Update README.md

richardwth · Apr 6, 2019 · e2cb214 · e2cb214
1 parent ac2226c
commit e2cb214
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/README.md b/README.md
@@ -69,7 +69,7 @@ Unlike the case of Wasserstein GAN, we do not encourage using the repulsive loss
 - both <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_D^{\text{rep}}" title="L_D^{\text{rep}}"/> and <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_G}" title="LG"/> may be close to 0 initially; this is because both G and D are weak.
 - <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_G}" title="L_G"/> may gradually increase during training; this is because it becomes harder for G to generate high quality samples and fool D (and G may not have the capacity to do so).
 
-For balanced and capable G and D, we would expect both <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_D^{\text{rep}}" title="L_D^{\text{rep}}"/> and <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_G}" title="L_G"/> to stay close to 0 during the whole training process and any kernel (i.e., <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(x,x)}" title="k_D(x,x)"/>, <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(y,y)}" title="k_D(y,y)"/> and <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(x,y)}" title="k_D(x,y)"/>) to be close to 0.6 and away from 0 or 1.
+For balanced and capable G and D, we would expect both <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_D^{\text{rep}}" title="L_D^{\text{rep}}"/> and <img src="https://latex.codecogs.com/gif.latex?\inline&space;L_G}" title="L_G"/> to stay close to 0 during the whole training process and any kernel (i.e., <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(x,x)}" title="k_D(x,x)"/>, <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(y,y)}" title="k_D(y,y)"/> and <img src="https://latex.codecogs.com/gif.latex?\inline&space;k_D(x,y)}" title="k_D(x,y)"/>) to be away from 0 or 1 and stay in the middle (e.g., 0.6).
 
 In some cases, you may find training using the repulsive loss diverges. Do not panic. It may be that the learning rate is not suitable. Please try other learning rate or the bounded kernel.