An experimental project using deep convolutional networks. It is based on Google's DeepDream program and was written using the SciPy ecosystem and Keras. The original DeepDream program is infamous for generating images of "dogslugs", because the network was trained on an dataset that has images of many different breeds of dogs. I wondered if it might be possible to generate "catslugs" by retuning the network on different data - specifically, the cat breed images from the Oxford-IIIT Pet Dataset.
The project is complete. Ultimately, the output was a bit disappointing - it was noticeably less doglike than the original DeepDream images, but did not produce "catslugs". Nevertheless, I learned a great deal about constructing and training deep neural networks.
First and foremost, the original Keras "DeepDream" project does not generate "dogslugs" to nearly the same extent that the original Caffe-based "DeepDream" project does. I can think of two possible reasons:
In addition, there may have been problems with the procedures I followed: