This is a tensorflow implementation of WGAN on mnist and SVHN.
Train: Use WGAN.ipynb, set the parameters in the second cell and choose the dataset you want to run on. You can use tensorboard to visualize the training.
Generation : Use generate_from_ckpt.ipynb, set
ckpt_dir in the second cell. Don't forget to change the dataset type accordingly.
All data will be downloaded automatically, the SVHN script is modified from this.
All parameters are set to the values the original paper recommends by default.
Ditersrepresents the number of critic updates in one step, in Original PyTorch version it was set to 5 unless iterstep < 25 or iterstep % 500 == 0 , I guess since the critic is free to fully optimized, it's reasonable to make more updates to critic at the beginning and every 500 steps, so I borrowed it without tuning. The learning rates for generator and critic are both set to 5e-5 , since during the training time the gradient norms are always relatively high(around 1e3), I suggest no drastic change on learning rates.
MLP version could take longer time to generate sharp image.
In this implementation, the critic loss is
tf.reduce_mean(fake_logit - true_logit), and generator loss is
tf.reduce_mean(-fake_logit). Actually, the whole system still works if you add a
-before both of them, it doesn't matter. Recall that the critic loss in duality form is , and the set is symmetric about the sign. Substitute $f$ with $-f$ gives us , the opposite number of original form. The original PyTorch implementation takes the second form, this implementation takes the first, both will work equally. You might want to add the
-and try it out.
Please set your device you want to run on in the code, search
tf.deviceand change accordingly. It runs on gpu:0 by default.
Inproved-WGAN is added, but somehow the gradient norm is close to 1, so the square-gradient normalizer doesn't work. Couldn't figure out why.