search

OpenGVLab / VideoMamba

VideoMamba: State Space Model for Efficient Video Understanding
https://arxiv.org/abs/2403.06977
Apache License 2.0
660 stars 47 forks source link

Augmentation difference between tiny and small variants #49

Open TalalWasim opened 1 month ago

TalalWasim commented 1 month ago

Hi,

I had a question regarding the augmentation strategy used between tiny and small variants. There are two different arguments used in tiny variant for the image model that are not used for the small variant as below:

--clip-grad 5.0 --no-repeated-aug --aa v0

Is there any particular reason for this?

Andy1621 commented 1 month ago

Good question~ For the tiny model (<10M), the best practice is to use weaker augmentation.

As for the --clip-grad, it's not a key.