tips.md

Hyperparameter Tuning

• Discriminator

  • We set the number of scales (n_scales_dis) to be a constant (4) in the paper. This is to verify that a good performance can be achieved for different tasks with a fixed network structure. We found in practice that a reduced number of scales (3) is often discriminative enough and more stabilized, especially for the tasks which require a significant shape deformation. In such case, the 4-th scale, which is the most discriminative one in the default setting, tends to be an overkill.
  • Reducing the number of base channels (ch_dis) is an effective way to accelerate the training process.

• Generator

  • To improve the inference speed, you may want to reduce the number of base channels (ch_gen) or the number of enhanced upsampling layers (n_enhanced_upsample_gen).

• Weak cycle

  • The weight for the weak cycle constraint (cyc_weight) can be adjusted on a per-task basis. A large value is generally more prohibitive for the shape deformation. On the other hand, it helps to keep the generated image correlated with the source image.
  • The input of the backward generator is resized by 1/(resize_factor_gen_bw). The number of downsampling and upsampling layers of the backward generator is set by n_updownsample_gen_bw. These two parameters can be adjusted to change the weak cycle to a full resolution forward cycle, or something in between. The effect is somewhat similar to increasing the cycle weight.

• Training with a higher resolution

  • You may want to adjust several parameters if the input / output resolution is higher than 256x256. Take 512x512 as an example. A good starting point will be setting img_size to 512, increasing n_downsample_init_dis from 2 to 3, and reducing ch_dis, ch_gen and ch_gen_bw by a factor of two.