Augmentation's randomness in classification code

[baek85]

Before asking my question, thank you for sharing your work!

I found that color augmentation is deterministic in your classification code.

In classification code(cls/dalib/modules/masking.py line 100-109), self.augmentation_params['color_jitter'] and self.augmentation_params['blur'] is sampled only at initialization of Masking class.
But, these parameters are related to randomness of augmentations(color jittering and gaussian blur).

I guess these parameters' sampling should be done in every forward step as with segmentation code(seg/mmseg/model/uda/masking_consistency_module.py line 119-128)

Thanks for your help

[lhoyer]

Hi @baek85,

Thank you for your interest in our work!

Yes, your are right. This is indeed an unintended behavior. However, to ensure consistency of the published results and provided source code, we have decided to keep the current version.

Best,
Lukas