YOLOv8损失函数解读
因为笔者目前做船舶感知检测的方向，想对yolov8进行改进创新。故从零开始学习yolov8。
参考博客：
YOLOv8详解:损失函数、Anchor-Free、样本分配策略；以及与v5的对比_yolov8损失函数-CSDN博客
YoloV8损失函数篇（代码加理论）-CSDN博客
一、损失函数
损失函数的权重在ultralytics/cfg/default.yaml中可以找到：
# Ultralytics AGPL-3.0 License - https://ultralytics.com/license
# Global configuration YAML with settings and hyperparameters for YOLO training, validation, prediction and export
# For documentation see https://docs.ultralytics.com/usage/cfg/
task: detect # (str) YOLO task, i.e. detect, segment, classify, pose, obb
mode: train # (str) YOLO mode, i.e. train, val, predict, export, track, benchmark
# Train settings -------------------------------------------------------------------------------------------------------
model: # (str, optional) path to model file, i.e. yolov8n.pt, yolov8n.yaml
data: # (str, optional) path to data file, i.e. coco8.yaml
epochs: 100 # (int) number of epochs to train for
time: # (float, optional) number of hours to train for, overrides epochs if supplied
patience: 100 # (int) epochs to wait for no observable improvement for early stopping of training
batch: 16 # (int) number of images per batch (-1 for AutoBatch)
imgsz: 640 # (int | list) input images size as int for train and val modes, or list[h,w] for predict and export modes
save: True # (bool) save train checkpoints and predict results
save_period: -1 # (int) Save checkpoint every x epochs (disabled if < 1)
cache: False # (bool) True/ram, disk or False. Use cache for data loading
device: # (int | str | list, optional) device to run on, i.e. cuda device=0 or device=0,1,2,3 or device=cpu
workers: 8 # (int) number of worker threads for data loading (per RANK if DDP)
project: # (str, optional) project name
name: # (str, optional) experiment name, results saved to 'project/name' directory
exist_ok: False # (bool) whether to overwrite existing experiment
pretrained: True # (bool | str) whether to use a pretrained model (bool) or a model to load weights from (str)
optimizer: auto # (str) optimizer to use, choices=[SGD, Adam, Adamax, AdamW, NAdam, RAdam, RMSProp, auto]
verbose: True # (bool) whether to print verbose output
seed: 0 # (int) random seed for reproducibility
deterministic: True # (bool) whether to enable deterministic mode
single_cls: False # (bool) train multi-class data as single-class
rect: False # (bool) rectangular training if mode='train' or rectangular validation if mode='val'
cos_lr: False # (bool) use cosine learning rate scheduler
close_mosaic: 10 # (int) disable mosaic augmentation for final epochs (0 to disable)
resume: False # (bool) resume training from last checkpoint
amp: True # (bool) Automatic Mixed Precision (AMP) training, choices=[True, False], True runs AMP check
fraction: 1.0 # (float) dataset fraction to train on (default is 1.0, all images in train set)
profile: False # (bool) profile ONNX and TensorRT speeds during training for loggers
freeze: None # (int | list, optional) freeze first n layers, or freeze list of layer indices during training
multi_scale: False # (bool) Whether to use multiscale during training
# Segmentation
overlap_mask: True # (bool) merge object masks into a single image mask during training (segment train only)
mask_ratio: 4 # (int) mask downsample ratio (segment train only)
# Classification
dropout: 0.0 # (float) use dropout regularization (classify train only)
# Val/Test settings ----------------------------------------------------------------------------------------------------
val: True # (bool) validate/test during training
split: val # (str) dataset split to use for validation, i.e. 'val', 'test' or 'train'
save_json: False # (bool) save results to JSON file
save_hybrid: False # (bool) save hybrid version of labels (labels + additional predictions)
conf: # (float, optional) object confidence threshold for detection (default 0.25 predict, 0.001 val)
iou: 0.7 # (float) intersection over union (IoU) threshold for NMS
max_det: 300 # (int) maximum number of detections per image
half: False # (bool) use half precision (FP16)
dnn: False # (bool) use OpenCV DNN for ONNX inference
plots: True # (bool) save plots and images during train/val
# Predict settings -----------------------------------------------------------------------------------------------------
source: # (str, optional) source directory for images or videos
vid_stride: 1 # (int) video frame-rate stride
stream_buffer: False # (bool) buffer all streaming frames (True) or return the most recent frame (False)
visualize: False # (bool) visualize model features
augment: False # (bool) apply image augmentation to prediction sources
agnostic_nms: False # (bool) class-agnostic NMS
classes: # (int | list[int], optional) filter results by class, i.e. classes=0, or classes=[0,2,3]
retina_masks: False # (bool) use high-resolution segmentation masks
embed: # (list[int], optional) return feature vectors/embeddings from given layers
# Visualize settings ---------------------------------------------------------------------------------------------------
show: False # (bool) show predicted images and videos if environment allows
save_frames: False # (bool) save predicted individual video frames
save_txt: False # (bool) save results as .txt file
save_conf: False # (bool) save results with confidence scores
save_crop: False # (bool) save cropped images with results
show_labels: True # (bool) show prediction labels, i.e. 'person'
show_conf: True # (bool) show prediction confidence, i.e. '0.99'
show_boxes: True # (bool) show prediction boxes
line_width: # (int, optional) line width of the bounding boxes. Scaled to image size if None.
# Export settings ------------------------------------------------------------------------------------------------------
format: torchscript # (str) format to export to, choices at https://docs.ultralytics.com/modes/export/#export-formats
keras: False # (bool) use Kera=s
optimize: False # (bool) TorchScript: optimize for mobile
int8: False # (bool) CoreML/TF INT8 quantization
dynamic: False # (bool) ONNX/TF/TensorRT: dynamic axes
simplify: True # (bool) ONNX: simplify model using `onnxslim`
opset: # (int, optional) ONNX: opset version
workspace: None # (float, optional) TensorRT: workspace size (GiB), `None` will let TensorRT auto-allocate memory
nms: False # (bool) CoreML: add NMS
# Hyperparameters ------------------------------------------------------------------------------------------------------
lr0: 0.01 # (float) initial learning rate (i.e. SGD=1E-2, Adam=1E-3)
lrf: 0.01 # (float) final learning rate (lr0 * lrf)
momentum: 0.937 # (float) SGD momentum/Adam beta1
weight_decay: 0.0005 # (float) optimizer weight decay 5e-4
warmup_epochs: 3.0 # (float) warmup epochs (fractions ok)
warmup_momentum: 0.8 # (float) warmup initial momentum
warmup_bias_lr: 0.1 # (float) warmup initial bias lr
box: 7.5 # (float) box loss gain
cls: 0.5 # (float) cls loss gain (scale with pixels)
dfl: 1.5 # (float) dfl loss gain
pose: 12.0 # (float) pose loss gain
kobj: 1.0 # (float) keypoint obj loss gain
nbs: 64 # (int) nominal batch size
hsv_h: 0.015 # (float) image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # (float) image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # (float) image HSV-Value augmentation (fraction)
degrees: 0.0 # (float) image rotation (+/- deg)
translate: 0.1 # (float) image translation (+/- fraction)
scale: 0.5 # (float) image scale (+/- gain)
shear: 0.0 # (float) image shear (+/- deg)
perspective: 0.0 # (float) image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # (float) image flip up-down (probability)
fliplr: 0.5 # (float) image flip left-right (probability)
bgr: 0.0 # (float) image channel BGR (probability)
mosaic: 1.0 # (float) image mosaic (probability)
mixup: 0.0 # (float) image mixup (probability)
copy_paste: 0.0 # (float) segment copy-paste (probability)
copy_paste_mode: "flip" # (str) the method to do copy_paste augmentation (flip, mixup)
auto_augment: randaugment # (str) auto augmentation policy for classification (randaugment, autoaugment, augmix)
erasing: 0.4 # (float) probability of random erasing during classification training (0-0.9), 0 means no erasing, must be less than 1.0.
crop_fraction: 1.0 # (float) image crop fraction for classification (0.1-1), 1.0 means no crop, must be greater than 0.
# Custom config.yaml ---------------------------------------------------------------------------------------------------
cfg: # (str, optional) for overriding defaults.yaml
# Tracker settings ------------------------------------------------------------------------------------------------------
tracker: botsort.yaml # (str) tracker type, choices=[botsort.yaml, bytetrack.yaml]
损失函数定义的文件在ultralytics/utils/loss.py
# Ultralytics AGPL-3.0 License - https://ultralytics.com/license
import torch
import torch.nn as nn
import torch.nn.functional as F
from ultralytics.utils.metrics import OKS_SIGMA
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
class VarifocalLoss(nn.Module):
"""
Varifocal loss by Zhang et al.
https://arxiv.org/abs/2008.13367.
"""
def __init__(self):
"""Initialize the VarifocalLoss class."""
super().__init__()
@staticmethod
def forward(pred_score, gt_score, label, alpha=0.75, gamma=2.0):
"""Computes varfocal loss."""
weight = alpha * pred_score.sigmoid().pow(gamma) * (1 - label) + gt_score * label
with autocast(enabled=False):
loss = (
(F.binary_cross_entropy_with_logits(pred_score.float(), gt_score.float(), reduction="none") * weight)
.mean(1)
.sum()
)
return loss
class FocalLoss(nn.Module):
"""Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)."""
def __init__(self):
"""Initializer for FocalLoss class with no parameters."""
super().__init__()
@staticmethod
def forward(pred, label, gamma=1.5, alpha=0.25):
"""Calculates and updates confusion matrix for object detection/classification tasks."""
loss = F.binary_cross_entropy_with_logits(pred, label, reduction="none")
# p_t = torch.exp(-loss)
# loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability
# TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
pred_prob = pred.sigmoid() # prob from logits
p_t = label * pred_prob + (1 - label) * (1 - pred_prob)
modulating_factor = (1.0 - p_t) ** gamma
loss *= modulating_factor
if alpha > 0:
alpha_factor = label * alpha + (1 - label) * (1 - alpha)
loss *= alpha_factor
return loss.mean(1).sum()
class DFLoss(nn.Module):
"""Criterion class for computing DFL losses during training."""
def __init__(self, reg_max=16) -> None:
"""Initialize the DFL module."""
super().__init__()
self.reg_max = reg_max
def __call__(self, pred_dist, target):
"""
Return sum of left and right DFL losses.
Distribution Focal Loss (DFL) proposed in Generalized Focal Loss
https://ieeexplore.ieee.org/document/9792391
"""
target = target.clamp_(0, self.reg_max - 1 - 0.01)
tl = target.long() # target left
tr = tl + 1 # target right
wl = tr - target # weight left
wr = 1 - wl # weight right
return (
F.cross_entropy(pred_dist, tl.view(-1), reduction="none").view(tl.shape) * wl
+ F.cross_entropy(pred_dist, tr.view(-1), reduction="none").view(tl.shape) * wr
).mean(-1, keepdim=True)
class BboxLoss(nn.Module):
"""Criterion class for computing training losses during training."""
def __init__(self, reg_max=16):
"""Initialize the BboxLoss module with regularization maximum and DFL settings."""
super().__init__()
self.dfl_loss = DFLoss(reg_max) if reg_max > 1 else None
def forward(self, pred_dist, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask):
"""IoU loss."""
weight = target_scores.sum(-1)[fg_mask].unsqueeze(-1)
iou = bbox_iou(pred_bboxes[fg_mask], target_bboxes[fg_mask], xywh=False, CIoU=True)
loss_iou = ((1.0 - iou) * weight).sum() / target_scores_sum
# DFL loss
if self.dfl_loss:
target_ltrb = bbox2dist(anchor_points, target_bboxes, 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)
return loss_iou, loss_dfl
class RotatedBboxLoss(BboxLoss):
"""Criterion class for computing training losses during training."""
def __init__(self, reg_max):
"""Initialize the BboxLoss module with regularization maximum and DFL settings."""
super().__init__(reg_max)
def forward(self, pred_dist, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask):
"""IoU loss."""
weight = target_scores.sum(-1)[fg_mask].unsqueeze(-1)
iou = probiou(pred_bboxes[fg_mask], target_bboxes[fg_mask])
loss_iou = ((1.0 - iou) * weight).sum() / target_scores_sum
# DFL loss
if self.dfl_loss:
target_ltrb = bbox2dist(anchor_points, xywh2xyxy(target_bboxes[..., :4]), 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)
return loss_iou, loss_dfl
class KeypointLoss(nn.Module):
"""Criterion class for computing training losses."""
def __init__(self, sigmas) -> None:
"""Initialize the KeypointLoss class."""
super().__init__()
self.sigmas = sigmas
def forward(self, pred_kpts, gt_kpts, kpt_mask, area):
"""Calculates keypoint loss factor and Euclidean distance loss for predicted and actual keypoints."""
d = (pred_kpts[..., 0] - gt_kpts[..., 0]).pow(2) + (pred_kpts[..., 1] - gt_kpts[..., 1]).pow(2)
kpt_loss_factor = kpt_mask.shape[1] / (torch.sum(kpt_mask != 0, dim=1) + 1e-9)
# e = d / (2 * (area * self.sigmas) ** 2 + 1e-9) # from formula
e = d / ((2 * self.sigmas).pow(2) * (area + 1e-9) * 2) # from cocoeval
return (kpt_loss_factor.view(-1, 1) * ((1 - torch.exp(-e)) * kpt_mask)).mean()
class v8DetectionLoss:
"""Criterion class for computing training losses."""
def __init__(self, model, tal_topk=10): # model must be de-paralleled
"""Initializes v8DetectionLoss with the model, defining model-related properties and BCE loss function."""
device = next(model.parameters()).device # get model device
h = model.args # hyperparameters
m = model.model[-1] # Detect() module
self.bce = nn.BCEWithLogitsLoss(reduction="none")
self.hyp = h
self.stride = m.stride # model strides
self.nc = m.nc # number of classes
self.no = m.nc + m.reg_max * 4
self.reg_max = m.reg_max
self.device = device
self.use_dfl = m.reg_max > 1
self.assigner = TaskAlignedAssigner(topk=tal_topk, num_classes=self.nc, alpha=0.5, beta=6.0)
self.bbox_loss = BboxLoss(m.reg_max).to(device)
self.proj = torch.arange(m.reg_max, dtype=torch.float, device=device)
def preprocess(self, targets, batch_size, scale_tensor):
"""Preprocesses the target counts and matches with the input batch size to output a tensor."""
nl, ne = targets.shape
if nl == 0:
out = torch.zeros(batch_size, 0, ne - 1, device=self.device)
else:
i = targets[:, 0] # image index
_, counts = i.unique(return_counts=True)
counts = counts.to(dtype=torch.int32)
out = torch.zeros(batch_size, counts.max(), ne - 1, device=self.device)
for j in range(batch_size):
matches = i == j
if n := matches.sum():
out[j, :n] = targets[matches, 1:]
out[..., 1:5] = xywh2xyxy(out[..., 1:5].mul_(scale_tensor))
return out
def bbox_decode(self, anchor_points, pred_dist):
"""Decode predicted object bounding box coordinates from anchor points and distribution."""
if self.use_dfl:
b, a, c = pred_dist.shape # batch, anchors, channels
pred_dist = pred_dist.view(b, a, 4, c // 4).softmax(3).matmul(self.proj.type(pred_dist.dtype))
# pred_dist = pred_dist.view(b, a, c // 4, 4).transpose(2,3).softmax(3).matmul(self.proj.type(pred_dist.dtype))
# 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, batch):
"""Calculate the sum of the loss for box, cls and dfl multiplied by batch size."""
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_scores = pred_scores.permute(0, 2, 1).contiguous()
pred_distri = pred_distri.permute(0, 2, 1).contiguous()
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)
# Targets
targets = torch.cat((batch["batch_idx"].view(-1, 1), batch["cls"].view(-1, 1), batch["bboxes"]), 1)
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,
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[1] = 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[2] = self.bbox_loss(
pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask
)
loss[0] *= self.hyp.box # box gain
loss[1] *= self.hyp.cls # cls gain
loss[2] *= self.hyp.dfl # dfl gain
return loss.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
class v8SegmentationLoss(v8DetectionLoss):
"""Criterion class for computing training losses."""
def __init__(self, model): # model must be de-paralleled
"""Initializes the v8SegmentationLoss class, taking a de-paralleled model as argument."""
super().__init__(model)
self.overlap = model.args.overlap_mask
def __call__(self, preds, batch):
"""Calculate and return the loss for the YOLO model."""
loss = torch.zeros(4, device=self.device) # box, 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.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)
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
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]
loss[1] = self.calculate_segmentation_loss(
fg_mask, masks, target_gt_idx, target_bboxes, batch_idx, proto, pred_masks, imgsz, self.overlap
)
# 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[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.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
@staticmethod
def single_mask_loss(
gt_mask: torch.Tensor, pred: torch.Tensor, proto: torch.Tensor, xyxy: torch.Tensor, area: torch.Tensor
) -> torch.Tensor:
"""
Compute the instance segmentation loss for a single image.
Args:
gt_mask (torch.Tensor): Ground truth mask of shape (n, H, W), where n is the number of objects.
pred (torch.Tensor): Predicted mask coefficients of shape (n, 32).
proto (torch.Tensor): Prototype masks of shape (32, H, W).
xyxy (torch.Tensor): Ground truth bounding boxes in xyxy format, normalized to [0, 1], of shape (n, 4).
area (torch.Tensor): Area of each ground truth bounding box of shape (n,).
Returns:
(torch.Tensor): The calculated mask loss for a single image.
Notes:
The function uses the equation pred_mask = torch.einsum('in,nhw->ihw', pred, proto) to produce the
predicted masks from the prototype masks and predicted mask coefficients.
"""
pred_mask = torch.einsum("in,nhw->ihw", pred, proto) # (n, 32) @ (32, 80, 80) -> (n, 80, 80)
loss = F.binary_cross_entropy_with_logits(pred_mask, gt_mask, reduction="none")
return (crop_mask(loss, xyxy).mean(dim=(1, 2)) / area).sum()
def calculate_segmentation_loss(
self,
fg_mask: torch.Tensor,
masks: torch.Tensor,
target_gt_idx: torch.Tensor,
target_bboxes: torch.Tensor,
batch_idx: torch.Tensor,
proto: torch.Tensor,
pred_masks: torch.Tensor,
imgsz: torch.Tensor,
overlap: bool,
) -> torch.Tensor:
"""
Calculate the loss for instance segmentation.
Args:
fg_mask (torch.Tensor): A binary tensor of shape (BS, N_anchors) indicating which anchors are positive.
masks (torch.Tensor): Ground truth masks of shape (BS, H, W) if `overlap` is False, otherwise (BS, ?, H, W).
target_gt_idx (torch.Tensor): Indexes of ground truth objects for each anchor of shape (BS, N_anchors).
target_bboxes (torch.Tensor): Ground truth bounding boxes for each anchor of shape (BS, N_anchors, 4).
batch_idx (torch.Tensor): Batch indices of shape (N_labels_in_batch, 1).
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.
Notes:
The batch loss can be computed for improved speed at higher memory usage.
For example, pred_mask can be computed as follows:
pred_mask = torch.einsum('in,nhw->ihw', pred, proto) # (i, 32) @ (32, 160, 160) -> (i, 160, 160)
"""
_, _, mask_h, mask_w = proto.shape
loss = 0
# Normalize to 0-1
target_bboxes_normalized = target_bboxes / imgsz[[1, 0, 1, 0]]
# Areas of target bboxes
marea = xyxy2xywh(target_bboxes_normalized)[..., 2:].prod(2)
# Normalize to mask size
mxyxy = target_bboxes_normalized * torch.tensor([mask_w, mask_h, mask_w, mask_h], device=proto.device)
for i, single_i in enumerate(zip(fg_mask, target_gt_idx, pred_masks, proto, mxyxy, marea, masks)):
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:
gt_mask = masks_i == (mask_idx + 1).view(-1, 1, 1)
gt_mask = gt_mask.float()
else:
gt_mask = masks[batch_idx.view(-1) == i][mask_idx]
loss += self.single_mask_loss(
gt_mask, pred_masks_i[fg_mask_i], proto_i, mxyxy_i[fg_mask_i], marea_i[fg_mask_i]
)
# WARNING: lines below prevents Multi-GPU DDP 'unused gradient' PyTorch errors, do not remove
else:
loss += (proto * 0).sum() + (pred_masks * 0).sum() # inf sums may lead to nan loss
return loss / fg_mask.sum()
class v8PoseLoss(v8DetectionLoss):
"""Criterion class for computing training losses."""
def __init__(self, model): # model must be de-paralleled
"""Initializes v8PoseLoss with model, sets keypoint variables and declares a keypoint loss instance."""
super().__init__(model)
self.kpt_shape = model.model[-1].kpt_shape
self.bce_pose = nn.BCEWithLogitsLoss()
is_pose = self.kpt_shape == [17, 3]
nkpt = self.kpt_shape[0] # number of keypoints
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, batch):
"""Calculate the total loss and detach it."""
loss = torch.zeros(5, device=self.device) # box, cls, dfl, kpt_location, kpt_visibility
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
)
# 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.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)
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
)
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.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
@staticmethod
def kpts_decode(anchor_points, pred_kpts):
"""Decodes predicted keypoints to image coordinates."""
y = pred_kpts.clone()
y[..., :2] *= 2.0
y[..., 0] += anchor_points[:, [0]] - 0.5
y[..., 1] += anchor_points[:, [1]] - 0.5
return y
def calculate_keypoints_loss(
self, masks, target_gt_idx, keypoints, batch_idx, stride_tensor, target_bboxes, pred_kpts
):
"""
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.
"""
batch_idx = batch_idx.flatten()
batch_size = len(masks)
# Find the maximum number of keypoints in a single image
max_kpts = torch.unique(batch_idx, return_counts=True)[1].max()
# Create a tensor to hold batched keypoints
batched_keypoints = torch.zeros(
(batch_size, max_kpts, keypoints.shape[1], keypoints.shape[2]), device=keypoints.device
)
# TODO: any idea how to vectorize this?
# Fill batched_keypoints with keypoints based on batch_idx
for i in range(batch_size):
keypoints_i = keypoints[batch_idx == i]
batched_keypoints[i, : keypoints_i.shape[0]] = keypoints_i
# Expand dimensions of target_gt_idx to match the shape of batched_keypoints
target_gt_idx_expanded = target_gt_idx.unsqueeze(-1).unsqueeze(-1)
# Use target_gt_idx_expanded to select keypoints from batched_keypoints
selected_keypoints = batched_keypoints.gather(
1, target_gt_idx_expanded.expand(-1, -1, keypoints.shape[1], keypoints.shape[2])
)
# Divide coordinates by stride
selected_keypoints /= stride_tensor.view(1, -1, 1, 1)
kpts_loss = 0
kpts_obj_loss = 0
if masks.any():
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 pred_kpt.shape[-1] == 3:
kpts_obj_loss = self.bce_pose(pred_kpt[..., 2], kpt_mask.float()) # keypoint obj loss
return kpts_loss, kpts_obj_loss
class v8ClassificationLoss:
"""Criterion class for computing training losses."""
def __call__(self, preds, batch):
"""Compute the classification loss between predictions and true labels."""
preds = preds[1] if isinstance(preds, (list, tuple)) else preds
loss = F.cross_entropy(preds, batch["cls"], reduction="mean")
loss_items = loss.detach()
return loss, loss_items
class v8OBBLoss(v8DetectionLoss):
"""Calculates losses for object detection, classification, and box distribution in rotated YOLO models."""
def __init__(self, model):
"""Initializes v8OBBLoss with model, assigner, and rotated bbox loss; note model must be de-paralleled."""
super().__init__(model)
self.assigner = RotatedTaskAlignedAssigner(topk=10, num_classes=self.nc, alpha=0.5, beta=6.0)
self.bbox_loss = RotatedBboxLoss(self.reg_max).to(self.device)
def preprocess(self, targets, batch_size, scale_tensor):
"""Preprocesses the target counts and matches with the input batch size to output a tensor."""
if targets.shape[0] == 0:
out = torch.zeros(batch_size, 0, 6, device=self.device)
else:
i = targets[:, 0] # image index
_, counts = i.unique(return_counts=True)
counts = counts.to(dtype=torch.int32)
out = torch.zeros(batch_size, counts.max(), 6, device=self.device)
for j in range(batch_size):
matches = i == j
if n := matches.sum():
bboxes = targets[matches, 2:]
bboxes[..., :4].mul_(scale_tensor)
out[j, :n] = torch.cat([targets[matches, 1:2], bboxes], dim=-1)
return out
def __call__(self, preds, batch):
"""Calculate and return the loss for the YOLO model."""
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, 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_angle = pred_angle.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"].view(-1, 5)), 1)
rw, rh = targets[:, 4] * imgsz[0].item(), targets[:, 5] * imgsz[1].item()
targets = targets[(rw >= 2) & (rh >= 2)] # filter rboxes of tiny size to stabilize training
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=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
# Pboxes
pred_bboxes = self.bbox_decode(anchor_points, pred_distri, pred_angle) # xyxy, (b, h*w, 4)
bboxes_for_assigner = pred_bboxes.clone().detach()
# Only the first four elements need to be scaled
bboxes_for_assigner[..., :4] *= stride_tensor
_, target_bboxes, target_scores, fg_mask, _ = self.assigner(
pred_scores.detach().sigmoid(),
bboxes_for_assigner.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[1] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum # BCE
# Bbox loss
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
)
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
return loss.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
def bbox_decode(self, anchor_points, pred_dist, pred_angle):
"""
Decode predicted object bounding box coordinates from anchor points and distribution.
Args:
anchor_points (torch.Tensor): Anchor points, (h*w, 2).
pred_dist (torch.Tensor): Predicted rotated distance, (bs, h*w, 4).
pred_angle (torch.Tensor): Predicted angle, (bs, h*w, 1).
Returns:
(torch.Tensor): Predicted rotated bounding boxes with angles, (bs, h*w, 5).
"""
if self.use_dfl:
b, a, c = pred_dist.shape # batch, anchors, channels
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)
class E2EDetectLoss:
"""Criterion class for computing training losses."""
def __init__(self, model):
"""Initialize E2EDetectLoss with one-to-many and one-to-one detection losses using the provided model."""
self.one2many = v8DetectionLoss(model, tal_topk=10)
self.one2one = v8DetectionLoss(model, tal_topk=1)
def __call__(self, preds, batch):
"""Calculate the sum of the loss for box, cls and dfl multiplied by batch size."""
preds = preds[1] if isinstance(preds, tuple) else preds
one2many = preds["one2many"]
loss_one2many = self.one2many(one2many, batch)
one2one = preds["one2one"]
loss_one2one = self.one2one(one2one, batch)
return loss_one2many[0] + loss_one2one[0], loss_one2many[1] + loss_one2one[1]
1.1 回归分类损失
1.1.1 DFL
全称 Distribution Focal Loss， 计算anchor point中心点到左上角和右下角的偏移量。
类定义如下：
class DFLoss(nn.Module):
"""Criterion class for computing DFL losses during training."""
def __init__(self, reg_max=16) -> None:
"""Initialize the DFL module."""
super().__init__()
self.reg_max = reg_max
def __call__(self, pred_dist, target):
"""
Return sum of left and right DFL losses.
Distribution Focal Loss (DFL) proposed in Generalized Focal Loss
https://ieeexplore.ieee.org/document/9792391
"""
target = target.clamp_(0, self.reg_max - 1 - 0.01)
tl = target.long() # target left
tr = tl + 1 # target right
wl = tr - target # weight left
wr = 1 - wl # weight right
return (
F.cross_entropy(pred_dist, tl.view(-1), reduction="none").view(tl.shape) * wl
+ F.cross_entropy(pred_dist, tr.view(-1), reduction="none").view(tl.shape) * wr
).mean(-1, keepdim=True)在初始化函数中，reg_max用于设置离散分布的最大值（默认为16），DFL损失函数就是用来处理回归问题的。它会将连续的目标值转化为离散的概率分布来优化模型。而reg_max就表示连续的目标值会被离散化到0~reg_max-1的整数区间内。
损失计算方法（call）中，target.clamp用于目标值截断，将目标值限制在[0,reg_max-1]的区间内，防止计算右整数tr超出索引值。
tl的计算得到左整数（3.7 ——> 3） tr就是在左整数基础上加一（3.7 ——> 3 ——>4）wl是左权重，表示目标值与右整数之间的距离，wr是右权重，表示目标值与左整数的距离。
在返回行中进行了交叉熵损失函数的计算，其中第一部分是左损失，计算预测分布pred_dist与tl的交叉熵，再乘以左权重wl，右损失的计算同理。其中，view(-1)将目标展平以适应交叉熵函数的输入格式，view(t;.shape)用来恢复原始形状以匹配权重矩阵。
最后.mean是用来执行加权求和与平均，左右损失相加沿最后一个维度求平均（保持维度），得到每个预测位置的DFL损失。
DFL计算公式如下：
调用DFL loss函数：
if self.use_dfl:
target_ltrb = bbox2dist(anchor_points, xywh2xyxy(target_bboxes[..., :4]), self.reg_max)
loss_dfl = self._df_loss(pred_dist[fg_mask].view(-1, self.reg_max + 1), target_ltrb[fg_mask]) * weight
loss_dfl = loss_dfl.sum() / target_scores_sum
首先调用bbox2dist将输入的坐标转换为距离中心的四个方向的距离l t r b
def bbox2dist(anchor_points, bbox, reg_max):
"""Transform bbox(xyxy) to dist(ltrb)."""
x1y1, x2y2 = bbox.chunk(2, -1)
return torch.cat((anchor_points - x1y1, x2y2 - anchor_points), -1).clamp_(0, reg_max - 0.01) # dist (lt, rb)
anchor_points - x1y1 计算的是锚点到左上角 (x1, y1) 的水平距离和垂直距离，即 l 和 t。
x2y2 - anchor_points 计算的是右下角 (x2, y2) 到锚点的水平距离和垂直距离，即 r 和 b。
如何获取网络输出的pred_dist:
首先需获得anchors和strides
self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
def make_anchors(feats, strides, grid_cell_offset=0.5):
feats：输入特征图列表，通常是从不同的特征层中提取的特征图。
strides：特征图的下采样步长列表，对应每个特征图。
grid_cell_offset：用于调整锚点位置的偏移量，默认为 0.5，即锚点位于网格单元的中心。
for i, stride in enumerate(strides):
_, _, h, w = feats[i].shape
sx = torch.arange(end=w, device=device, dtype=dtype) + grid_cell_offset # shift x
sy = torch.arange(end=h, device=device, dtype=dtype) + grid_cell_offset # shift y
sy, sx = torch.meshgrid(sy, sx, indexing="ij") if TORCH_1_10 else torch.meshgrid(sy, sx)
anchor_points.append(torch.stack((sx, sy), -1).view(-1, 2))
stride_tensor.append(torch.full((h * w, 1), stride, dtype=dtype, device=device))
return torch.cat(anchor_points), torch.cat(stride_tensor)
h 和 w 分别表示特征图的高度和宽度。
sx 和 sy 分别是水平和垂直方向上的网格坐标（加上偏移量后，中心在网格单元中心）。
torch.meshgrid 用于生成二维网格坐标（sx 和 sy），这些坐标将构成锚点。
torch.stack((sx, sy), -1).view(-1, 2) 将 sx 和 sy 坐标组合为 (x, y) 坐标对，并展平为一个二维的锚点列表。
torch.full((h * w, 1), stride, dtype=dtype, device=device) 生成一个步长张量，与锚点数量匹配。
所以他也是有初始的框的只不过单纯用于训练给定一个初始值
对box进行卷积得到分布信息：
class DFL(nn.Module):
def __init__(self, c1=16):
super().__init__()
self.conv = nn.Conv2d(c1, 1, 1, bias=False).requires_grad_(False)
x = torch.arange(c1, dtype=torch.float)
self.conv.weight.data[:] = nn.Parameter(x.view(1, c1, 1, 1))
self.c1 = c1
def forward(self, x):
"""Applies a transformer layer on input tensor 'x' and returns a tensor."""
b, _, a = x.shape # batch, channels, anchors
return self.conv(x.view(b, 4, self.c1, a).transpose(2, 1).softmax(1)).view(b, 4, a)
# return self.conv(x.view(b, self.c1, 4, a).softmax(1)).view(b, 4, a)
conv不进行训练，权重是默认的，self.conv.weight.data[:] = nn.Parameter(x.view(1, c1, 1, 1))，即[0, 1, 2, ..., c1-1]
forward:
这里将输入张量 x 重新调整形状为 (batch_size, 4, c1, num_anchors)。4 表示每个锚点的4个回归值（通常是 l, t, r, b，即到边界的距离）。
transpose(2, 1)：交换维度，将张量变为 (batch_size, c1, 4, num_anchors) 的形状。这使得通道维度 c1 排在第二位。
.softmax(1)：对 c1 维度（即原来的通道维度）应用 softmax 操作。softmax 将每个类别的预测转换为概率分布，这在 DFL 中用于对每个边界框的预测进行更加细粒度的调整。
self.conv(...)：使用 1x1 卷积对经过 softmax 的输出进行处理，实际上是对 softmax 结果的加权平均。由于卷积层的权重被初始化为 0 到 c1-1 的线性值，这一步相当于计算 softmax 结果的期望值，输出的每个通道的值可以被解释为最终预测的偏移量。
view(b, 4, a)：将最终的输出张量调整回形状 (batch_size, 4, num_anchors)，即每个锚点有4个回归值。
1.1.2 IoU Loss
定位损失采用CIoU loss，只计算正样本的定位损失。
target_bboxes /= stride_tensor
loss[0], loss[2] = self.bbox_loss(
pred_distri, pred_bboxes, anchor_points, target_bboxes, target_scores, target_scores_sum, fg_mask
)
1.2 分类损失
BCE loss，只计算正样本的分类损失。
loss[1] = self.bce(pred_scores, target_scores.to(dtype)).sum() / target_scores_sum
分类和定位权衡问题中，可以在ultralytics/utils/tal.py中修改其规则：
其中s和u分别代表分类得分和iou值，α和β用于控制锚点对其指标中两个任务（分类和定位）的影响。
align_metric = bbox_scores.pow(self.alpha) * overlaps.pow(self.beta)
默认值：
alpha:1.0
beta:6.0
分配规则选择训练样本： 对于每个实例，选择m个具有最大t值的锚点作为正样本，而将剩余的锚点作为负样本。
1.3 CIOU loss
调用loss方法：
"""IoU loss."""
weight = target_scores.sum(-1)[fg_mask].unsqueeze(-1)
iou = bbox_iou(pred_bboxes[fg_mask], target_bboxes[fg_mask], xywh=False, CIoU=True)
loss_iou = ((1.0 - iou) * weight).sum() / target_scores_sum
fg_mask可以理解为区分前景和背景，其值为布尔类型的true和false。
pred_bboxes[fg_mask]会将batch维度去掉，例如B，N1，5转变成N2，5.传入一个box数组，将一个batch所有的框进行整合。
需要修改iou计算方法时要直到输入，v8并没有batch这个维度。
iou输出是一个列向量，一旦修改计算方法返回的就是行向量，有可能导致计算错误。
bbox_iou:
def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
box1: 表示一个边界框，形状为(1, 4)的Tensor。
box2: 表示n个边界框，形状为(n, 4)的Tensor。
xywh: 如果为True，表示输入的框格式为(x, y, w, h)（中心点坐标和宽高）；如果为False，则输入格式为(x1, y1, x2, y2)（左上角和右下角坐标）。
GIoU, DIoU, CIoU: 控制是否计算相应的IoU扩展版本。
eps: 一个小值，用于避免除零错误。
IoU就是用来计算预测框与真实框的交并比。
1、计算得到交集h和w相乘
inter = (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp_(0) * (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1) ).clamp_(0)
2、计算得到并集
union = w1 * h1 + w2 * h2 - inter + eps
3、计算IOU（交集/并集）
iou = inter / union4、其中YOLO使用的CIOU是IOU基础上进行的计算，IOU部分相同。
首先还是计算IOU，接着计算包围两个边界框的最小矩阵s的h和w。
cw = b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1) # convex (smallest enclosing box) width
ch = b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1) # convex height
之后计算s的对角线的平方c2，两个边界框中心点之间距离的评分rho2
c2 = cw.pow(2) + ch.pow(2) + eps # convex diagonal squared
rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2).pow(2) + (b2_y1 + b2_y2 - b1_y1 - b1_y2).pow(2)) / 4 # center dist**2
rho2推理过程：
中心点cx:cx = (x1 + x2) / 2
中心点cy:cy = (y1 + y2) / 2
两个框中心点分别是 (cx1, cy1) 和 (cx2, cy2)距离为：根号下(cx2-cx1)^2 + (cy2-cy1)^2
距离的平方：cho2 = (cx2-cx1)^2 + (cy2-cy1)^2
展开得：((x1 + x2) / 2-(x1_2 + x2_2) / 2)^2...
整理得：1/4 * ((x1 + x2) -(x1_2 + x2_2))与代码一致
最后计算一个与边界框宽高比相关的v，根据v计算权重alpha得到最终CIOU：
v = (4 / math.pi**2) * ((w2 / h2).atan() - (w1 / h1).atan()).pow(2)
with torch.no_grad():
alpha = v / (v - iou + (1 + eps))
return iou - (rho2 / c2 + v * alpha) # CIoU
这就完成了IOU loss的计算。