Neural Style Transfer with TensorFlow

Overview

1. The code for Neural Style Transfer is in the NST_Tensorflow Jupyter Notebook. The Python file nst_setup, which builds up the Neural Network, is automatically called by the Notebook
2. The provided content and style images reside in the challenge folder
3. All 10 images generated while the model was running (1 after every 10th iteration, 100 iterations in total) are in the Generated_Images folder. The final output image matching both content and style is named iter100, and is in the same folder

Model used

VGG 19 pre-trained on ImageNet data, max-pooling replaced with average-pooling, last 3 FC layers removed

Conclusion

Neural Style transfer was successfully implemented, and this can be seen by going through the generated images in sequence (iter10 to iter100). Only 100 update iterations were carried out due to the limited computing power of my CPU. Still, the final image matches both the content of the Japanese garden and the style of the Picasso portrait.

Papers

• VGGNet: https://arxiv.org/abs/1409.1556
• Neural Style Transfer: https://arxiv.org/abs/1508.06576

Modules used

• PIL: https://pillow.readthedocs.io/en/stable/
• matplotlib: https://matplotlib.org/api/pyplot_api.html
• numpy: https://numpy.org/
• tensorflow: https://www.tensorflow.org/versions/r1.14/api_docs/python (tf version 1.14 used)
• keras.applications: https://keras.io/applications/