search

WongKinYiu / PyTorch_YOLOv4

PyTorch implementation of YOLOv4
1.87k stars 585 forks source link

我使用yolov4-tiny训练voc2007+2012 ，在darknet上的结果是map0.5 0.70以上 但是在此代码上训练得到的结果是0.60 结果为啥差距这么大 #133

Open mr-xiao-zuo opened 3 years ago

WongKinYiu commented 3 years ago

您用的是哪個branch?

mr-xiao-zuo commented 3 years ago

您用的是哪個branch?

Aug 10, 2020 用的这个版本里的yolov4-tiny.yaml训练的

mr-xiao-zuo commented 3 years ago

您用的是哪個branch?

是我的参数设置的有问题 我一直不太明白差距这么大。。

WongKinYiu commented 3 years ago

branch指的是https://github.com/WongKinYiu/PyTorch_YOLOv4/branches/all

Alex-afka commented 3 years ago

branch指的是https://github.com/WongKinYiu/PyTorch_YOLOv4/branches/all

u5_mish_preview 这个branch

WongKinYiu commented 3 years ago

u5 branch and u5_mish_preview branch do not support yolov4-tiny. and it better to use u3_preview branch for training yolov4-tiny. in my experiments, u5 branch series get worse performance on almost all of small/tiny models than u3 branch.

Alex-afka commented 3 years ago

u5 branch and u5_mish_preview branch do not support yolov4-tiny. and it better to use u3_preview branch for training yolov4-tiny. in my experiments, u5 branch series get worse performance on almost all of small/tiny models than u3 branch.

如果我在 u3_preview这个branch训练voc2007+2012，我需要如何map0.5达到0.7以上，难道也要使用预训练权重吗，从头训练不使用预训练权重效果是不是达不到0.7

WongKinYiu commented 3 years ago

in my experiments. voc: baseline voc with coco pretrained: +3% AP voc transfer from coco: +3.2% AP

mr-xiao-zuo commented 3 years ago

您用的是哪個branch?

in my experiments. voc: baseline voc with coco pretrained: +3% AP voc transfer from coco: +3.2% AP

有coco预训练权重吗？具体设置的参数大小是什么呢？比如batch_size，epoch，anchor，lr等您设置的是多少呢？另外您最后在VOC2007+2012上的mAP0.5是多少呢？

WongKinYiu commented 3 years ago

我的結果是tiny-3l的結果, 在log檔裡512x512分別是70.5, 73.5, 73.6. 運行tes.pyt會比這個結果略高, 不過目前沒有空的gpu能測.

若要用u5訓練voc的話, 可以嘗試以下hyper parameter.

lr0: 0.0032
lrf: 0.12
momentum: 0.843
weight_decay: 0.00036
warmup_epochs: 2.0
warmup_momentum: 0.5
warmup_bias_lr: 0.05
box: 0.0296
cls: 0.243
cls_pw: 0.631
obj: 0.301
obj_pw: 0.911
iou_t: 0.2
anchor_t: 2.91
# anchors: 3.63
fl_gamma: 0.0
hsv_h: 0.0138
hsv_s: 0.664
hsv_v: 0.464
degrees: 0.373
translate: 0.245
scale: 0.898
shear: 0.602
perspective: 0.0
flipud: 0.00856
fliplr: 0.5
mosaic: 1.0
mixup: 0.243
mr-xiao-zuo commented 3 years ago

我的結果是tiny-3l的結果, 在log檔裡512x512分別是70.5, 73.5, 73.6. 運行tes.pyt會比這個結果略高, 不過目前沒有空的gpu能測.

若要用u5訓練voc的話, 可以嘗試以下hyper parameter.

lr0: 0.0032
lrf: 0.12
momentum: 0.843
weight_decay: 0.00036
warmup_epochs: 2.0
warmup_momentum: 0.5
warmup_bias_lr: 0.05
box: 0.0296
cls: 0.243
cls_pw: 0.631
obj: 0.301
obj_pw: 0.911
iou_t: 0.2
anchor_t: 2.91
# anchors: 3.63
fl_gamma: 0.0
hsv_h: 0.0138
hsv_s: 0.664
hsv_v: 0.464
degrees: 0.373
translate: 0.245
scale: 0.898
shear: 0.602
perspective: 0.0
flipud: 0.00856
fliplr: 0.5
mosaic: 1.0
mixup: 0.243

您跑的最好VOC效果的原版代码能给我发一份吗？因为现在网上的源码我都没有得到很理想的结果

WongKinYiu commented 3 years ago

https://github.com/WongKinYiu/PyTorch_YOLOv4/tree/u3_preview

mr-xiao-zuo commented 3 years ago

https://github.com/WongKinYiu/PyTorch_YOLOv4/tree/u3_preview

您在这个源码中使用的coco预训练权重能发我一个吗？

WongKinYiu commented 3 years ago

https://drive.google.com/file/d/1aQKcCvTAl1uOWzzHVE9Z8Ixgikc3AuYQ/view

mr-xiao-zuo commented 3 years ago

https://drive.google.com/file/d/1aQKcCvTAl1uOWzzHVE9Z8Ixgikc3AuYQ/view

非常感谢！

Alex-afka commented 3 years ago

https://drive.google.com/file/d/1aQKcCvTAl1uOWzzHVE9Z8Ixgikc3AuYQ/view

i train voc2007+2012 test是voc2007测试集 batchasize=16 epoch是400 最后结果为map为0.65 是我的参数设置的不对吗 您能把您设置的参数分享给我吗 谢谢

xiaozhangya commented 3 years ago

https://drive.google.com/file/d/1aQKcCvTAl1uOWzzHVE9Z8Ixgikc3AuYQ/view

非常感谢！

你好！你试过u3版本的yolov4吗？他好像不支持重新加载模型继续训练。有点坑

WENKONG01 commented 3 years ago

https://drive.google.com/file/d/1aQKcCvTAl1uOWzzHVE9Z8Ixgikc3AuYQ/view

非常感谢！

你好！你试过u3版本的yolov4吗？他好像不支持重新加载模型继续训练。有点坑

您好，这个链接我下载不了，您可以不可以给我分享一下，邮箱wenwenkong0902@163.com,谢谢