322f71f0f8c06a738c47cf825f8726ee5f1f496a,lib/discriminators/pyramid_nostride_discriminator.py,,discriminator,#Any#Any#Any#Any#Any#,6

Before Change

      filter_size_h = 2
      if(i==depth-1):
          filter_size_w=int(result.get_shape()[1])
          filter_size_h=int(result.get_shape()[2])
      filter = [1,filter_size_w,filter_size_h,1]
      stride = [1,filter_size_w,filter_size_h,1]
      result = conv2d(result, int(int(result.get_shape()[3])*depth_increase), name="d_expand_layer"+str(i), k_w=3, k_h=3, d_h=1, d_w=1)
      result = tf.nn.avg_pool(result, ksize=filter, strides=stride, padding="SAME")

After Change

      // APPEND xs[i] and gs[i]
      if(i < len(xs) and i > 0):
        xg = tf.concat(0, [xs[i], gs[i]])
        xg += tf.random_normal(xg.get_shape(), mean=0, stddev=config["discriminator.noise_stddev"]*i, dtype=config["dtype"])

        xgs.append(xg)

        mxg = conv2d(xg, 6*(i), name="d_add_xg"+str(i), k_w=3, k_h=3, d_h=1, d_w=1)
        mxg = batch_norm(config["batch_size"], name="d_add_xg_bn_"+str(i))(mxg)
        mxg = activation(mxg)

        xgs_conv.append(mxg)
  
        net = tf.concat(3, [net, xg])

      filter_size_w = 2
      filter_size_h = 2
      filter = [1,filter_size_w,filter_size_h,1]
      stride = [1,filter_size_w,filter_size_h,1]
      net = conv2d(net, int(int(net.get_shape()[3])*depth_increase), name="d_expand_layer"+str(i), k_w=3, k_h=3, d_h=1, d_w=1)
      net = tf.nn.avg_pool(net, ksize=filter, strides=stride, padding="SAME")

      print("Discriminator pyramid layer:", net)

    k=-1
    net = batch_norm(config["batch_size"]*2, name="d_expand_bn_end_"+str(i))(net)
    net = activation(net)
    net = tf.reshape(net, [batch_size, -1])
    net = linear(net, int(1024*1.5), scope="d_fc_end1")
    net = batch_norm(config["batch_size"]*2, name="d_bn_end1")(net)
    net = activation(net)
    net = linear(net, 1024, scope="d_fc_end2")
    net = batch_norm(config["batch_size"]*2, name="d_bn_end2")(net)

Instances