MatthieuCourbariaux
MatthieuCourbariaux
Yes, it is on my TODO list ;-)
I had a working build 3 years ago.
Back then, on a GTX Titan Black, each epoch took about 22 seconds. The whole training lasted about 6h. Unfortunately, I did not keep the trained parameters.
1. I do not remember. 2. e is a small number to prevent instability... maybe 10^-6. I do not remember.
I think you have to keep N a power of 2 (e.g. 1024 and 2048 might work). Our XNOR kernel indeed only works for some limited settings. It was only...
Your issue may be that the targets are not onehot encoded. You may want to use this code to onehot the targets: https://github.com/MatthieuCourbariaux/BinaryNet/blob/master/Train-time/mnist.py#L109
> It seems to be reasonable, but when we really try to realize this in one bit, how should we represent -1? Should we convert it into 0? In practice,...
Yes, we do intend to train a BinaryNet on Imagenet by the end of June. Besides, a very [recent paper](http://arxiv.org/abs/1603.05279) already tried to train some "XNOR-Nets" on Imagenet.