nn.py 文件源码

def deconv(inp, name, filter_size, out_channels, stride=1,
           padding='SAME', nonlinearity=None, init_scale=1.0):
    """ Deconvolution layer. See `conv`"""
    with tf.variable_scope(name):
        strides = [1, stride, stride, 1]
        [N, H, W, in_channels] = inp.get_shape().as_list()
        if padding == 'SAME':
            target_shape = [N, H * stride, W * stride, out_channels]
        else:
            target_shape = [N, H * stride + filter_size[0] - 1, W * stride + filter_size[1] - 1, out_channels]
        target_shape = tf.constant(target_shape, dtype=tf.int32)

        if tf.GLOBAL['init']:
            V = get_variable('V', shape=filter_size + (out_channels, in_channels), dtype=tf.float32,
                             initializer=tf.random_normal_initializer(0, 0.05), trainable=True)
            V_norm = tf.nn.l2_normalize(V.initialized_value(), [0, 1, 3])
            out = tf.nn.conv2d_transpose(inp, V_norm, target_shape, strides, padding)
            m_init, v_init = tf.nn.moments(out, [0, 1, 2])
            scale_init = init_scale / tf.sqrt(v_init + 1e-8)
            g = get_variable('g', shape=None, dtype=tf.float32, initializer=scale_init, trainable=True, regularizer=tf.contrib.layers.l2_regularizer(tf.GLOBAL['reg']))
            b = get_variable('b', shape=None, dtype=tf.float32, initializer=-m_init * scale_init, trainable=True, regularizer=tf.contrib.layers.l2_regularizer(tf.GLOBAL['reg']))
            out = tf.reshape(scale_init, [1, 1, 1, out_channels]) * (out - tf.reshape(m_init, [1, 1, 1, out_channels]))
            if nonlinearity is not None:
                out = nonlinearity(out)

        else:
            V, g, b = get_variable('V'), get_variable('g'), get_variable('b')
            tf.assert_variables_initialized([V, g, b])
            W = g[None, None, :, None] * tf.nn.l2_normalize(V, [0, 1, 3])
            out = tf.nn.conv2d_transpose(inp, W, target_shape, strides, padding) + b[None, None, None]
            if nonlinearity is not None:
                out = nonlinearity(out)

        return out