search

facebookresearch / kill-the-bits

Code for: "And the bit goes down: Revisiting the quantization of neural networks"
Other
636 stars 123 forks source link

train acc is about 60%, but test acc is only 8% #12

Closed xiaodao2049 closed 5 years ago

xiaodao2049 commented 5 years ago

Hi，thanks for your work. I want to reproduce the result in the paper, from the log file, I found that during the training phase, the Top1 acc can reach 60%, but the validation acc is only 8%.It's so werid...

pierrestock commented 5 years ago

Hey xiaodao2049,

Thanks for reaching out! Could you paste here the exact command you used to get this result?

Also, do you refer to the validation accuracy as the one given by running inference.py on the compressed model obtained by running quantize.py? In this case, issue #9 should help you (you have to patch two lines of code in inference.py).

Please let me know if this works for you!

Pierre

xiaodao2049 commented 5 years ago

Thank you，pierrestock. Issue #9 does work for me. And the val acc in my log is so low due to my val data is not correct, everything is fine now.