xnor_net.py 文件源码

def convolve(self, input, deterministic=False, **kwargs):
        """ Binary convolution. Both inputs and weights are binary (+1 or -1)
        This overrides convolve operation from Conv2DLayer implementation
        """
        if(self.xnor):
            # compute the binary inputs H and the scaling matrix K
            input, K = binarize_conv_input(input, self.beta_filter)

            # Compute the binarized filters are the scaling matrix
            self.Wb, alpha = binarize_conv_filters(self.W)
            if not deterministic:
                old_alpha = theano.clone(self.xalpha, share_inputs=False)
                old_alpha.default_update = alpha
                alpha += 0*old_alpha
            else:
                alpha = self.xalpha 

            # TODO: Use XNOR ops for the convolution. As of now using Lasagne's convolution for
            # functionality verification.
            # approx weight tensor
            #W_full_precision = self.Wb * alpha.dimshuffle(0, 'x', 'x', 'x')
            Wr = self.W

            self.W = self.Wb

            feat_maps = super(Conv2DLayer, self).convolve(input, **kwargs)
            # restore the approx full precision weight for gradiant computation
            #self.W = W_full_precision
            self.W = Wr

            # scale by K and alpha
            # FIXME: Actually we are scaling after adding bias here. Need to scale first and then add bias.
            # The super class method automatically adds bias. Somehow need to overcome this..
            # may subtract the bias, scale by alpha and beta ans then add bias ?
            feat_maps = feat_maps * K

            feat_maps = feat_maps * alpha.dimshuffle('x', 0, 'x', 'x')
        else:
            feat_maps = super(Conv2DLayer, self).convolve(input, **kwargs)

        return feat_maps