Class Model

examples/GAN/BEGAN.py:29–129

class Model(GANModelDesc):
    def inputs(self):
        return [tf.TensorSpec((None, args.final_size, args.final_size, 3), tf.float32, 'input')]

    @auto_reuse_variable_scope
    def decoder(self, z):
        l = FullyConnected('fc', z, NF * 8 * 8)
        l = tf.reshape(l, [-1, 8, 8, NF])

        with argscope(Conv2D, activation=tf.nn.elu, kernel_size=3, strides=1):
            l = (LinearWrap(l)
                 .Conv2D('conv1.1', NF)
                 .Conv2D('conv1.2', NF)
                 .tf.image.resize_nearest_neighbor([16, 16], align_corners=True)
                 .Conv2D('conv2.1', NF)
                 .Conv2D('conv2.2', NF)
                 .tf.image.resize_nearest_neighbor([32, 32], align_corners=True)
                 .Conv2D('conv3.1', NF)
                 .Conv2D('conv3.2', NF)
                 .tf.image.resize_nearest_neighbor([64, 64], align_corners=True)
                 .Conv2D('conv4.1', NF)
                 .Conv2D('conv4.2', NF)
                 .Conv2D('conv4.3', 3, activation=tf.identity)())
        return l

    @auto_reuse_variable_scope
    def encoder(self, imgs):
        with argscope(Conv2D, activation=tf.nn.elu, kernel_size=3, strides=1):
            l = (LinearWrap(imgs)
                 .Conv2D('conv1.1', NF)
                 .Conv2D('conv1.2', NF)
                 .Conv2D('conv1.3', NF * 2)
                 .AvgPooling('pool1', 2)
                 # 32
                 .Conv2D('conv2.1', NF * 2)
                 .Conv2D('conv2.2', NF * 3)
                 .AvgPooling('pool2', 2)
                 # 16
                 .Conv2D('conv3.1', NF * 3)
                 .Conv2D('conv3.2', NF * 4)
                 .AvgPooling('pool3', 2)
                 # 8
                 .Conv2D('conv4.1', NF * 4)
                 .Conv2D('conv4.2', NF * 4)

                 .FullyConnected('fc', NH)())
        return l

    def build_graph(self, image_pos):
        image_pos = image_pos / 128.0 - 1

        z = tf.random_uniform([args.batch, args.z_dim], minval=-1, maxval=1, name='z_train')
        z = tf.placeholder_with_default(z, [None, args.z_dim], name='z')

        def summary_image(name, x):
            x = (x + 1.0) * 128.0
            x = tf.clip_by_value(x, 0, 255)
            tf.summary.image(name, tf.cast(x, tf.uint8), max_outputs=30)