search

quic / aimet

AIMET is a library that provides advanced quantization and compression techniques for trained neural network models.
https://quic.github.io/aimet-pages/index.html
Other
2.12k stars 379 forks source link

Model compression for generative models (GAN or VAE) #820

Closed gulaoshi closed 1 year ago

gulaoshi commented 2 years ago

AIMET is designed for CNN compression, but how about generative models (GAN or VAE)? Can it be used to compress the Generator in GAN, or Decoder in VAE?

quic-ssiddego commented 2 years ago

Hi @gulaoshi Thank you for your query. Yes, AIMET compression has been tested with CNNs. GAN or VAE models with conv layers may be compressed. But, we have not tested such models. It would be great if you could contribute back your findings to us, if you give this a try.

Silk760 commented 2 years ago

it can be compressed yes, but the model will lose a significant accuracy. I have tried with another type of pruning, GANS and VAE are not as robust as CNNs. If you want to prune such a model like these you have to understand the dynamic of the neural network and the sensitivity of the layers. I am working to test data-free quantization on these models seems to me is the most promising method for these type of models. I hope Aimet team helps me to solve the issue with libpymo until I can test the method.

quic-ssiddego commented 2 years ago

@Silk760 Thank you for sharing the note on compression. Just wanted to check back on the second part above - applying DFQ on such models. Do you need any help or have further queries? Please let me know.

quic-mangal commented 1 year ago

Closing this issue due to inactivity. Please re-open it/ create a new issue if you need further help.