Ternary net how to store the model

[luhaofang]

No description provided.

[fengfu-chris]

You need to store the ternary weights and the scaling factor alpha. For
X * (alpha * Wt) = (alpha * X) * Wt
this would hold the potential of speeding up the forward propagation.

[mathmanu]

@fengfu-chris
Why not do it as:
(X * Wt) * alpha
So convolution is just add/substract and finally a scaling is done to the output.
Is my understanding correct - is it possible to implement it this way?

[fengfu-chris]

@mathmanu That is correct. Both methods can be implemented with an additional Scale layer in Caffe.

[manwu1994]

@fengfu-chris
Why not do it as:
(X * Wt) * alpha
So convolution is just add/substract and finally a scaling is done to the output.
Is my understanding correct - is it possible to implement it this way?

I also agree with ((X * Wt) * alpha) that speeding up the inference in hardware, do you have any method to extract the alpha parameters of each neurons? Thank you...