Classifier - CutMix: Regularization Strategy to Train Strong Classifiers

[AlexeyAB]

CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features

• paper: https://arxiv.org/abs/1905.04899v2
• source: https://github.com/clovaai/CutMix-PyTorch

Use for Classifier training:

• cutmix=1 - will be used CutMix (36%-91%, 9%-64%)
• mixup=1 - will be used MixUp (50%, 50%)

Run training with flag -show_imgs to see how images are changed (in separate windows and saved to files aug_... .jpg) and how labels are changed (see the console).

./darknet classifier train cfg/imagenet1k_c.data cfg/efficientnet_b0.cfg -topk -show_imgs

---

---

Results of ResNet50-Faster-RCNN: clovaai/CutMix-PyTorch#13

Backbone	mAP
Baseline	33.27
Mixup-Pretrained	34.18
Cutout-Pretrained	34.28
CutMix-Pretrained	35.16

CutMix-Pretrained shows a meaningful improvement in MS-COCO detection!

[AlexeyAB]

@WongKinYiu

Fixed CutMix: 4c26117
removed boundary-aware constraint: clovaai/CutMix-PyTorch#14

No, our code does not sample the cropping region with any boundary-aware constraint.

Use -show_imgs flag to see how does it work.
darknet.exe classifier train cfg/imagenet1k_c.data cfg/efficientnet_b0.cfg -topk -show_imgs

[WongKinYiu]

@AlexeyAB OK, I will re-train the models.

[WongKinYiu]

@AlexeyAB cutmix=1 gets Segmentation fault (core dump) in latest repo.

[AlexeyAB]

@WongKinYiu After how many iterations it leads to Segmentation fault?
Did you use only cutmix=1 or together with mixup=1 mosaic=1?

[WongKinYiu]

@AlexeyAB 0~2 iterations, use together.

[AlexeyAB]

@WongKinYiu
I tried cutmix=1 and cutmix=1 mosaic=1 but I can't reproduce this bug.

• Did you use the latest version of this repo https://github.com/AlexeyAB/darknet or did you move CutMix/Mosaic data augmentation features to your repo?
• Also check bad.list file

[WongKinYiu]

I used yesterday's repo and trained a detector.

[AlexeyAB]

@WongKinYiu Detector doesn't support CutMix yet:
I added assert yesterday: efc5478#diff-2ceac7e68fdac00b370188285ab286f7R891

[WongKinYiu]

@AlexeyAB OK, thank you!

[WongKinYiu]

@AlexeyAB

mixup does not get better performance in my experiments but cutmix does, for example:

Model	Augmentation	Top-1	Top-5
CSPResNeXt-50	Normal	77.9	94.0
CSPResNeXt-50	Mixup	77.2	94.0
CSPResNeXt-50	Cutmix	78.0	94.3
CSPResNeXt-50	Cutmix+Mixup	77.7	94.4

[AlexeyAB]

@WongKinYiu Thanks! Did you test mosaic and cutmix+mosaic?

[WongKinYiu]

@AlexeyAB

Not yet, I just get some available gpus today.

I will test mosaic=1 and blur=1 first.