이제 드디어 YOLO 를 분석할 시기가 온듯..
YOLOWorld
라니..task_map
architecturemodel, trainer, validator, predictor 를 task 별로 정해서 가져온다.
"classify": {
"model": ClassificationModel,
"trainer": yolo.classify.ClassificationTrainer,
"validator": yolo.classify.ClassificationValidator,
"predictor": yolo.classify.ClassificationPredictor,
},
"detect": {
"model": DetectionModel,
"trainer": yolo.detect.DetectionTrainer,
"validator": yolo.detect.DetectionValidator,
"predictor": yolo.detect.DetectionPredictor,
},
"segment": {
"model": SegmentationModel,
"trainer": yolo.segment.SegmentationTrainer,
"validator": yolo.segment.SegmentationValidator,
"predictor": yolo.segment.SegmentationPredictor,
},
"pose": {
"model": PoseModel,
"trainer": yolo.pose.PoseTrainer,
"validator": yolo.pose.PoseValidator,
"predictor": yolo.pose.PosePredictor,
},
"obb": {
"model": OBBModel,
"trainer": yolo.obb.OBBTrainer,
"validator": yolo.obb.OBBValidator,
"predictor": yolo.obb.OBBPredictor,
},
model = YOLO("yolov8x.pt")
mapping architecture
yolov8.yaml
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# Parameters
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.33, 0.25, 1024] # YOLOv8n summary: 225 layers, 3157200 parameters, 3157184 gradients, 8.9 GFLOPs
s: [0.33, 0.50, 1024] # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients, 28.8 GFLOPs
m: [0.67, 0.75, 768] # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients, 79.3 GFLOPs
l: [1.00, 1.00, 512] # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
x: [1.00, 1.25, 512] # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs
# YOLOv8.0n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
- [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
- [-1, 3, C2f, [128, True]]
- [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
- [-1, 6, C2f, [256, True]]
- [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
- [-1, 6, C2f, [512, True]]
- [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
- [-1, 3, C2f, [1024, True]]
- [-1, 1, SPPF, [1024, 5]] # 9
# YOLOv8.0n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 6], 1, Concat, [1]] # cat backbone P4
- [-1, 3, C2f, [512]] # 12
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 4], 1, Concat, [1]] # cat backbone P3
- [-1, 3, C2f, [256]] # 15 (P3/8-small)
- [-1, 1, Conv, [256, 3, 2]]
- [[-1, 12], 1, Concat, [1]] # cat head P4
- [-1, 3, C2f, [512]] # 18 (P4/16-medium)
- [-1, 1, Conv, [512, 3, 2]]
- [[-1, 9], 1, Concat, [1]] # cat head P5
- [-1, 3, C2f, [1024]] # 21 (P5/32-large)
- [[15, 18, 21], 1, Detect, [nc]] # Detect(P3, P4, P5)
Segmentation, Pose, OBB model 은 DetectionModel
을 상속받는다.
Classification 과 Detection 모델은 BaseModel
을 상속 받고,
load model 은 parse_model
이 메인 인 것 같다.
non-linear equation 을 back propagation 할 때,
loss[0], loss[2]
: bbox loss (iou = box loss, dfl loss)# IoU
iou <= inter / union
loss[1]
: cls loss (varifocal_loss -> BCEWithLogitsLoss)True / False
로 구분하기 위한 loss.return loss.sum() * batch_size, loss.detach() # loss(box, cls, dfl)
class TaskAlignedAssigner(nn.Module):
class TaskAlignedAssigner(nn.Module):
...
def forward(self,
pd_scores, pd_bboxes, anc_points, gt_labels, gt_bboxes, mask_gt):
...
return target_labels,
target_bboxes,
target_scores,
fg_mask.bool(),
target_gt_idx
using assigner
... 와_, target_bboxes, target_scores, fg_mask, _ = 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,
)
loss[2]
: BCEWithLogitsLoss 로 cls loss. loss[0], loss[3]
: bbox loss ( iou = box loss , dfl loss )loss[1]
: segmentation loss ( anchor box )pose loss 라니.. 설레이는데..
loss[3]
: BCEWithLogitsLoss 로 cls loss.loss[0], loss[4]
: bbox loss (iou = box loss, dfl loss )loss[1], loss[2]
: keypoints loss ( keypoints, keypoints object )Oriented Bounding Box
loss[1]
: cls loss (varifocal_loss -> BCEWithLogitsLoss)loss[0], loss[2]
: bbox loss (iou = box loss, dfl loss) 이 코드 구조, 내가 원했던 구조. 내 코드의 간략 리뷰를 하자면..
이젠 너무 task 에 대한 weight 이 커져서, scratch 학습은 사실 ImageNet 1k
정도가 maximum 이지 않을 까 한다. ImageNet 21k
를 해봤지만.. 그 비용과 효율보다는 weight 을 가져다 사용하면서, 필요한 새로운 데이터를 튜닝하는것이 ..