CS:4980 Deep Learning Assignment

CS:4980:006 Deep Learning Assignment 9: Due 11/16/2018

In this assignment, we will create some photos of digital art: deep dream and neural style photos.

For deep dream photos, we use deepdream.py. This program uses Google's inception network (a variant of GoogLeNet) and its trained network can be found here: tensorflow_inception_graph.pb. For example, if we use trump.jpg as the input, the output picture would be trump2.jpg.

For neural style photos, we use the code style-py.tgz obtained from https://github.com/anishathalye/neural-style, which consists of three python files. The trained network is VGG-19 convolutional neural network . The network is available here . For example, the content picture could be pitt.jpg, the style picture could be stars.jpg, and the output picture would look like pitt-output.jpg. The script to run this program can be found here: style.sh.

At first, make the codes working on your machine.

You need to finish the following tasks in this assignment:

Take the picture trump.jpg as the input, select in turn the following layers
```
     mixed3a_3x3_bottleneck_pre_relu 96
     mixed4a_3x3_bottleneck_pre_relu 96
     mixed5a_3x3_bottleneck_pre_relu 160
```
with channel number = 10, 40, 90, respectively, in deepdream.py. Collect the last photo produced in each run and compare them. Submit these photosphotos with your observations and comments.
Pick one of your portaits as the content picture and pick any other photo of a modern painting as the style picture, run style-py.tgz on the Argon with the following additional parameters:
```
content-weight     style-weight  
5.0                500.0            (default)
20.0               100.0
50.0               50.0
2.0                1000.0
1.0                2000.0
```
In style.sh, you may call python as
```
python neural_style.py --content-weight 50 --style-weight 50 ...
```
Collect the output photo produced in each run and compare them. Submit these photos plus the content and style photos with your observations and comments.

If you like to run deepdrem.py on Argon, you need to copy the function imsave from neural_style.py to deepdrem.py and comment off the show function in deepdream.py.

Please submit everything required in the ICON dropbox for Assignment 9 before the deadline.

Thank you!