nnUtils.py 文件源码

def BinarizedAffine(nOutputPlane, bias=True, name=None, reuse=None):
    def b_affineLayer(x, is_training=True):
        with tf.variable_op_scope([x], name, 'Affine', reuse=reuse):
            '''
            Note that we use binarized version of the input (bin_x) and the weights (bin_w). Since the binarized function uses STE
            we calculate the gradients using bin_x and bin_w but we update w (the full precition version).
            '''
            bin_x = binarize(x)
            reshaped = tf.reshape(bin_x, [x.get_shape().as_list()[0], -1])
            nInputPlane = reshaped.get_shape().as_list()[1]
            w = tf.get_variable('weight', [nInputPlane, nOutputPlane], initializer=tf.contrib.layers.xavier_initializer())
            bin_w = binarize(w)
            output = tf.matmul(reshaped, bin_w)
            if bias:
                b = tf.get_variable('bias', [nOutputPlane],initializer=tf.zeros_initializer)
                output = tf.nn.bias_add(output, b)
        return output
    return b_affineLayer