稍微总结一下常见的训练时和验证时数据增强(Test time augmentation)技术。
https://discuss.pytorch.org/t/test-time-augmentation/8828
A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012.
- 模型:
- AlexNet
- 训练:
- scale(256)
- random crop(224)
- horizontal flips
- PCA (reduces the top-1 error rate by over 1%)
- 测试:
- 10 crop(224)
- average the softmax output
K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015.
- 模型:
- VGG
- 训练:
- batch_size 256, momentum 0.9
- weight decay 5e-4
- dropout