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
2.08k stars 373 forks source link

the accuracy during quantization aware training decreased and keep at 0.1 #1517

Open roycyh opened 2 years ago

roycyh commented 2 years ago

hello, when I running the example quantization aware training or range learning, the top 1 accuracy decreases to 0.1 and keeping unchanged. Did anyone meet this problem?

quic-mangal commented 1 year ago

@roycyh, thanks for reaching out. When you create a sim object and perform sim.compute_encodings and perform evaluation, do you still get this accuracy drop?

One recommendation is to visualize the weight ranges (you can use AIMET APIs to do the same) and if the weight ranges are very large you can try using the Cross Layer Equalization feature.