Class CondGAN

stageI/model.py:11–224

class CondGAN(object):
    def __init__(self, image_shape):
        self.batch_size = cfg.TRAIN.BATCH_SIZE
        self.network_type = cfg.GAN.NETWORK_TYPE
        self.image_shape = image_shape
        self.gf_dim = cfg.GAN.GF_DIM
        self.df_dim = cfg.GAN.DF_DIM
        self.ef_dim = cfg.GAN.EMBEDDING_DIM

        self.image_shape = image_shape
        self.s = image_shape[0]
        self.s2, self.s4, self.s8, self.s16 =\
            int(self.s / 2), int(self.s / 4), int(self.s / 8), int(self.s / 16)

        # Since D is only used during training, we build a template
        # for safe reuse the variables during computing loss for fake/real/wrong images
        # We do not do this for G,
        # because batch_norm needs different options for training and testing
        if cfg.GAN.NETWORK_TYPE == "default":
            with tf.variable_scope("d_net"):
                self.d_encode_img_template = self.d_encode_image()
                self.d_context_template = self.context_embedding()
                self.discriminator_template = self.discriminator()
        elif cfg.GAN.NETWORK_TYPE == "simple":
            with tf.variable_scope("d_net"):
                self.d_encode_img_template = self.d_encode_image_simple()
                self.d_context_template = self.context_embedding()
                self.discriminator_template = self.discriminator()
        else:
            raise NotImplementedError

    # g-net
    def generate_condition(self, c_var):
        conditions =\
            (pt.wrap(c_var).
             flatten().
             custom_fully_connected(self.ef_dim * 2).
             apply(leaky_rectify, leakiness=0.2))
        mean = conditions[:, :self.ef_dim]
        log_sigma = conditions[:, self.ef_dim:]
        return [mean, log_sigma]

    def generator(self, z_var):
        node1_0 =\
            (pt.wrap(z_var).
             flatten().
             custom_fully_connected(self.s16 * self.s16 * self.gf_dim * 8).
             fc_batch_norm().
             reshape([-1, self.s16, self.s16, self.gf_dim * 8]))
        node1_1 = \
            (node1_0.
             custom_conv2d(self.gf_dim * 2, k_h=1, k_w=1, d_h=1, d_w=1).
             conv_batch_norm().
             apply(tf.nn.relu).
             custom_conv2d(self.gf_dim * 2, k_h=3, k_w=3, d_h=1, d_w=1).
             conv_batch_norm().
             apply(tf.nn.relu).
             custom_conv2d(self.gf_dim * 8, k_h=3, k_w=3, d_h=1, d_w=1).