AutoBNN
AutoBNN copied to clipboard
Model sizes of ImageNet models
Hi,
I read your paper and think it is very interesting. You already provided the number of FLOPs of Res18-Auto-A and Res18-Auto-B.
Did you also calculate or measure the model size (assuming we store the binary weights as 1 bit) of Res18-Auto-A and Res18-Auto-B for the ImageNet models? That would be very interesting for comparison purposes.
Best regards, Joseph
Hi Joseph,
Thanks for your interest. The statistics of AutoBNN for ImageNet is as follows:
| model | model size | top-1 acc |
|---|---|---|
| Res18-Auto-A | ~15MB | 68.64% |
| Res18-Auto-B | ~19MB | 69.65% |
Best, Kai