_world_size():
    if not is_dist_avail_and_initialized():
        return 1
    return dist.get_world_size()