The discriminator and GAN loss values for 100 iterations can be plotted as follows:
From the preceding plot, we can observe the following:
- By increasing the number of layers, the fluctuation in the loss values for the discriminator and GAN network has reduced compared to the results we obtained earlier.
- The spikes or high loss values that have been observed for some of the iterations indicate the corresponding network struggling, while competing against the other network.
- The variability in the GAN loss values continues to be higher compared to those for discriminator network-related loss.
The following plot is of the first fake image in each of the 100 iterations:
From the preceding images, we can observe the following:
- With additional convolutional layers in the generator and discriminator networks, the network begins to generate images replicating the handwritten digit five much earlier.
- In the previous network, fake images that consistently looked like handwritten digit five did not appear until about 70-80 iterations.
- Due to the use of additional layers, we can see the digit five being formed more or less consistently after about 20-30 iterations, which suggests an improvement.
Next, we will try to use this network to generate another handwritten digit.