Problem in training a GAN network

I have used a DCGAN network from this tutorial to create my simple images. The images are (35,35,3) which there is a colored circle with random radius in them.

During the training, the discriminator works well and the loss decreases in each step. However, there is a problem with Generator which the output is not desirable and it just makes some noises and in each epoch the loss increases dramatically from the beginning.
I have tried tuning hyper-params , changing the Generator model and increasing the samples but all the same.
May I ask how is your idea and where the problem is?