minoring/deepdream

Experiment that visualizes the patterns learned by a neural network.

deepdream

Experiment that visualizes the patterns learned by a neural network.

Usage

python main.py --num_steps=900 --learning_rate=0.001 --scaling=False

python main.py --num_steps=300 --learning_rate=0.001 --scaling=True --num_octaves=3

python main.py --num_steps=300 \
               --learning_rate=0.001 \
               --scaling=True \
               --URL=1024px-Van_Gogh_-_Starry_Night_-_Google_Art_Project.jpg

Experiment

Training Without scaling	Same number of steps but add scaling

You can see that without scaling's output

• Noisy.
• Low resolution.
• Patterns appear like they're all happening at the same granularity.

Addresses these problems by applying gradient ascent at different scales. This will allow patterns generated at smaller scales to be incorporated into patterns at higher scales and filled in with additional detail.

References

blogs

• Inceptionism: Going Deeper into Neural Networks

code

• https://www.tensorflow.org/tutorials/generative/deepdream

Model Architecture

Code: Inception V3 model for Keras

Paper: Rethinking the Inception Architecture for Computer Vision (CVPR 2016)