5,321
edits
Line 33: | Line 33: | ||
Each generator consists of 5 convolutional blocks:<br> | Each generator consists of 5 convolutional blocks:<br> | ||
Conv(<math>3 \times 3</math>)-BatchNorm-LeakyReLU.<br> | Conv(<math>3 \times 3</math>)-BatchNorm-LeakyReLU.<br> | ||
Note: This generator is similar to pix2pix.<br> | |||
They use 32 kernels per block at the coarsest scale and increase <math>2 \times</math> every 4 scales.<br> | They use 32 kernels per block at the coarsest scale and increase <math>2 \times</math> every 4 scales.<br> | ||
Line 39: | Line 40: | ||
* Internal Covariate Shift - the change in distribution of network activations as network parameters change. | * Internal Covariate Shift - the change in distribution of network activations as network parameters change. | ||
* Whitening | * Whitening | ||
====Leaky Relu==== | |||
Relu is | |||
<math>\begin{cases} | |||
x & \text{if }x > 0\\ | |||
0 & \text{if }x <= 0 | |||
\end{cases} | |||
</math>. <br> | |||
If the input is <math><=0</math> then any gradient through that neuron will always be 0.<br> | |||
This leads to dead neurons which remain dead if the neurons below never output a positive number.<br> | |||
That is, you get neurons which always output <math>0</math> throughout the training process.<br> | |||
Leaky relu: <math> | |||
\begin{cases} | |||
x & \text{if }x > 0\\ | |||
0.01x & \text{if }x <= 0 | |||
\end{cases} | |||
</math> | |||
always has a gradient so neurons below will always be updated. | |||
===Discriminator=== | ===Discriminator=== |