search

google-research / uda

Unsupervised Data Augmentation (UDA)
https://arxiv.org/abs/1904.12848
Apache License 2.0
2.17k stars 313 forks source link

code to reproduce the ImageNet results #25

Open chaithu536 opened 5 years ago

chaithu536 commented 5 years ago

I enjoyed reading the paper and very promising results. Are you planning to share the code of ImageNet experiments in semi-supervised setting ? It would be a great help to reproduce the results.

michaelpulsewidth commented 5 years ago

Thanks! We don't plan to release the ImageNet code because it requires a lot of code cleaning. The hyperparameters for ImageNet experiments are available here. The augmentation policy is available here.

chaithu536 commented 5 years ago

Thanks for the links. Is it possible to share the ImageNet trained model or weights ? How did you choose the 10% labelled data from ImageNet ? Did you pick random images from each class ?

michaelpulsewidth commented 5 years ago

Hi, we used random images from the whole ImageNet and didn't keep the class balanced. We used this script to preprocess the ImageNet and saved the data into 1024 files. We used the first 102 files as the training data. Tensorflow is a bit different from pytorch. If we only share the model weights without the code, it cannot be executed.

chaithu536 commented 5 years ago

Thank you for providing link to the script. I just want to confirm that the data from only these 102 files are used for fully supervised learning to obtain 55.09% Top-1 accuracy as reported in the paper (Table 3). Is this correct ? Which network architecture is used to train on this dataset ?

michaelpulsewidth commented 5 years ago

Yes, that's correct. We use ResNet-50 as our baseline model. The code of our baseline was obtained here.