Using PyTorch to apply artistic styles to pictures
This is my first ML-Project using the PyTorch library. Furthermore, I want to gain some experience with advanced ConvNet technologies and wanted to have some fun with the creative parts of machine learning.
The style-transfer script is built to work like any other command-line application. You call it using python and pass some of the following arguments to it:
| Argument | Explanation | Required? | Default |
|---|---|---|---|
| -h, --help | Displays the help message | ❌ | - |
| -c, --content | Specifies the location of the content-image | ✔️ | - |
| -s, --style | Specifies the location of the style-image | ✔️ | - |
| -d, --destination | Specifies the destination of the resulting image | ❌ | - |
| -v, --verbose | Flag; Should intermediate results be displayed? | ❌ | False |
| -i, --interval | Defines the interval in which current results should be displayed | ❌ | 10 |
| --lr | Defines the learning rate | ❌ | 0.1 |
| --content-w | Provides the content weight | ❌ | 1 |
| --style-w | Provides the style weight | ❌ | 105 |
| --tv-w | Provides the total variation weight | ❌ | 3 * 10-3 |
After a couple of bug fixes, here are some of the final results:
| Content | Style | Result |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
Various examples
| ... | August Macke - "Four Girls" | Gustav Klimt - "The Kiss" | Hiroshige - "Toyokawa Bridge" | The Tarantula Nebula | Vincent van Gogh - "Starry Night" | Edvard Munch - "The Scream" |
|---|---|---|---|---|---|---|
 |
 |
|
|
|
 |
|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Various styles applied to content images
Over all, I'm very happy with how this project turned out. Not only did I learn many new things about the PyTorch library, but I was also able to improve my skills in working with convolutional neural networks and finetuning learning processes.
Furthermore, I'm really glad about the end results. With a different GPU, one with more memory, I'd probably be able to make the pictures look even better, as it would allow for neural style transfer with larger images.
... MattMoony (September, 2019)