nn.py 文件源码

def deconv2d(x, num_filters, filter_size=[3,3], stride=[1,1], pad='SAME', nonlinearity=None, init_scale=1., counters={}, init=False, ema=None, **kwargs):
    ''' transposed convolutional layer '''
    name = get_name('deconv2d', counters)
    xs = int_shape(x)
    if pad=='SAME':
        target_shape = [xs[0], xs[1]*stride[0], xs[2]*stride[1], num_filters]
    else:
        target_shape = [xs[0], xs[1]*stride[0] + filter_size[0]-1, xs[2]*stride[1] + filter_size[1]-1, num_filters]
    with tf.variable_scope(name):
        if init:
            # data based initialization of parameters
            V = tf.get_variable('V', filter_size+[num_filters,int(x.get_shape()[-1])], tf.float32, tf.random_normal_initializer(0, 0.05), trainable=True)
            V_norm = tf.nn.l2_normalize(V.initialized_value(), [0,1,3])
            x_init = tf.nn.conv2d_transpose(x, V_norm, target_shape, [1]+stride+[1], padding=pad)
            m_init, v_init = tf.nn.moments(x_init, [0,1,2])
            scale_init = init_scale/tf.sqrt(v_init + 1e-8)
            g = tf.get_variable('g', dtype=tf.float32, initializer=scale_init, trainable=True)
            b = tf.get_variable('b', dtype=tf.float32, initializer=-m_init*scale_init, trainable=True)
            x_init = tf.reshape(scale_init,[1,1,1,num_filters])*(x_init-tf.reshape(m_init,[1,1,1,num_filters]))
            if nonlinearity is not None:
                x_init = nonlinearity(x_init)
            return x_init

        else:
            V, g, b = get_vars_maybe_avg(['V', 'g', 'b'], ema)
            tf.assert_variables_initialized([V,g,b])

            # use weight normalization (Salimans & Kingma, 2016)
            W = tf.reshape(g,[1,1,num_filters,1])*tf.nn.l2_normalize(V,[0,1,3])

            # calculate convolutional layer output
            x = tf.nn.conv2d_transpose(x, W, target_shape, [1]+stride+[1], padding=pad)
            x = tf.nn.bias_add(x, b)

            # apply nonlinearity
            if nonlinearity is not None:
                x = nonlinearity(x)
            return x