search

DataDistillation / DataDAM

[ICCV 2023] DataDAM: Efficient Dataset Distillation with Attention Matching
28 stars 4 forks source link

Reproduction issue to the code #8

Closed gihyunkim closed 6 months ago

gihyunkim commented 6 months ago

Hi, I am very interested in dataset condensation and your paper. I have tried to reproduce the performance of your code, however, I found that the code has weird point on "get_network" function.

You set network setting from "get_default_convnet_setting" function but fix net_norm as "batch_norm" for ConvNet for some reason.

I ran the code with "instance_norm" as DM and couldn't get high performance (about 63%) as your paper.

I'd appreciate if you explain this situation and what is my problem for reproduction.

Thank you.

DataDistillation commented 6 months ago

Hello, Thank you for your interest in our work. As in the code, please use batch-based normalization as it provides stable results.