Bag of Tricks for Image Classification with Convolutional Neural Networks

论文地址

resnet50+CUB_200_2011

• 1: 未使用预训练权重，使用kaiming初始化，训练450 epochs，Top1-AC: 47.64%
• 2: 使用了ImageNet预训练权重，100 epochs，Top1-AC：66.45%
• 3: 将ResNet改为了ResNet-B(论文中有说明)的形式，100 epochs，Top1-AC：72.82%(使用预训练权重)
• 4: 将ResNet改为了ResNet-C(论文中有说明)的形式，100 epcohs，Top1-AC：55.99%
• 5: 将ResNet改为了ResNet-C(论文中有说明)的形式，450 epochs，Top1-AC：20.95%(未使用预训练权重)
• 6: ResNet-B+ColorJitter+centerCrop(test)，100epochs，Top1-AC：76.91%
• 7: ResNet-B+ColorJitter+Resize(test), 100epochs, Top1-AC：69.8%
• 8: ResNet-B+centerCrop(test), 100 epochs, Top1-AC:79.06%
• 9: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std, Top1-AC:79.1%
• 10: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+NAG, Top1-AC: 79.32%
• 11: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+NAG+warm up(5), Top1-AC: 78.75%
• 12: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+warm up(5)+cosine, Top1-AC: 79.06%
• 13: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+cosine, Top1-AC: 78.91%
• 14: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+cosine+LabelSmooth(0.1), Top1-AC: 81.43%
• 15: ResNet-B+centerCrop(test), 100 epcohs, ImageNet的mean和std+NAG+LabelSmooth(0.1), Top1-AC: 80.69%
• 16: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+cosine+warm up(5)+LabelSmooth(0.1), Top1-AC: 81.81%
• 17: ResNet-B+centerCrop(test), 100 epochs, ImageNet的mean和std+cosine+warm up(5)+labelSmooth(0.1)+mixup, Top1-AC: 81.34%
• 18: ResNet-B+centerCrop(test), 175 epochs, ImageNet的mean和std+cosine+warm up(5)+labelSmooth(0.1)+mixup, Top1-AC: 82.27%

ResNet结构(预训练权重)

Model	Top1-AC
ResNet(init)	66.45%
ResNet-B	72.82%
ResNet-C	55.99%

ResNet-B+tricks

model	epochs	ColorJitter	NAG	warm up(5)	cosine	LabelSmooth(0.1)	mixup	Top1-AC
1	100	×	×	×	×	×	×	79.1%
2	100	√	×	×	×	×	×	76.91%
3	100	×	√	×	×	×	×	79.32%
4	100	×	√	√	×	×	×	78.75%
5	100	×	×	√	√	×	×	79.06%
6	100	×	√	×	×	√	×	80.69%
7	100	×	×	×	√	√	×	81.43%
8	100	×	×	√	√	√	×	81.81%
9	175	×	×	√	√	√	√	82.27%