search

icoz69 / DeepEMD

Code for paper "DeepEMD: Few-Shot Image Classification with Differentiable Earth Mover's Distance and Structured Classifiers", CVPR2020
MIT License
577 stars 82 forks source link

Configs for training 5-shot models #20

Closed LouieYang closed 4 years ago

LouieYang commented 4 years ago

Hello, thanks for your great work first.

I want to reproduce some results on 5-shot tasks but found the training process quite slow compared to 1-shot tasks (e.g. 10-20 seconds/iterations), is it normal? or could you share the training configs for 5-shot tasks.

btw, I also have questions about how you improve your numbers on miniimagenet 5-way 1-shot task from CVPR version (65.91) to arxiv version? any suggestion is appreciated.

Thanks Yang

icoz69 commented 4 years ago

Hi Yang Actually, we don't train a 5-shot model. We directly use the 1-shot model to test.

I don't quite understand your 2nd question. The difference between two versions is DeepEMD-Grid and DeepEMD-Sampling.

LouieYang commented 4 years ago

In CVPR version, the results of DeepEMD-FCN for 1-shot tasks on miniimagenet is 65.91 in Table 4(a) while in DeepEMD v2 the results for DeepEMD-FCM becomes 66.50 in Table 6(a). So I wonder where the improvement comes from?

icoz69 commented 4 years ago

Hi, the improvements come from the pyramid structure. You may check the details in the paper.

LouieYang commented 4 years ago

ok, I see the differences

Many thanks.