search

google-research / uda

Unsupervised Data Augmentation (UDA)
https://arxiv.org/abs/1904.12848
Apache License 2.0
2.18k stars 312 forks source link

A question about augmentation in labeled data #62

Open heartInsert opened 5 years ago

heartInsert commented 5 years ago

Sorry for my noob question . I noticed in the train graph you decribe in the paper's Figure 1 . The labeled img didn't transfomed by Randaugment. I think maybe it is because the lacking of labeled data , we have to matain its fidelity . But now , I have at about 9k labeled data and 7k unlabeled data , and there is also very little gap between labeled and unlabeled , I want to confirm wherther it will cause side effect if I use Randaugment in labeled data . Thanks.

michaelpulsewidth commented 5 years ago

I can't say for sure. I haven't tried similar settings. It's worth a try.

heartInsert commented 5 years ago

Thanks ,I will try them both later. There are two more questions . 1: In your paper section 3.4 SCALABILITY TEST ON THE IMAGENET DATASET , when use UDA in full ImageNet ,you selected 1.3M images from JFT .Sorry for English is not my mother tongue , so is 1.3M stands for 1.3e+6 ? 2: I think I have read your paper carefully , but you didn't mention it that , when use UDA in full ImageNet , if the transfomations for supervised data is also just simple augmentation (with cropping and flipping ) like in cifar-10 ? Thanks.

michaelpulsewidth commented 5 years ago
  1. Sorry for the confusion. 1.3M means 1.3 Million
  2. Right.