Deep_Dream

Implementing deep dream using pretrained Google Inception-v3

Dependencies

• numpy (http://www.numpy.org/)
• functools
• PIL (http://stackoverflow.com/questions/20060096/installing-pil-with-pip)
• tensorflow (https://www.tensorflow.org/install/#pip-installation)
• matplotlib (http://matplotlib.org/1.5.1/users/installing.html)
• urllib (https://pypi.python.org/pypi/urllib3)
• os
• zipfile

Usage

Once you have your dependencies installed via pip, run the demo script in terminal via

python deep_dream.py --image=<IMAGE_PATH> --layer=<LIST_OF_INCEPTION_LAYERS>

Arguments:

-i or --image 		

The path of input image

-l or --layers 		

List of all layers of inception-v3 network to test for generating deep dream images.

eg.- mixed3a_pool_reduce_pre_relu, mixed4b_pool_reduce_pre_relu, etc.

Refer the file 'Inception_layers.txt' for names of layers in the Inception-v3 architecture

Results

Experimenting across different activation layers/feature maps in the Inception-V3 architecture, I generated some interesting visualisations. I generated a GIF animation of the results for same image across different layers to view the projections of specific activatons of a given feature map on the input image. Let us visualise the results below:

Brad Pitt

Monalisa

Natural Scene

Sky

Credits

1. Google Research blog on Deep Dream
2. Siraj Raval's excellent video on Deep Dream in TensorFlow