a02527e58d09b7f86575f5b790db1fa9dbc8e4e6,lib/gan.py,,discriminator,#Any#Any#Any#Any#Any#Any#,46

Before Change

    last_layer = tf.reshape(last_layer, [batch_size, -1])
    last_layer = tf.slice(last_layer, [single_batch_size, 0], [single_batch_size, -1])

    net = tf.reshape(net, [config["batch_size"]*2, -1])
    net = tf.concat(1, [net]+regularizers)
    s = [int(x) for x in net.get_shape()]
    net = tf.reshape(net, [s[0], 1, 1, s[1]])
    net = conv2d(net, int(net.get_shape()[3]), name="d_endd2", k_w=1, k_h=1, d_h=1, d_w=1, stddev=0.1)
    net = tf.reshape(net, [config["batch_size"]*2, -1])
    net = tf.reduce_mean(net, 1)
    net = tf.reshape(net,  [config["batch_size"]*2, 1])
    class_logits = net
    gan_logits = net
    return [tf.slice(class_logits, [0, 0], [single_batch_size, 1]),
                tf.slice(gan_logits, [0,0], [single_batch_size,1]),
                tf.slice(class_logits, [single_batch_size, 0], [single_batch_size, 1]),
                tf.slice(gan_logits, [single_batch_size,0], [single_batch_size,1]),
                last_layer]

    net = linear(net, config["y_dims"]+1, scope="d_proj", stddev=0.002)

    def build_logits(class_logits, num_classes):
        generated_class_logits = tf.squeeze(tf.slice(class_logits, [0, num_classes - 1], [batch_size, 1]))
        positive_class_logits = tf.slice(class_logits, [0, 0], [batch_size, num_classes - 1])

After Change

    regularizers = []
    for regularizer in config["discriminator.regularizers"]:
        regs = regularizer(config, net)
        regularizers += [tf.reshape(r,[int(r.get_shape()[0]),1,1,int(r.get_shape()[1])]) for r in regs]

    net = tf.concat(3, [net]+regularizers)
    net = conv2d(net, net.get_shape()[3], name="d_endd3", k_w=1, k_h=1, d_h=1, d_w=1, stddev=0.3)

Instances