haofusheng / saml Goto Github PK

我运行你的代码时出现了一些问题：

1. task_generator_test.py里缺少了transforms.Resize((224,224)),导致读取的图片尺寸不一样而报错。
2. main.py里整个main函数从头到尾,完全不分函数写，不注重模块化。train dataset和test dataset数据处理是一样的，但是代码完全复制粘贴，代码顺序却还不一样。
3. 最后希望作者能借鉴他人代码的写法，注重模块化，划分函数编写代码。
   https://github.com/WenbinLee/DN4

I try to adapt the model to the general setting of few-shot learning, which adopts ResNet-12 as backbone and image size 84*84. I get the 54% 5-way-1-shot acc without pre-training and 56% 5-way-1-shot acc with pre-training, following the protocol of Meta-Baseline[1]. Such results are disappointing. So, is there any suggestion?

[1] Chen, Y., Wang, X., Liu, Z., Xu, H., & Darrell, T. (2020). A new meta-baseline for few-shot learning. arXiv preprint arXiv:2003.04390.

I didn't change all the factors in your code, just fixed bugs that can not affect on results and I tested your code on CUB / miniImagenet datasets for image size 84 X 84 and 224 X 224 both.
However, I could not reach the results in your paper and the difference was quite lot.
Could you explain why this happens?