dcgan.py 文件源码

def build_summaries(gan):
    """
    """
    g_summaries = []
    d_summaries = []

    g_summaries.append(
        tf.summary.scalar('generator loss', gan['generator_loss']))

    d_summaries.append(
        tf.summary.scalar('discriminator loss', gan['discriminator_loss']))

    for vg in gan['generator_variables_gradients']:
        variable_name = '{}/variable'.format(vg[0].name)
        gradient_name = '{}/gradient'.format(vg[0].name)

        g_summaries.append(tf.summary.histogram(variable_name, vg[0]))
        g_summaries.append(tf.summary.histogram(gradient_name, vg[1]))

    for vg in gan['discriminator_variables_gradients']:
        variable_name = '{}/variable'.format(vg[0].name)
        gradient_name = '{}/gradient'.format(vg[0].name)

        d_summaries.append(tf.summary.histogram(variable_name, vg[0]))
        d_summaries.append(tf.summary.histogram(gradient_name, vg[1]))

    # fake image
    image_width, image_depth = (64, 3) if FLAGS.use_lsun else (32, 1)

    fake_grid = tf.reshape(
        gan['generator_fake'],
        [1, FLAGS.batch_size * image_width, image_width, image_depth])
    fake_grid = tf.split(fake_grid, FLAGS.summary_col_size, axis=1)
    fake_grid = tf.concat(fake_grid, axis=2)
    fake_grid = tf.saturate_cast(fake_grid * 127.5 + 127.5, tf.uint8)

    summary_generator_fake = tf.summary.image(
        'generated image', fake_grid, max_outputs=1)

    g_summaries_plus = g_summaries + [summary_generator_fake]

    return {
        'summary_generator': tf.summary.merge(g_summaries),
        'summary_generator_plus': tf.summary.merge(g_summaries_plus),
        'summary_discriminator': tf.summary.merge(d_summaries),
    }