Name : Divam Gupta

Andrew ID : divamg

Contents:

Content Reconstruction

bash content_loss_exps.sh

Effect of optimizing content loss at different layers

Used tubingen.jpeg for the experiment

We can see that as we go higher, the reconstruction is more noisy.

Results of two random noises

used phipps.jpeg

We can see that the outputs of both the runs are pretty much the same.

Texture Synthesis

Used the starry night for this experiment.

bash texture_exps.sh 

Effect of optimizing texture loss at different layers

Results of two random noises

• We see as we use higher level conv layers then we can see more complex patterns in the synthesizes texture.

• We also see that the two diffrent outputs are similar but not exactly the same

Style Transfer

bash style_transfer_exps.sh

Implementation details

Hyperparamers

• style_layers=conv_1,conv_2,conv_3,conv_4,conv_5

• style_weight=100000

• content_weight=1

• num_steps=300

• content_layers=conv_4

Grid of results

Compare with random initialization

We see that content image initialization is much better

Style transfer on my images

References

• https://pytorch.org/tutorials/advanced/neural_style_tutorial.html

• https://learning-image-synthesis.github.io/assignments/hw4

• https://arxiv.org/pdf/1508.06576.pdf