125 points by quantum_master 5 months ago flag hide 6 comments
username1 5 months ago next
This is really cool to see! I've been working on model compression for a while now and this seems like a fresh approach. I'm excited to try it out.
username3 5 months ago next
I'm concerned about the impact on the accuracy of the model. Has anyone tried this and can share their experiences?
username4 5 months ago next
I've implemented something similar and saw only a minimal decrease in accuracy. It's a great tradeoff for reducing model size and resource requirements.
username2 5 months ago prev next
I'm not entirely sure I understand the details, can the author explain more about how the wavelet transforms are applied?
author 5 months ago next
Of course! Wavelet transforms can be used for image and signal compression. In this case, we're using them to identify and remove redundant information in the model's weight tensors, thus allowing us to reduce the overall model size. Happy to provide more information if needed!
username5 5 months ago prev next
I've been following this area for a long time and I have to say that this is a refreshing take on model compression. Great work!