10月4日 炼丹

罗宇鑫

日志记录： 10月4日下午 13:00 代号：001

:::info 数据：
cls 调整到0.04
命令行：
python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —data dataSet/clothes.yaml —hyp data/hyps/hyp.clothe014.yaml —batch-size 36 —epochs 70 —device 4 :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 1.0 # cls loss gain
cls: 0.04 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.5 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录： 10月4日下午 13:00 代号：002

:::info 数据：
cls 调整到0.25
命令行：
python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —data dataSet/clothes.yaml —hyp data/hyps/hyp.clothe015.yaml —batch-size 36 —epochs 70 —device 2 :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.1 # cls loss gain
cls: 0.25 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.5 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录： 10月4日下午 15:00 代号：003

:::info 数据：
anchor_t 调整到3
命令行：
python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —data dataSet/clothes.yaml —hyp data/hyps/hyp.clothe016.yaml —batch-size 24 —epochs 70 —device 1 ::: �
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
cls: 0.25 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.5 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
#anchor_t: 4 # anchor-multiple threshold
anchor_t: 3 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)
�

日志记录： 10月4日下午 15:00 代号：004

:::info 数据：
anchor_t 调整到 2.5
命令行：
python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —data dataSet/clothes.yaml —hyp data/hyps/hyp.clothe018.yaml —batch-size 24 —epochs 70 —device 3 ::: �lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.1 # cls loss gain
cls: 0.25 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.5 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
#anchor_t: 4 # anchor-multiple threshold
anchor_t: 2.5 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)
�

李志颖

日志记录： 10月4日上午11时 代号：001

:::info 数据：
obj 从0.5调整到0.8
命令行：
python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —hyp data/hyps/hyp.clothes34.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 4,6 —adam :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.5 # cls loss gain
cls: 0.3 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.8 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录：10月4日上午12时 代号：002

:::info 数据：obj从0.8调整到0.45
命令行：
python train.py —weights weights/yolov5s6.pt —cfg models/yolov5s6-bifpn_atten.yaml —hyp da️ta/hyps/hyp.clothes35.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 2,4 —adam :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.5 # cls loss gain
cls: 0.3 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
obj: 0.45 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录：10月4日上午15时 代号：003

:::info 数据：obj从0.45调整到了0.4
命令行：python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —hyp dat️a/hyps/hyp.clothes36.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 2,4 —adam :::
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.5 # cls loss gain
cls: 0.3 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
#obj: 0.45 # obj loss gain (scale with pixels)
obj: 0.40 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

:::tips 实验004-实验005小结：
004采用了adam优化器，005采用的是SGD优化器+Momentum，效果上能看到差距，虽说Adam优化器是在SGD基础上进行整合的，但现在很多模型所沿用的还是SGD优化器 :::

日志记录：10月4日上午16时 代号：004

:::info 数据：obj从0.4调整到了0.55
命令行：python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —hyp dat️️a/hyps/hyp.clothes37.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 2,4 —adam :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.5 # cls loss gain
cls: 0.3 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
#obj: 0.45 # obj loss gain (scale with pixels)
#obj: 0.40 # obj loss gain (scale with pixels)
obj: 0.55 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录：10月4日上午16时 代号：005

:::info 数据：obj从0.4调整到了0.55
命令行：python train.py —weights weights/yolov5s.pt —cfg models/yolov5s6-bifpn_atten.yaml —hyp dat️️a/hyps/hyp.clothes37.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 2,4 :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.05 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.900 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.055 # warmup initial bias lr
box: 0.05 # box loss gain
#cls: 0.5 # cls loss gain
cls: 0.3 # cls loss gain
cls_pw: 1.0 # cls BCELoss positive_weight
#obj: 0.45 # obj loss gain (scale with pixels)
#obj: 0.40 # obj loss gain (scale with pixels)
obj: 0.55 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positive_weight
iou_t: 0.30 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)

日志记录：10月4日上午22时 代号：006

:::info 数据：采用的是yolov5s6.pt权重文件，跟yolov5.pt权重文件差距很多，数据量的差距很大，超参数文件，采用的是原始的hyp.scratch.yaml
命令行：python train.py —weights weights/yolov5s6.pt —cfg models/yolov5s6-bifpnatten.yaml —hyp data/hyps/hyp.scratch.yaml —data dataSet/clothes.yaml —batch-size 16 —epochs 70 —device 2,4 :::
lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
lrf: 0.2 # final OneCycleLR learning rate (lr0 * lrf)
momentum: 0.937 # SGD momentum/Adam beta1
weight_decay: 0.0005 # optimizer weight decay 5e-4
warmup_epochs: 3.0 # warmup epochs (fractions ok)
warmup_momentum: 0.8 # warmup initial momentum
warmup_bias_lr: 0.1 # warmup initial bias lr
box: 0.05 # box loss gain
cls: 0.5 # cls loss gain
cls_pw: 1.0 # cls BCELoss positiveweight
_obj: 1.0 # obj loss gain (scale with pixels)
obj_pw: 1.0 # obj BCELoss positiveweight
_iou_t: 0.20 # IoU training threshold
anchor_t: 4.0 # anchor-multiple threshold
# anchors: 3 # anchors per output layer (0 to ignore)
fl_eiou_gamma: 0.0 #focal eiou loss gamma
iou_aware: 0.0
fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
hsv_h: 0.015 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4 # image HSV-Value augmentation (fraction)
degrees: 0.0 # image rotation (+/- deg)
translate: 0.1 # image translation (+/- fraction)
scale: 0.5 # image scale (+/- gain)
shear: 0.0 # image shear (+/- deg)
perspective: 0.0 # image perspective (+/- fraction), range 0-0.001
flipud: 0.0 # image flip up-down (probability)
fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)_