ultralytics-opencv-headless 8.3.253__py3-none-any.whl → 8.4.0__py3-none-any.whl

ultralytics/utils/loss.py

@@ -2,19 +2,20 @@
 
 from __future__ import annotations
 
+import math
 from typing import Any
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
-from ultralytics.utils.metrics import OKS_SIGMA
+from ultralytics.utils.metrics import OKS_SIGMA, RLE_WEIGHT
 from ultralytics.utils.ops import crop_mask, xywh2xyxy, xyxy2xywh
 from ultralytics.utils.tal import RotatedTaskAlignedAssigner, TaskAlignedAssigner, dist2bbox, dist2rbox, make_anchors
 from ultralytics.utils.torch_utils import autocast
 
 from .metrics import bbox_iou, probiou
-from .tal import bbox2dist
+from .tal import bbox2dist, rbox2dist
 
 
 class VarifocalLoss(nn.Module):
@@ -122,6 +123,8 @@ class BboxLoss(nn.Module):
         target_scores: torch.Tensor,
         target_scores_sum: torch.Tensor,
         fg_mask: torch.Tensor,
+        imgsz: torch.Tensor,
+        stride: torch.Tensor,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """Compute IoU and DFL losses for bounding boxes."""
         weight = target_scores.sum(-1)[fg_mask].unsqueeze(-1)
@@ -134,11 +137,75 @@ class BboxLoss(nn.Module):
             loss_dfl = self.dfl_loss(pred_dist[fg_mask].view(-1, self.dfl_loss.reg_max), target_ltrb[fg_mask]) * weight
             loss_dfl = loss_dfl.sum() / target_scores_sum
         else:
-            loss_dfl = torch.tensor(0.0).to(pred_dist.device)
+            target_ltrb = bbox2dist(anchor_points, target_bboxes)
+            # normalize ltrb by image size
+            target_ltrb = target_ltrb * stride
+            target_ltrb[..., 0::2] /= imgsz[1]
+            target_ltrb[..., 1::2] /= imgsz[0]
+            pred_dist = pred_dist * stride
+            pred_dist[..., 0::2] /= imgsz[1]
+            pred_dist[..., 1::2] /= imgsz[0]
+            loss_dfl = (
+                F.l1_loss(pred_dist[fg_mask], target_ltrb[fg_mask], reduction="none").mean(-1, keepdim=True) * weight
+            )
+            loss_dfl = loss_dfl.sum() / target_scores_sum
 
         return loss_iou, loss_dfl
 
 
+class RLELoss(nn.Module):
+    """Residual Log-Likelihood Estimation Loss.
+
+    Args:
+        use_target_weight (bool): Option to use weighted loss.
+        size_average (bool): Option to average the loss by the batch_size.
+        residual (bool): Option to add L1 loss and let the flow learn the residual error distribution.
+
+    References:
+        https://arxiv.org/abs/2107.11291
+    """
+
+    def __init__(self, use_target_weight: bool = True, size_average: bool = True, residual: bool = True):
+        """Initialize RLELoss with target weight and residual options.
+
+        Args:
+            use_target_weight (bool): Whether to use target weights for loss calculation.
+            size_average (bool): Whether to average the loss over elements.
+            residual (bool): Whether to include residual log-likelihood term.
+        """
+        super().__init__()
+        self.size_average = size_average
+        self.use_target_weight = use_target_weight
+        self.residual = residual
+
+    def forward(
+        self, sigma: torch.Tensor, log_phi: torch.Tensor, error: torch.Tensor, target_weight: torch.Tensor = None
+    ) -> torch.Tensor:
+        """
+        Args:
+            sigma (torch.Tensor): Output sigma, shape (N, D).
+            log_phi (torch.Tensor): Output log_phi, shape (N).
+            error (torch.Tensor): Error, shape (N, D).
+            target_weight (torch.Tensor): Weights across different joint types, shape (N).
+        """
+        log_sigma = torch.log(sigma)
+        loss = log_sigma - log_phi.unsqueeze(1)
+
+        if self.residual:
+            loss += torch.log(sigma * 2) + torch.abs(error)
+
+        if self.use_target_weight:
+            assert target_weight is not None, "'target_weight' should not be None when 'use_target_weight' is True."
+            if target_weight.dim() == 1:
+                target_weight = target_weight.unsqueeze(1)
+            loss *= target_weight
+
+        if self.size_average:
+            loss /= len(loss)
+
+        return loss.sum()
+
+
 class RotatedBboxLoss(BboxLoss):
     """Criterion class for computing training losses for rotated bounding boxes."""
 
@@ -155,6 +222,8 @@ class RotatedBboxLoss(BboxLoss):
         target_scores: torch.Tensor,
         target_scores_sum: torch.Tensor,
         fg_mask: torch.Tensor,
+        imgsz: torch.Tensor,
+        stride: torch.Tensor,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """Compute IoU and DFL losses for rotated bounding boxes."""
         weight = target_scores.sum(-1)[fg_mask].unsqueeze(-1)
@@ -163,15 +232,84 @@ class RotatedBboxLoss(BboxLoss):
 
         # DFL loss
         if self.dfl_loss:
-            target_ltrb = bbox2dist(anchor_points, xywh2xyxy(target_bboxes[..., :4]), self.dfl_loss.reg_max - 1)
+            target_ltrb = rbox2dist(
+                target_bboxes[..., :4], anchor_points, target_bboxes[..., 4:5], reg_max=self.dfl_loss.reg_max - 1
+            )
             loss_dfl = self.dfl_loss(pred_dist[fg_mask].view(-1, self.dfl_loss.reg_max), target_ltrb[fg_mask]) * weight
             loss_dfl = loss_dfl.sum() / target_scores_sum
         else:
-            loss_dfl = torch.tensor(0.0).to(pred_dist.device)
+            target_ltrb = rbox2dist(target_bboxes[..., :4], anchor_points, target_bboxes[..., 4:5])
+            target_ltrb = target_ltrb * stride
+            target_ltrb[..., 0::2] /= imgsz[1]
+            target_ltrb[..., 1::2] /= imgsz[0]
+            pred_dist = pred_dist * stride
+            pred_dist[..., 0::2] /= imgsz[1]
+            pred_dist[..., 1::2] /= imgsz[0]
+            loss_dfl = (
+                F.l1_loss(pred_dist[fg_mask], target_ltrb[fg_mask], reduction="none").mean(-1, keepdim=True) * weight
+            )
+            loss_dfl = loss_dfl.sum() / target_scores_sum
 
         return loss_iou, loss_dfl
 
 
+class MultiChannelDiceLoss(nn.Module):
+    """Criterion class for computing multi-channel Dice losses."""
+
+    def __init__(self, smooth: float = 1e-6, reduction: str = "mean"):
+        """Initialize MultiChannelDiceLoss with smoothing and reduction options.
+
+        Args:
+            smooth (float): Smoothing factor to avoid division by zero.
+            reduction (str): Reduction method ('mean', 'sum', or 'none').
+        """
+        super().__init__()
+        self.smooth = smooth
+        self.reduction = reduction
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """Calculate multi-channel Dice loss between predictions and targets."""
+        assert pred.size() == target.size(), "the size of predict and target must be equal."
+
+        pred = pred.sigmoid()
+        intersection = (pred * target).sum(dim=(2, 3))
+        union = pred.sum(dim=(2, 3)) + target.sum(dim=(2, 3))
+        dice = (2.0 * intersection + self.smooth) / (union + self.smooth)
+        dice_loss = 1.0 - dice
+        dice_loss = dice_loss.mean(dim=1)
+
+        if self.reduction == "mean":
+            return dice_loss.mean()
+        elif self.reduction == "sum":
+            return dice_loss.sum()
+        else:
+            return dice_loss
+
+
+class BCEDiceLoss(nn.Module):
+    """Criterion class for computing combined BCE and Dice losses."""
+
+    def __init__(self, weight_bce: float = 0.5, weight_dice: float = 0.5):
+        """Initialize BCEDiceLoss with BCE and Dice weight factors.
+
+        Args:
+            weight_bce (float): Weight factor for BCE loss component.
+            weight_dice (float): Weight factor for Dice loss component.
+        """
+        super().__init__()
+        self.weight_bce = weight_bce
+        self.weight_dice = weight_dice
+        self.bce = nn.BCEWithLogitsLoss()
+        self.dice = MultiChannelDiceLoss(smooth=1)
+
+    def forward(self, pred: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """Calculate combined BCE and Dice loss between predictions and targets."""
+        _, _, mask_h, mask_w = pred.shape
+        if tuple(target.shape[-2:]) != (mask_h, mask_w):  # downsample to the same size as pred
+            target = F.interpolate(target, (mask_h, mask_w), mode="nearest")
+        return self.weight_bce * self.bce(pred, target) + self.weight_dice * self.dice(pred, target)
+
+
 class KeypointLoss(nn.Module):
     """Criterion class for computing keypoint losses."""
 
@@ -194,7 +332,7 @@ class KeypointLoss(nn.Module):
 class v8DetectionLoss:
     """Criterion class for computing training losses for YOLOv8 object detection."""
 
-    def __init__(self, model, tal_topk: int = 10):  # model must be de-paralleled
+    def __init__(self, model, tal_topk: int = 10, tal_topk2: int | None = None):  # model must be de-paralleled
         """Initialize v8DetectionLoss with model parameters and task-aligned assignment settings."""
         device = next(model.parameters()).device  # get model device
         h = model.args  # hyperparameters
@@ -210,7 +348,14 @@ class v8DetectionLoss:
 
         self.use_dfl = m.reg_max > 1
 
-        self.assigner = TaskAlignedAssigner(topk=tal_topk, num_classes=self.nc, alpha=0.5, beta=6.0)
+        self.assigner = TaskAlignedAssigner(
+            topk=tal_topk,
+            num_classes=self.nc,
+            alpha=0.5,
+            beta=6.0,
+            stride=self.stride.tolist(),
+            topk2=tal_topk2,
+        )
         self.bbox_loss = BboxLoss(m.reg_max).to(device)
         self.proj = torch.arange(m.reg_max, dtype=torch.float, device=device)
 
@@ -240,35 +385,31 @@ class v8DetectionLoss:
             # pred_dist = (pred_dist.view(b, a, c // 4, 4).softmax(2) * self.proj.type(pred_dist.dtype).view(1, 1, -1, 1)).sum(2)
         return dist2bbox(pred_dist, anchor_points, xywh=False)
 
-    def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
-        """Calculate the sum of the loss for box, cls and dfl multiplied by batch size."""
+    def get_assigned_targets_and_loss(self, preds: dict[str, torch.Tensor], batch: dict[str, Any]) -> tuple:
+        """Calculate the sum of the loss for box, cls and dfl multiplied by batch size and return foreground mask and
+        target indices.
+        """
         loss = torch.zeros(3, device=self.device)  # box, cls, dfl
-        feats = preds[1] if isinstance(preds, tuple) else preds
-        pred_distri, pred_scores = torch.cat([xi.view(feats[0].shape[0], self.no, -1) for xi in feats], 2).split(
-            (self.reg_max * 4, self.nc), 1
+        pred_distri, pred_scores = (
+            preds["boxes"].permute(0, 2, 1).contiguous(),
+            preds["scores"].permute(0, 2, 1).contiguous(),
         )
-
-        pred_scores = pred_scores.permute(0, 2, 1).contiguous()
-        pred_distri = pred_distri.permute(0, 2, 1).contiguous()
+        anchor_points, stride_tensor = make_anchors(preds["feats"], self.stride, 0.5)
 
         dtype = pred_scores.dtype
         batch_size = pred_scores.shape[0]
-        imgsz = torch.tensor(feats[0].shape[2:], device=self.device, dtype=dtype) * self.stride[0]  # image size (h,w)
-        anchor_points, stride_tensor = make_anchors(feats, self.stride, 0.5)
+        imgsz = torch.tensor(preds["feats"][0].shape[2:], device=self.device, dtype=dtype) * self.stride[0]
 
         # Targets
         targets = torch.cat((batch["batch_idx"].view(-1, 1), batch["cls"].view(-1, 1), batch["bboxes"]), 1)
-        targets = self.preprocess(targets, batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
+        targets = self.preprocess(targets.to(self.device), batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
         gt_labels, gt_bboxes = targets.split((1, 4), 2)  # cls, xyxy
         mask_gt = gt_bboxes.sum(2, keepdim=True).gt_(0.0)
 
         # Pboxes
         pred_bboxes = self.bbox_decode(anchor_points, pred_distri)  # xyxy, (b, h*w, 4)
-        # dfl_conf = pred_distri.view(batch_size, -1, 4, self.reg_max).detach().softmax(-1)
-        # dfl_conf = (dfl_conf.amax(-1).mean(-1) + dfl_conf.amax(-1).amin(-1)) / 2
 
-        _, target_bboxes, target_scores, fg_mask, _ = self.assigner(
-            # pred_scores.detach().sigmoid() * 0.8 + dfl_conf.unsqueeze(-1) * 0.2,
+        _, target_bboxes, target_scores, fg_mask, target_gt_idx = self.assigner(
             pred_scores.detach().sigmoid(),
             (pred_bboxes.detach() * stride_tensor).type(gt_bboxes.dtype),
             anchor_points * stride_tensor,
@@ -280,7 +421,6 @@ class v8DetectionLoss:
         target_scores_sum = max(target_scores.sum(), 1)
 
         # Cls loss
-        # loss[1] = self.varifocal_loss(pred_scores, target_scores, target_labels) / target_scores_sum  # VFL way
         loss[1] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum  # BCE
 
         # Bbox loss
@@ -293,105 +433,97 @@ class v8DetectionLoss:
                 target_scores,
                 target_scores_sum,
                 fg_mask,
+                imgsz,
+                stride_tensor,
             )
 
         loss[0] *= self.hyp.box  # box gain
         loss[1] *= self.hyp.cls  # cls gain
         loss[2] *= self.hyp.dfl  # dfl gain
+        return (
+            (fg_mask, target_gt_idx, target_bboxes, anchor_points, stride_tensor),
+            loss,
+            loss.detach(),
+        )  # loss(box, cls, dfl)
 
-        return loss * batch_size, loss.detach()  # loss(box, cls, dfl)
+    def parse_output(
+        self, preds: dict[str, torch.Tensor] | tuple[torch.Tensor, dict[str, torch.Tensor]]
+    ) -> torch.Tensor:
+        """Parse model predictions to extract features."""
+        return preds[1] if isinstance(preds, tuple) else preds
+
+    def __call__(
+        self,
+        preds: dict[str, torch.Tensor] | tuple[torch.Tensor, dict[str, torch.Tensor]],
+        batch: dict[str, torch.Tensor],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the sum of the loss for box, cls and dfl multiplied by batch size."""
+        return self.loss(self.parse_output(preds), batch)
+
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+        """A wrapper for get_assigned_targets_and_loss and parse_output."""
+        batch_size = preds["boxes"].shape[0]
+        loss, loss_detach = self.get_assigned_targets_and_loss(preds, batch)[1:]
+        return loss * batch_size, loss_detach
 
 
 class v8SegmentationLoss(v8DetectionLoss):
     """Criterion class for computing training losses for YOLOv8 segmentation."""
 
-    def __init__(self, model):  # model must be de-paralleled
+    def __init__(self, model, tal_topk: int = 10, tal_topk2: int | None = None):  # model must be de-paralleled
         """Initialize the v8SegmentationLoss class with model parameters and mask overlap setting."""
-        super().__init__(model)
+        super().__init__(model, tal_topk, tal_topk2)
         self.overlap = model.args.overlap_mask
+        self.bcedice_loss = BCEDiceLoss(weight_bce=0.5, weight_dice=0.5)
 
-    def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         """Calculate and return the combined loss for detection and segmentation."""
-        loss = torch.zeros(4, device=self.device)  # box, seg, cls, dfl
-        feats, pred_masks, proto = preds if len(preds) == 3 else preds[1]
-        batch_size, _, mask_h, mask_w = proto.shape  # batch size, number of masks, mask height, mask width
-        pred_distri, pred_scores = torch.cat([xi.view(feats[0].shape[0], self.no, -1) for xi in feats], 2).split(
-            (self.reg_max * 4, self.nc), 1
-        )
-
-        # B, grids, ..
-        pred_scores = pred_scores.permute(0, 2, 1).contiguous()
-        pred_distri = pred_distri.permute(0, 2, 1).contiguous()
-        pred_masks = pred_masks.permute(0, 2, 1).contiguous()
-
-        dtype = pred_scores.dtype
-        imgsz = torch.tensor(feats[0].shape[2:], device=self.device, dtype=dtype) * self.stride[0]  # image size (h,w)
-        anchor_points, stride_tensor = make_anchors(feats, self.stride, 0.5)
-
-        # Targets
-        try:
-            batch_idx = batch["batch_idx"].view(-1, 1)
-            targets = torch.cat((batch_idx, batch["cls"].view(-1, 1), batch["bboxes"]), 1)
-            targets = self.preprocess(targets, batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
-            gt_labels, gt_bboxes = targets.split((1, 4), 2)  # cls, xyxy
-            mask_gt = gt_bboxes.sum(2, keepdim=True).gt_(0.0)
-        except RuntimeError as e:
-            raise TypeError(
-                "ERROR ❌ segment dataset incorrectly formatted or not a segment dataset.\n"
-                "This error can occur when incorrectly training a 'segment' model on a 'detect' dataset, "
-                "i.e. 'yolo train model=yolo11n-seg.pt data=coco8.yaml'.\nVerify your dataset is a "
-                "correctly formatted 'segment' dataset using 'data=coco8-seg.yaml' "
-                "as an example.\nSee https://docs.ultralytics.com/datasets/segment/ for help."
-            ) from e
-
-        # Pboxes
-        pred_bboxes = self.bbox_decode(anchor_points, pred_distri)  # xyxy, (b, h*w, 4)
-
-        _, target_bboxes, target_scores, fg_mask, target_gt_idx = self.assigner(
-            pred_scores.detach().sigmoid(),
-            (pred_bboxes.detach() * stride_tensor).type(gt_bboxes.dtype),
-            anchor_points * stride_tensor,
-            gt_labels,
-            gt_bboxes,
-            mask_gt,
-        )
-
-        target_scores_sum = max(target_scores.sum(), 1)
-
-        # Cls loss
-        # loss[1] = self.varifocal_loss(pred_scores, target_scores, target_labels) / target_scores_sum  # VFL way
-        loss[2] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum  # BCE
+        pred_masks, proto = preds["mask_coefficient"].permute(0, 2, 1).contiguous(), preds["proto"]
+        loss = torch.zeros(5, device=self.device)  # box, seg, cls, dfl
+        if len(proto) == 2:
+            proto, pred_semseg = proto
+        else:
+            pred_semseg = None
+        (fg_mask, target_gt_idx, target_bboxes, _, _), det_loss, _ = self.get_assigned_targets_and_loss(preds, batch)
+        # NOTE: re-assign index for consistency for now. Need to be removed in the future.
+        loss[0], loss[2], loss[3] = det_loss[0], det_loss[1], det_loss[2]
 
+        batch_size, _, mask_h, mask_w = proto.shape  # batch size, number of masks, mask height, mask width
         if fg_mask.sum():
-            # Bbox loss
-            loss[0], loss[3] = self.bbox_loss(
-                pred_distri,
-                pred_bboxes,
-                anchor_points,
-                target_bboxes / stride_tensor,
-                target_scores,
-                target_scores_sum,
-                fg_mask,
-            )
             # Masks loss
             masks = batch["masks"].to(self.device).float()
             if tuple(masks.shape[-2:]) != (mask_h, mask_w):  # downsample
-                masks = F.interpolate(masks[None], (mask_h, mask_w), mode="nearest")[0]
+                # masks = F.interpolate(masks[None], (mask_h, mask_w), mode="nearest")[0]
+                proto = F.interpolate(proto, masks.shape[-2:], mode="bilinear", align_corners=False)
 
+            imgsz = (
+                torch.tensor(preds["feats"][0].shape[2:], device=self.device, dtype=pred_masks.dtype) * self.stride[0]
+            )
             loss[1] = self.calculate_segmentation_loss(
-                fg_mask, masks, target_gt_idx, target_bboxes, batch_idx, proto, pred_masks, imgsz, self.overlap
+                fg_mask,
+                masks,
+                target_gt_idx,
+                target_bboxes,
+                batch["batch_idx"].view(-1, 1),
+                proto,
+                pred_masks,
+                imgsz,
             )
+            if pred_semseg is not None:
+                sem_masks = batch["sem_masks"].to(self.device)  # NxHxW
+                mask_zero = sem_masks == 0  # NxHxW
+                sem_masks = F.one_hot(sem_masks.long(), num_classes=self.nc).permute(0, 3, 1, 2).float()  # NxCxHxW
+                sem_masks[mask_zero.unsqueeze(1).expand_as(sem_masks)] = 0
+                loss[4] = self.bcedice_loss(pred_semseg, sem_masks)
+                loss[4] *= self.hyp.box  # seg gain
 
         # WARNING: lines below prevent Multi-GPU DDP 'unused gradient' PyTorch errors, do not remove
         else:
             loss[1] += (proto * 0).sum() + (pred_masks * 0).sum()  # inf sums may lead to nan loss
+            loss[4] += (pred_semseg * 0).sum() + (sem_masks * 0).sum()
 
-        loss[0] *= self.hyp.box  # box gain
         loss[1] *= self.hyp.box  # seg gain
-        loss[2] *= self.hyp.cls  # cls gain
-        loss[3] *= self.hyp.dfl  # dfl gain
-
-        return loss * batch_size, loss.detach()  # loss(box, seg, cls, dfl)
+        return loss * batch_size, loss.detach()  # loss(box, cls, dfl)
 
     @staticmethod
     def single_mask_loss(
@@ -427,7 +559,6 @@ class v8SegmentationLoss(v8DetectionLoss):
         proto: torch.Tensor,
         pred_masks: torch.Tensor,
         imgsz: torch.Tensor,
-        overlap: bool,
     ) -> torch.Tensor:
         """Calculate the loss for instance segmentation.
 
@@ -440,7 +571,6 @@ class v8SegmentationLoss(v8DetectionLoss):
             proto (torch.Tensor): Prototype masks of shape (BS, 32, H, W).
             pred_masks (torch.Tensor): Predicted masks for each anchor of shape (BS, N_anchors, 32).
             imgsz (torch.Tensor): Size of the input image as a tensor of shape (2), i.e., (H, W).
-            overlap (bool): Whether the masks in `masks` tensor overlap.
 
         Returns:
             (torch.Tensor): The calculated loss for instance segmentation.
@@ -466,7 +596,7 @@ class v8SegmentationLoss(v8DetectionLoss):
             fg_mask_i, target_gt_idx_i, pred_masks_i, proto_i, mxyxy_i, marea_i, masks_i = single_i
             if fg_mask_i.any():
                 mask_idx = target_gt_idx_i[fg_mask_i]
-                if overlap:
+                if self.overlap:
                     gt_mask = masks_i == (mask_idx + 1).view(-1, 1, 1)
                     gt_mask = gt_mask.float()
                 else:
@@ -486,9 +616,9 @@ class v8SegmentationLoss(v8DetectionLoss):
 class v8PoseLoss(v8DetectionLoss):
     """Criterion class for computing training losses for YOLOv8 pose estimation."""
 
-    def __init__(self, model):  # model must be de-paralleled
+    def __init__(self, model, tal_topk: int = 10, tal_topk2: int = 10):  # model must be de-paralleled
         """Initialize v8PoseLoss with model parameters and keypoint-specific loss functions."""
-        super().__init__(model)
+        super().__init__(model, tal_topk, tal_topk2)
         self.kpt_shape = model.model[-1].kpt_shape
         self.bce_pose = nn.BCEWithLogitsLoss()
         is_pose = self.kpt_shape == [17, 3]
@@ -496,69 +626,40 @@ class v8PoseLoss(v8DetectionLoss):
         sigmas = torch.from_numpy(OKS_SIGMA).to(self.device) if is_pose else torch.ones(nkpt, device=self.device) / nkpt
         self.keypoint_loss = KeypointLoss(sigmas=sigmas)
 
-    def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         """Calculate the total loss and detach it for pose estimation."""
-        loss = torch.zeros(5, device=self.device)  # box, pose, kobj, cls, dfl
-        feats, pred_kpts = preds if isinstance(preds[0], list) else preds[1]
-        pred_distri, pred_scores = torch.cat([xi.view(feats[0].shape[0], self.no, -1) for xi in feats], 2).split(
-            (self.reg_max * 4, self.nc), 1
+        pred_kpts = preds["kpts"].permute(0, 2, 1).contiguous()
+        loss = torch.zeros(5, device=self.device)  # box, cls, dfl, kpt_location, kpt_visibility
+        (fg_mask, target_gt_idx, target_bboxes, anchor_points, stride_tensor), det_loss, _ = (
+            self.get_assigned_targets_and_loss(preds, batch)
         )
+        # NOTE: re-assign index for consistency for now. Need to be removed in the future.
+        loss[0], loss[3], loss[4] = det_loss[0], det_loss[1], det_loss[2]
 
-        # B, grids, ..
-        pred_scores = pred_scores.permute(0, 2, 1).contiguous()
-        pred_distri = pred_distri.permute(0, 2, 1).contiguous()
-        pred_kpts = pred_kpts.permute(0, 2, 1).contiguous()
-
-        dtype = pred_scores.dtype
-        imgsz = torch.tensor(feats[0].shape[2:], device=self.device, dtype=dtype) * self.stride[0]  # image size (h,w)
-        anchor_points, stride_tensor = make_anchors(feats, self.stride, 0.5)
-
-        # Targets
-        batch_size = pred_scores.shape[0]
-        batch_idx = batch["batch_idx"].view(-1, 1)
-        targets = torch.cat((batch_idx, batch["cls"].view(-1, 1), batch["bboxes"]), 1)
-        targets = self.preprocess(targets, batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
-        gt_labels, gt_bboxes = targets.split((1, 4), 2)  # cls, xyxy
-        mask_gt = gt_bboxes.sum(2, keepdim=True).gt_(0.0)
+        batch_size = pred_kpts.shape[0]
+        imgsz = torch.tensor(preds["feats"][0].shape[2:], device=self.device, dtype=pred_kpts.dtype) * self.stride[0]
 
         # Pboxes
-        pred_bboxes = self.bbox_decode(anchor_points, pred_distri)  # xyxy, (b, h*w, 4)
         pred_kpts = self.kpts_decode(anchor_points, pred_kpts.view(batch_size, -1, *self.kpt_shape))  # (b, h*w, 17, 3)
 
-        _, target_bboxes, target_scores, fg_mask, target_gt_idx = self.assigner(
-            pred_scores.detach().sigmoid(),
-            (pred_bboxes.detach() * stride_tensor).type(gt_bboxes.dtype),
-            anchor_points * stride_tensor,
-            gt_labels,
-            gt_bboxes,
-            mask_gt,
-        )
-
-        target_scores_sum = max(target_scores.sum(), 1)
-
-        # Cls loss
-        # loss[1] = self.varifocal_loss(pred_scores, target_scores, target_labels) / target_scores_sum  # VFL way
-        loss[3] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum  # BCE
-
         # Bbox loss
         if fg_mask.sum():
-            target_bboxes /= stride_tensor
-            loss[0], loss[4] = self.bbox_loss(
-                pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask
-            )
             keypoints = batch["keypoints"].to(self.device).float().clone()
             keypoints[..., 0] *= imgsz[1]
             keypoints[..., 1] *= imgsz[0]
 
             loss[1], loss[2] = self.calculate_keypoints_loss(
-                fg_mask, target_gt_idx, keypoints, batch_idx, stride_tensor, target_bboxes, pred_kpts
+                fg_mask,
+                target_gt_idx,
+                keypoints,
+                batch["batch_idx"].view(-1, 1),
+                stride_tensor,
+                target_bboxes,
+                pred_kpts,
             )
 
-        loss[0] *= self.hyp.box  # box gain
         loss[1] *= self.hyp.pose  # pose gain
         loss[2] *= self.hyp.kobj  # kobj gain
-        loss[3] *= self.hyp.cls  # cls gain
-        loss[4] *= self.hyp.dfl  # dfl gain
 
         return loss * batch_size, loss.detach()  # loss(box, pose, kobj, cls, dfl)
 
@@ -571,34 +672,23 @@ class v8PoseLoss(v8DetectionLoss):
         y[..., 1] += anchor_points[:, [1]] - 0.5
         return y
 
-    def calculate_keypoints_loss(
+    def _select_target_keypoints(
         self,
-        masks: torch.Tensor,
-        target_gt_idx: torch.Tensor,
         keypoints: torch.Tensor,
         batch_idx: torch.Tensor,
-        stride_tensor: torch.Tensor,
-        target_bboxes: torch.Tensor,
-        pred_kpts: torch.Tensor,
-    ) -> tuple[torch.Tensor, torch.Tensor]:
-        """Calculate the keypoints loss for the model.
-
-        This function calculates the keypoints loss and keypoints object loss for a given batch. The keypoints loss is
-        based on the difference between the predicted keypoints and ground truth keypoints. The keypoints object loss is
-        a binary classification loss that classifies whether a keypoint is present or not.
+        target_gt_idx: torch.Tensor,
+        masks: torch.Tensor,
+    ) -> torch.Tensor:
+        """Select target keypoints for each anchor based on batch index and target ground truth index.
 
         Args:
-            masks (torch.Tensor): Binary mask tensor indicating object presence, shape (BS, N_anchors).
-            target_gt_idx (torch.Tensor): Index tensor mapping anchors to ground truth objects, shape (BS, N_anchors).
             keypoints (torch.Tensor): Ground truth keypoints, shape (N_kpts_in_batch, N_kpts_per_object, kpts_dim).
             batch_idx (torch.Tensor): Batch index tensor for keypoints, shape (N_kpts_in_batch, 1).
-            stride_tensor (torch.Tensor): Stride tensor for anchors, shape (N_anchors, 1).
-            target_bboxes (torch.Tensor): Ground truth boxes in (x1, y1, x2, y2) format, shape (BS, N_anchors, 4).
-            pred_kpts (torch.Tensor): Predicted keypoints, shape (BS, N_anchors, N_kpts_per_object, kpts_dim).
+            target_gt_idx (torch.Tensor): Index tensor mapping anchors to ground truth objects, shape (BS, N_anchors).
+            masks (torch.Tensor): Binary mask tensor indicating object presence, shape (BS, N_anchors).
 
         Returns:
-            kpts_loss (torch.Tensor): The keypoints loss.
-            kpts_obj_loss (torch.Tensor): The keypoints object loss.
+            (torch.Tensor): Selected keypoints tensor, shape (BS, N_anchors, N_kpts_per_object, kpts_dim).
         """
         batch_idx = batch_idx.flatten()
         batch_size = len(masks)
@@ -625,6 +715,40 @@ class v8PoseLoss(v8DetectionLoss):
             1, target_gt_idx_expanded.expand(-1, -1, keypoints.shape[1], keypoints.shape[2])
         )
 
+        return selected_keypoints
+
+    def calculate_keypoints_loss(
+        self,
+        masks: torch.Tensor,
+        target_gt_idx: torch.Tensor,
+        keypoints: torch.Tensor,
+        batch_idx: torch.Tensor,
+        stride_tensor: torch.Tensor,
+        target_bboxes: torch.Tensor,
+        pred_kpts: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the keypoints loss for the model.
+
+        This function calculates the keypoints loss and keypoints object loss for a given batch. The keypoints loss is
+        based on the difference between the predicted keypoints and ground truth keypoints. The keypoints object loss is
+        a binary classification loss that classifies whether a keypoint is present or not.
+
+        Args:
+            masks (torch.Tensor): Binary mask tensor indicating object presence, shape (BS, N_anchors).
+            target_gt_idx (torch.Tensor): Index tensor mapping anchors to ground truth objects, shape (BS, N_anchors).
+            keypoints (torch.Tensor): Ground truth keypoints, shape (N_kpts_in_batch, N_kpts_per_object, kpts_dim).
+            batch_idx (torch.Tensor): Batch index tensor for keypoints, shape (N_kpts_in_batch, 1).
+            stride_tensor (torch.Tensor): Stride tensor for anchors, shape (N_anchors, 1).
+            target_bboxes (torch.Tensor): Ground truth boxes in (x1, y1, x2, y2) format, shape (BS, N_anchors, 4).
+            pred_kpts (torch.Tensor): Predicted keypoints, shape (BS, N_anchors, N_kpts_per_object, kpts_dim).
+
+        Returns:
+            kpts_loss (torch.Tensor): The keypoints loss.
+            kpts_obj_loss (torch.Tensor): The keypoints object loss.
+        """
+        # Select target keypoints using helper method
+        selected_keypoints = self._select_target_keypoints(keypoints, batch_idx, target_gt_idx, masks)
+
         # Divide coordinates by stride
         selected_keypoints[..., :2] /= stride_tensor.view(1, -1, 1, 1)
 
@@ -632,6 +756,7 @@ class v8PoseLoss(v8DetectionLoss):
         kpts_obj_loss = 0
 
         if masks.any():
+            target_bboxes /= stride_tensor
             gt_kpt = selected_keypoints[masks]
             area = xyxy2xywh(target_bboxes[masks])[:, 2:].prod(1, keepdim=True)
             pred_kpt = pred_kpts[masks]
@@ -644,6 +769,171 @@ class v8PoseLoss(v8DetectionLoss):
         return kpts_loss, kpts_obj_loss
 
 
+class PoseLoss26(v8PoseLoss):
+    """Criterion class for computing training losses for YOLOv8 pose estimation with RLE loss support."""
+
+    def __init__(self, model, tal_topk: int = 10, tal_topk2: int | None = None):  # model must be de-paralleled
+        """Initialize PoseLoss26 with model parameters and keypoint-specific loss functions including RLE loss."""
+        super().__init__(model, tal_topk, tal_topk2)
+        is_pose = self.kpt_shape == [17, 3]
+        nkpt = self.kpt_shape[0]  # number of keypoints
+        self.rle_loss = None
+        self.flow_model = model.model[-1].flow_model if hasattr(model.model[-1], "flow_model") else None
+        if self.flow_model is not None:
+            self.rle_loss = RLELoss(use_target_weight=True).to(self.device)
+            self.target_weights = (
+                torch.from_numpy(RLE_WEIGHT).to(self.device) if is_pose else torch.ones(nkpt, device=self.device)
+            )
+
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the total loss and detach it for pose estimation."""
+        pred_kpts = preds["kpts"].permute(0, 2, 1).contiguous()
+        loss = torch.zeros(6 if self.rle_loss else 5, device=self.device)  # box, cls, dfl, kpt_location, kpt_visibility
+        (fg_mask, target_gt_idx, target_bboxes, anchor_points, stride_tensor), det_loss, _ = (
+            self.get_assigned_targets_and_loss(preds, batch)
+        )
+        # NOTE: re-assign index for consistency for now. Need to be removed in the future.
+        loss[0], loss[3], loss[4] = det_loss[0], det_loss[1], det_loss[2]
+
+        batch_size = pred_kpts.shape[0]
+        imgsz = torch.tensor(batch["resized_shape"][0], device=self.device, dtype=pred_kpts.dtype)  # image size (h,w)
+
+        pred_kpts = pred_kpts.view(batch_size, -1, *self.kpt_shape)  # (b, h*w, 17, 3)
+
+        if self.rle_loss and preds.get("kpts_sigma", None) is not None:
+            pred_sigma = preds["kpts_sigma"].permute(0, 2, 1).contiguous()
+            pred_sigma = pred_sigma.view(batch_size, -1, self.kpt_shape[0], 2)  # (b, h*w, 17, 2)
+            pred_kpts = torch.cat([pred_kpts, pred_sigma], dim=-1)  # (b, h*w, 17, 5)
+
+        pred_kpts = self.kpts_decode(anchor_points, pred_kpts)
+
+        # Bbox loss
+        if fg_mask.sum():
+            keypoints = batch["keypoints"].to(self.device).float().clone()
+            keypoints[..., 0] *= imgsz[1]
+            keypoints[..., 1] *= imgsz[0]
+
+            keypoints_loss = self.calculate_keypoints_loss(
+                fg_mask,
+                target_gt_idx,
+                keypoints,
+                batch["batch_idx"].view(-1, 1),
+                stride_tensor,
+                target_bboxes,
+                pred_kpts,
+            )
+            loss[1] = keypoints_loss[0]
+            loss[2] = keypoints_loss[1]
+            if self.rle_loss is not None:
+                loss[5] = keypoints_loss[2]
+
+        loss[1] *= self.hyp.pose  # pose gain
+        loss[2] *= self.hyp.kobj  # kobj gain
+        if self.rle_loss is not None:
+            loss[5] *= self.hyp.rle  # rle gain
+
+        return loss * batch_size, loss.detach()  # loss(box, cls, dfl)
+
+    @staticmethod
+    def kpts_decode(anchor_points: torch.Tensor, pred_kpts: torch.Tensor) -> torch.Tensor:
+        """Decode predicted keypoints to image coordinates."""
+        y = pred_kpts.clone()
+        y[..., 0] += anchor_points[:, [0]]
+        y[..., 1] += anchor_points[:, [1]]
+        return y
+
+    def calculate_rle_loss(self, pred_kpt: torch.Tensor, gt_kpt: torch.Tensor, kpt_mask: torch.Tensor) -> torch.Tensor:
+        """Calculate the RLE (Residual Log-likelihood Estimation) loss for keypoints.
+
+        Args:
+            pred_kpt (torch.Tensor): Predicted keypoints with sigma, shape (N, kpts_dim) where kpts_dim >= 4.
+            gt_kpt (torch.Tensor): Ground truth keypoints, shape (N, kpts_dim).
+            kpt_mask (torch.Tensor): Mask for valid keypoints, shape (N, num_keypoints).
+
+        Returns:
+            (torch.Tensor): The RLE loss.
+        """
+        pred_kpt_visible = pred_kpt[kpt_mask]
+        gt_kpt_visible = gt_kpt[kpt_mask]
+        pred_coords = pred_kpt_visible[:, 0:2]
+        pred_sigma = pred_kpt_visible[:, -2:]
+        gt_coords = gt_kpt_visible[:, 0:2]
+
+        target_weights = self.target_weights.unsqueeze(0).repeat(kpt_mask.shape[0], 1)
+        target_weights = target_weights[kpt_mask]
+
+        pred_sigma = pred_sigma.sigmoid()
+        error = (pred_coords - gt_coords) / (pred_sigma + 1e-9)
+
+        # Filter out NaN values to prevent MultivariateNormal validation errors (can occur with small images)
+        valid_mask = ~torch.isnan(error).any(dim=-1)
+        if not valid_mask.any():
+            return torch.tensor(0.0, device=pred_kpt.device)
+
+        error = error[valid_mask]
+        pred_sigma = pred_sigma[valid_mask]
+        target_weights = target_weights[valid_mask]
+
+        log_phi = self.flow_model.log_prob(error)
+
+        return self.rle_loss(pred_sigma, log_phi, error, target_weights)
+
+    def calculate_keypoints_loss(
+        self,
+        masks: torch.Tensor,
+        target_gt_idx: torch.Tensor,
+        keypoints: torch.Tensor,
+        batch_idx: torch.Tensor,
+        stride_tensor: torch.Tensor,
+        target_bboxes: torch.Tensor,
+        pred_kpts: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Calculate the keypoints loss for the model.
+
+        This function calculates the keypoints loss and keypoints object loss for a given batch. The keypoints loss is
+        based on the difference between the predicted keypoints and ground truth keypoints. The keypoints object loss is
+        a binary classification loss that classifies whether a keypoint is present or not.
+
+        Args:
+            masks (torch.Tensor): Binary mask tensor indicating object presence, shape (BS, N_anchors).
+            target_gt_idx (torch.Tensor): Index tensor mapping anchors to ground truth objects, shape (BS, N_anchors).
+            keypoints (torch.Tensor): Ground truth keypoints, shape (N_kpts_in_batch, N_kpts_per_object, kpts_dim).
+            batch_idx (torch.Tensor): Batch index tensor for keypoints, shape (N_kpts_in_batch, 1).
+            stride_tensor (torch.Tensor): Stride tensor for anchors, shape (N_anchors, 1).
+            target_bboxes (torch.Tensor): Ground truth boxes in (x1, y1, x2, y2) format, shape (BS, N_anchors, 4).
+            pred_kpts (torch.Tensor): Predicted keypoints, shape (BS, N_anchors, N_kpts_per_object, kpts_dim).
+
+        Returns:
+            kpts_loss (torch.Tensor): The keypoints loss.
+            kpts_obj_loss (torch.Tensor): The keypoints object loss.
+            rle_loss (torch.Tensor): The RLE loss.
+        """
+        # Select target keypoints using inherited helper method
+        selected_keypoints = self._select_target_keypoints(keypoints, batch_idx, target_gt_idx, masks)
+
+        # Divide coordinates by stride
+        selected_keypoints[..., :2] /= stride_tensor.view(1, -1, 1, 1)
+
+        kpts_loss = 0
+        kpts_obj_loss = 0
+        rle_loss = 0
+
+        if masks.any():
+            target_bboxes /= stride_tensor
+            gt_kpt = selected_keypoints[masks]
+            area = xyxy2xywh(target_bboxes[masks])[:, 2:].prod(1, keepdim=True)
+            pred_kpt = pred_kpts[masks]
+            kpt_mask = gt_kpt[..., 2] != 0 if gt_kpt.shape[-1] == 3 else torch.full_like(gt_kpt[..., 0], True)
+            kpts_loss = self.keypoint_loss(pred_kpt, gt_kpt, kpt_mask, area)  # pose loss
+
+            if self.rle_loss is not None and (pred_kpt.shape[-1] == 4 or pred_kpt.shape[-1] == 5):
+                rle_loss = self.calculate_rle_loss(pred_kpt, gt_kpt, kpt_mask)
+            if pred_kpt.shape[-1] == 3 or pred_kpt.shape[-1] == 5:
+                kpts_obj_loss = self.bce_pose(pred_kpt[..., 2], kpt_mask.float())  # keypoint obj loss
+
+        return kpts_loss, kpts_obj_loss, rle_loss
+
+
 class v8ClassificationLoss:
     """Criterion class for computing training losses for classification."""
 
@@ -657,10 +947,17 @@ class v8ClassificationLoss:
 class v8OBBLoss(v8DetectionLoss):
     """Calculates losses for object detection, classification, and box distribution in rotated YOLO models."""
 
-    def __init__(self, model):
+    def __init__(self, model, tal_topk=10, tal_topk2: int | None = None):
         """Initialize v8OBBLoss with model, assigner, and rotated bbox loss; model must be de-paralleled."""
-        super().__init__(model)
-        self.assigner = RotatedTaskAlignedAssigner(topk=10, num_classes=self.nc, alpha=0.5, beta=6.0)
+        super().__init__(model, tal_topk=tal_topk)
+        self.assigner = RotatedTaskAlignedAssigner(
+            topk=tal_topk,
+            num_classes=self.nc,
+            alpha=0.5,
+            beta=6.0,
+            stride=self.stride.tolist(),
+            topk2=tal_topk2,
+        )
         self.bbox_loss = RotatedBboxLoss(self.reg_max).to(self.device)
 
     def preprocess(self, targets: torch.Tensor, batch_size: int, scale_tensor: torch.Tensor) -> torch.Tensor:
@@ -680,23 +977,19 @@ class v8OBBLoss(v8DetectionLoss):
                     out[j, :n] = torch.cat([targets[matches, 1:2], bboxes], dim=-1)
         return out
 
-    def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         """Calculate and return the loss for oriented bounding box detection."""
-        loss = torch.zeros(3, device=self.device)  # box, cls, dfl
-        feats, pred_angle = preds if isinstance(preds[0], list) else preds[1]
-        batch_size = pred_angle.shape[0]  # batch size
-        pred_distri, pred_scores = torch.cat([xi.view(feats[0].shape[0], self.no, -1) for xi in feats], 2).split(
-            (self.reg_max * 4, self.nc), 1
+        loss = torch.zeros(4, device=self.device)  # box, cls, dfl
+        pred_distri, pred_scores, pred_angle = (
+            preds["boxes"].permute(0, 2, 1).contiguous(),
+            preds["scores"].permute(0, 2, 1).contiguous(),
+            preds["angle"].permute(0, 2, 1).contiguous(),
         )
-
-        # b, grids, ..
-        pred_scores = pred_scores.permute(0, 2, 1).contiguous()
-        pred_distri = pred_distri.permute(0, 2, 1).contiguous()
-        pred_angle = pred_angle.permute(0, 2, 1).contiguous()
+        anchor_points, stride_tensor = make_anchors(preds["feats"], self.stride, 0.5)
+        batch_size = pred_angle.shape[0]  # batch size, number of masks, mask height, mask width
 
         dtype = pred_scores.dtype
-        imgsz = torch.tensor(feats[0].shape[2:], device=self.device, dtype=dtype) * self.stride[0]  # image size (h,w)
-        anchor_points, stride_tensor = make_anchors(feats, self.stride, 0.5)
+        imgsz = torch.tensor(batch["resized_shape"][0], device=self.device, dtype=dtype)  # image size (h,w)
 
         # targets
         try:
@@ -704,14 +997,14 @@ class v8OBBLoss(v8DetectionLoss):
             targets = torch.cat((batch_idx, batch["cls"].view(-1, 1), batch["bboxes"].view(-1, 5)), 1)
             rw, rh = targets[:, 4] * float(imgsz[1]), targets[:, 5] * float(imgsz[0])
             targets = targets[(rw >= 2) & (rh >= 2)]  # filter rboxes of tiny size to stabilize training
-            targets = self.preprocess(targets, batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
+            targets = self.preprocess(targets.to(self.device), batch_size, scale_tensor=imgsz[[1, 0, 1, 0]])
             gt_labels, gt_bboxes = targets.split((1, 5), 2)  # cls, xywhr
             mask_gt = gt_bboxes.sum(2, keepdim=True).gt_(0.0)
         except RuntimeError as e:
             raise TypeError(
                 "ERROR ❌ OBB dataset incorrectly formatted or not a OBB dataset.\n"
                 "This error can occur when incorrectly training a 'OBB' model on a 'detect' dataset, "
-                "i.e. 'yolo train model=yolo11n-obb.pt data=coco8.yaml'.\nVerify your dataset is a "
+                "i.e. 'yolo train model=yolo11n-obb.pt data=dota8.yaml'.\nVerify your dataset is a "
                 "correctly formatted 'OBB' dataset using 'data=dota8.yaml' "
                 "as an example.\nSee https://docs.ultralytics.com/datasets/obb/ for help."
             ) from e
@@ -741,16 +1034,29 @@ class v8OBBLoss(v8DetectionLoss):
         if fg_mask.sum():
             target_bboxes[..., :4] /= stride_tensor
             loss[0], loss[2] = self.bbox_loss(
-                pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask
+                pred_distri,
+                pred_bboxes,
+                anchor_points,
+                target_bboxes,
+                target_scores,
+                target_scores_sum,
+                fg_mask,
+                imgsz,
+                stride_tensor,
             )
+            weight = target_scores.sum(-1)[fg_mask]
+            loss[3] = self.calculate_angle_loss(
+                pred_bboxes, target_bboxes, fg_mask, weight, target_scores_sum
+            )  # angle loss
         else:
             loss[0] += (pred_angle * 0).sum()
 
         loss[0] *= self.hyp.box  # box gain
         loss[1] *= self.hyp.cls  # cls gain
         loss[2] *= self.hyp.dfl  # dfl gain
+        loss[3] *= self.hyp.angle  # angle gain
 
-        return loss * batch_size, loss.detach()  # loss(box, cls, dfl)
+        return loss * batch_size, loss.detach()  # loss(box, cls, dfl, angle)
 
     def bbox_decode(
         self, anchor_points: torch.Tensor, pred_dist: torch.Tensor, pred_angle: torch.Tensor
@@ -770,6 +1076,34 @@ class v8OBBLoss(v8DetectionLoss):
             pred_dist = pred_dist.view(b, a, 4, c // 4).softmax(3).matmul(self.proj.type(pred_dist.dtype))
         return torch.cat((dist2rbox(pred_dist, pred_angle, anchor_points), pred_angle), dim=-1)
 
+    def calculate_angle_loss(self, pred_bboxes, target_bboxes, fg_mask, weight, target_scores_sum, lambda_val=3):
+        """Calculate oriented angle loss.
+
+        Args:
+            pred_bboxes: [N, 5] (x, y, w, h, theta).
+            target_bboxes: [N, 5] (x, y, w, h, theta).
+            fg_mask: Foreground mask indicating valid predictions.
+            weight: Loss weights for each prediction.
+            target_scores_sum: Sum of target scores for normalization.
+            lambda_val: control the sensitivity to aspect ratio.
+        """
+        w_gt = target_bboxes[..., 2]
+        h_gt = target_bboxes[..., 3]
+        pred_theta = pred_bboxes[..., 4]
+        target_theta = target_bboxes[..., 4]
+
+        log_ar = torch.log(w_gt / h_gt)
+        scale_weight = torch.exp(-(log_ar**2) / (lambda_val**2))
+
+        delta_theta = pred_theta - target_theta
+        delta_theta_wrapped = delta_theta - torch.round(delta_theta / math.pi) * math.pi
+        ang_loss = torch.sin(2 * delta_theta_wrapped[fg_mask]) ** 2
+
+        ang_loss = scale_weight[fg_mask] * ang_loss
+        ang_loss = ang_loss * weight
+
+        return ang_loss.sum() / target_scores_sum
+
 
 class E2EDetectLoss:
     """Criterion class for computing training losses for end-to-end detection."""
@@ -789,61 +1123,108 @@ class E2EDetectLoss:
         return loss_one2many[0] + loss_one2one[0], loss_one2many[1] + loss_one2one[1]
 
 
+class E2ELoss:
+    """Criterion class for computing training losses for end-to-end detection."""
+
+    def __init__(self, model, loss_fn=v8DetectionLoss):
+        """Initialize E2ELoss with one-to-many and one-to-one detection losses using the provided model."""
+        self.one2many = loss_fn(model, tal_topk=10)
+        self.one2one = loss_fn(model, tal_topk=7, tal_topk2=1)
+        self.updates = 0
+        self.total = 1.0
+        # init gain
+        self.o2m = 0.8
+        self.o2o = self.total - self.o2m
+        self.o2m_copy = self.o2m
+        # final gain
+        self.final_o2m = 0.1
+
+    def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the sum of the loss for box, cls and dfl multiplied by batch size."""
+        preds = self.one2many.parse_output(preds)
+        one2many, one2one = preds["one2many"], preds["one2one"]
+        loss_one2many = self.one2many.loss(one2many, batch)
+        loss_one2one = self.one2one.loss(one2one, batch)
+        return loss_one2many[0] * self.o2m + loss_one2one[0] * self.o2o, loss_one2one[1]
+
+    def update(self) -> None:
+        """Update the weights for one-to-many and one-to-one losses based on the decay schedule."""
+        self.updates += 1
+        self.o2m = self.decay(self.updates)
+        self.o2o = max(self.total - self.o2m, 0)
+
+    def decay(self, x) -> float:
+        """Calculate the decayed weight for one-to-many loss based on the current update step."""
+        return max(1 - x / max(self.one2one.hyp.epochs - 1, 1), 0) * (self.o2m_copy - self.final_o2m) + self.final_o2m
+
+
 class TVPDetectLoss:
     """Criterion class for computing training losses for text-visual prompt detection."""
 
-    def __init__(self, model):
+    def __init__(self, model, tal_topk=10):
         """Initialize TVPDetectLoss with task-prompt and visual-prompt criteria using the provided model."""
-        self.vp_criterion = v8DetectionLoss(model)
+        self.vp_criterion = v8DetectionLoss(model, tal_topk)
         # NOTE: store following info as it's changeable in __call__
+        self.hyp = self.vp_criterion.hyp
         self.ori_nc = self.vp_criterion.nc
         self.ori_no = self.vp_criterion.no
         self.ori_reg_max = self.vp_criterion.reg_max
 
+    def parse_output(self, preds) -> dict[str, torch.Tensor]:
+        """Parse model predictions to extract features."""
+        return self.vp_criterion.parse_output(preds)
+
     def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         """Calculate the loss for text-visual prompt detection."""
-        feats = preds[1] if isinstance(preds, tuple) else preds
+        return self.loss(self.parse_output(preds), batch)
+
+    def loss(self, preds: dict[str, torch.Tensor], batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the loss for text-visual prompt detection."""
+        assert self.ori_reg_max == self.vp_criterion.reg_max  # TODO: remove it
 
-        if self.ori_reg_max * 4 + self.ori_nc == feats[0].shape[1]:
+        if self.ori_nc == preds["scores"].shape[1]:
             loss = torch.zeros(3, device=self.vp_criterion.device, requires_grad=True)
             return loss, loss.detach()
 
-        vp_feats = self._get_vp_features(feats)
-        vp_loss = self.vp_criterion(vp_feats, batch)
-        cls_loss = vp_loss[0][1]
-        return cls_loss, vp_loss[1]
+        preds["scores"] = self._get_vp_features(preds)
+        vp_loss = self.vp_criterion(preds, batch)
+        box_loss = vp_loss[0][1]
+        return box_loss, vp_loss[1]
 
-    def _get_vp_features(self, feats: list[torch.Tensor]) -> list[torch.Tensor]:
+    def _get_vp_features(self, preds: dict[str, torch.Tensor]) -> list[torch.Tensor]:
         """Extract visual-prompt features from the model output."""
-        vnc = feats[0].shape[1] - self.ori_reg_max * 4 - self.ori_nc
+        # NOTE: remove empty placeholder
+        scores = preds["scores"][:, self.ori_nc :, :]
+        vnc = scores.shape[1]
 
         self.vp_criterion.nc = vnc
         self.vp_criterion.no = vnc + self.vp_criterion.reg_max * 4
         self.vp_criterion.assigner.num_classes = vnc
-
-        return [
-            torch.cat((box, cls_vp), dim=1)
-            for box, _, cls_vp in [xi.split((self.ori_reg_max * 4, self.ori_nc, vnc), dim=1) for xi in feats]
-        ]
+        return scores
 
 
 class TVPSegmentLoss(TVPDetectLoss):
     """Criterion class for computing training losses for text-visual prompt segmentation."""
 
-    def __init__(self, model):
+    def __init__(self, model, tal_topk=10):
         """Initialize TVPSegmentLoss with task-prompt and visual-prompt criteria using the provided model."""
         super().__init__(model)
-        self.vp_criterion = v8SegmentationLoss(model)
+        self.vp_criterion = v8SegmentationLoss(model, tal_topk)
+        self.hyp = self.vp_criterion.hyp
 
     def __call__(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         """Calculate the loss for text-visual prompt segmentation."""
-        feats, pred_masks, proto = preds if len(preds) == 3 else preds[1]
+        return self.loss(self.parse_output(preds), batch)
+
+    def loss(self, preds: Any, batch: dict[str, torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
+        """Calculate the loss for text-visual prompt detection."""
+        assert self.ori_reg_max == self.vp_criterion.reg_max  # TODO: remove it
 
-        if self.ori_reg_max * 4 + self.ori_nc == feats[0].shape[1]:
+        if self.ori_nc == preds["scores"].shape[1]:
             loss = torch.zeros(4, device=self.vp_criterion.device, requires_grad=True)
             return loss, loss.detach()
 
-        vp_feats = self._get_vp_features(feats)
-        vp_loss = self.vp_criterion((vp_feats, pred_masks, proto), batch)
+        preds["scores"] = self._get_vp_features(preds)
+        vp_loss = self.vp_criterion(preds, batch)
         cls_loss = vp_loss[0][2]
         return cls_loss, vp_loss[1]