nnUtils.py 文件源码

def BinarizedSpatialConvolution(nOutputPlane, kW, kH, dW=1, dH=1,
        padding='VALID', bias=True, reuse=None, name='BinarizedSpatialConvolution'):
    def b_conv2d(x, is_training=True):
        nInputPlane = x.get_shape().as_list()[3]
        with tf.variable_op_scope([x], None, name, reuse=reuse):
            w = tf.get_variable('weight', [kH, kW, nInputPlane, nOutputPlane],
                            initializer=tf.contrib.layers.xavier_initializer_conv2d())
            bin_w = binarize(w)
            bin_x = binarize(x)
            '''
            Note that we use binarized version of the input and the weights. Since the binarized function uses STE
            we calculate the gradients using bin_x and bin_w but we update w (the full precition version).
            '''
            out = tf.nn.conv2d(bin_x, bin_w, strides=[1, dH, dW, 1], padding=padding)
            if bias:
                b = tf.get_variable('bias', [nOutputPlane],initializer=tf.zeros_initializer)
                out = tf.nn.bias_add(out, b)
            return out
    return b_conv2d