Week 11 Assignment: Explore BigGAN or deepDream – Ziying Wang (Jamie)

For this week’s project, I tweaked the parameters in the DeepDream model. 

The tried the examples in the images folder first, then discover that id photo or human portrait works the best with deep dream model. Therefore, I used the picture of Robert Downey Jr. for my deepDream project.

The original picture:

The first set of parameters I changed was the mixed figures.

From left to right, I gradually changed the mixed number from 0 to 0.2, which means for the fourth picture, all four mixed figures in the features become 0.2. The pattern on the picture slowly reduces as the mixed number of each layer increased.

Then I kept the features as they were in the fourth picture, and changed the step size from 0.01 to 0.05 and 0.08 (from left to right). The effect increases as the step size increases, as well as the pattern’s complexity.

The third parameter I tweaked was the number of the octave, I changed it from 3 to 5. The opacity increases while the complexity decreases.

Finally, I increased the iterations from 20 to 50. The opacity of the effect remains the same while the complexity lessens.

After this short tweak with terminal, I uploaded the original picture of rdj to my google cloud and used the deepdream_video_hack to generate the deepDream video.

The video’s frame length depends on the zooming rate and the dreaming rate. Since I didn’t change these rates at first, the video ended quickly. I then changed the dreaming rate to 50 and the zooming rate to 100, then I changed the fps to 150, which should get me a 1 s video, I ended up getting the 1 s fast zooming DeepDream video.

To observe the transformation better, I changed the zooming rate and changing rate to 100, then reduced the fps to 24 frames/s, and generated a 13 seconds long deepDream video that demonstrates the change well.

The zoom_factor of the previous video was too high and the zooming rate was too big, the video was zooming endlessly and the ending became boring. I, therefore, decreased both factors as well as the dreaming rate, so that the time before the zooming can reduce ass well. I ended up getting this video which is quite satisfying.

Leave a Reply