Question about the dimension of set encoder

[weijiandeng]

Hi, thanks for sharing the code. It looks good!

I might miss something, so I am a little confused about the dimension in the set encoder:

MetaD2A/MetaD2A_mobilenetV3/generator/generator_model.py

Line 254 in 6023a5b

for x in X: # X.shape: [32, 400, 512]

Take ImageNet32 as an example, for each image, its size is 3x32x32. For each class, the code samples "N=self.num_sample" images.

Then, why reshape x to be x.view(-1, self.num_sample, 512)).squeeze(1).

MetaD2A/MetaD2A_mobilenetV3/generator/generator_model.py

Line 256 in 6023a5b

x.view(-1, self.num_sample, 512)).squeeze(1)

What is the meaning of 512? In my review, I think the code reshapes the x to be "num_class, self.num_sample (for each class), features". But I did not know the meaning of 512. Did the code extract the feature for each image (e.g., for each image, 512 is its feature size--> its feature is a 512-d vector)? However, I did not find the code to extract features.

Could you please help me address this question? Thanks in advance!

[seanie12]

During preprocessing, as written in appendix, they use pre-trained resnet 18 to extract 512 dimensional feature vector from a 32 x 32 image. So 512 is the dimension of the extracted feature. @HayeonLee