search

baiksung / MeTAL

Official PyTorch implementation of "Meta-Learning with Task-Adaptive Loss Function for Few-Shot Learning" (ICCV2021 Oral)
62 stars 13 forks source link

请问class StepLossAdapter(nn.Module)类的forward #1

Open 9876K opened 2 years ago

9876K commented 2 years ago

生成的仿射变换参数为什么要分割？ 还有仿射变换参数在后面又乘了可学习参数 self.multiplier_bias和 self.offset_bias，这是为什么呢？

baiksung commented 2 years ago

Hi,

Thank you for your interests in my paper. But, I'm sorry to tell you that I don't speak Chinese. It would be great if you post your questions in English, and I'd be happy to help you.

Best, Sungyong

Zlixu commented 2 years ago

Hi,

Your idea is very interesting and inspire me. I use your code ,dataset(mini imagenet) and original experiment config (5shot and 5 way) to run, and get high train acc (more than 90%) and val acc(more than 90%) in logs. However the test acc is no more than 53%.So I don't know why. Could you help me? Look forward to your rely.

Best， Zhang Lixu

baiksung commented 2 years ago

Hi Zhang Lixu,

I don't think you should get train acc & val acc more than 90% with this code. What is the exact command you executed? Did you, by any chance, use multi GPUs? If so, this code does not fully support multi-gpu settings and may cause bugs, such as yours. I would recommend running on a single GPU to reproduce results.

Best, Sungyong