search

allenai / XNOR-Net

ImageNet classification using binary Convolutional Neural Networks
https://xnor.ai/
Other
856 stars 239 forks source link

About the paper #42

Open ycjing opened 4 years ago

ycjing commented 4 years ago

Hi @mrastegari

Thank you for the nice paper and code! The work is really impressive. I have a question about the paper. As mentioned in other issues #11 , Torch does not support bit operations. But in the paper, there is a statement: "This results in 58× faster convolutional operations (in terms of number of the high precision operations) and 32× memory savings". I would appreciate it if you could explain the way to compute these values (i.e., 58x and 32x). Thank you! Wish you all the best!

Best, Yongcheng

resurgo97 commented 3 years ago

The inference speed is increased on particular accelerators that support bit-wise operations. Of course it doesn't make it fast to train/inference the model on GPUs.

ycjing commented 3 years ago

Hi @resurgo97

Thank you for the response! I appreciate it. Yes, I understand that the improvement in speed can be only achieved on devices that support bit-wise operations. But I wonder how to derive the theoretical improvement in speed (e.g., 58x and 32x). Thank you!

Best, Yongcheng