search

bcmi / F2GAN-Few-Shot-Image-Generation

Fusing-and-Filling GAN (F2GAN) for few-shot image generation, ACM MM2020
78 stars 11 forks source link

Question about the dataset setting #6

Closed edward3862 closed 3 years ago

edward3862 commented 3 years ago

Hello, thanks for this repo. The dataset split seems to differ from the description in the paper.

In the paper:

For EMNIST, following MatchingGAN[16], we randomly select 28 categories from total 48 categories as training seen categories and select 10 categories from remaining categories as unseen testing categories

which suppose be 28:10:10 for train, val and test.

In the code:

x_train, x_val, x_test = self.x[:35], self.x[35:42], self.x[42:]

which is 35:7:5 for train, val and test.

Which is the setting to re-implement the result in the paper?

hy-zpg commented 3 years ago

The setting of dataset split should be as described in the paper, and I have updated data_with_matchingclassifier.py.