kktv_datagame_1711_cnn.py

def build_model():
    """
    """
    eigens = tf.placeholder(shape=[None, 32, 28, 1], dtype=tf.float32)
    labels = tf.placeholder(shape=[None, 28], dtype=tf.float32)

    flow = eigens

    weights_initializer = tf.truncated_normal_initializer(stddev=0.02)

    # conv
    for i in xrange(FLAGS.body_conv_layers):
        flow = tf.contrib.layers.convolution2d(
            inputs=flow,
            num_outputs=FLAGS.body_conv_output,
            kernel_size=FLAGS.body_conv_kernel,
            stride=1,
            padding='SAME',
            activation_fn=leaky_relu,
            normalizer_fn=instance_norm,
            weights_initializer=weights_initializer,
            biases_initializer=None,
            scope='body_conv_{}'.format(i))

    flow = tf.contrib.layers.flatten(flow)

    for j in xrange(FLAGS.fc_layers):
        flow = tf.contrib.layers.fully_connected(
            inputs=flow,
            num_outputs=FLAGS.fc_output,
            activation_fn=leaky_relu,
            weights_initializer=weights_initializer,
            scope='fc_{}'.format(j))

    flow = tf.contrib.layers.fully_connected(
        inputs=flow,
        num_outputs=28,
        activation_fn=None,
        weights_initializer=weights_initializer,
        scope='fc_end')

    guess = tf.nn.sigmoid(flow)

    loss = my_loss(flow, labels)

    learning_rate = tf.placeholder(shape=[], dtype=tf.float32)

    trainer = tf.train \
        .AdamOptimizer(learning_rate=learning_rate, beta1=0.5) \
        .minimize(loss)

    return {
        'eigens': eigens,
        'labels': labels,
        'guess': guess,
        'loss': loss,
        'trainer': trainer,
        'learning_rate': learning_rate,
    }