search

Sha-Lab / FEAT

The code repository for "Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions"
MIT License
418 stars 84 forks source link

a question about resnet-18/12 kernel size #34

Closed Sword-keeper closed 4 years ago

Sword-keeper commented 4 years ago

Hi that's really a nice code! However, i have a question about resnet. In the resnet paper, resnet's first conv layer's kernel size is 7. But in your code , I found the first conv kernel size is 3. I also found other codes which use resnet-12 or 18. They also set the kernel size is 3.
What's more, I read the TADAM's paper which design the resnet-12(maybe not). It just use blocks, didn't add conv layer before the blocks. Could you tell me why you change conv kernel size ? Thank you !

Han-Jia commented 4 years ago

We follow the ResNet-12 of MetaOptNet and ResNet-18 of SimpleShot (results with ResNet-18 are not reported in the paper).

The main reason is that the images used in FSL experiments are small images (84x84), so we should remove some downsampling operations and do not need a large kernel size.