Keras deep dream implementation, based on https://github.com/fchollet/keras/blob/master/examples/deep_dream.py, with two new features:
- the ability to choose any pre-trained network from https://github.com/fchollet/deep-learning-models for dreaming
- implementing the 'octaves' iterative re-scaling functionality that was missing from the Keras example.
https://github.com/fchollet/deep-learning-models is not (yet) a python module, so the easiest way
to set things up is to clone that repo, copy deep_dream.py from here to there,
and run it there. Note that https://github.com/fchollet/deep-learning-models requires
Keras version>=1.07.
python deep_dream.py [ vgg16 | vgg19 | resnet50 | inception_v3 ] input.png output_filename_prefix [ layer ]
Like in the original example, most of the parameter tuning happens
in the setting dict variable in the source code. If the
optional positional layer argument is not provided,
it's taken from settings['features'].
The optimal parameter settings differ for the different pre-trained networks, providing sensible defaults for each of them is on my TODO list.
The algorithm compiles a new Keras model for each image size, and for the larger networks (resnet50, inception_v3), compilation is very slow with the Theano backend, slower than neural computation on a GPU. Does anyone know how to avoid this? In Caffe it's just https://github.com/google/deepdream/blame/master/dream.ipynb#L208
The deep dream algorithm is quite 2015 now, but I always wanted to see it running on residual networks, and with https://github.com/fchollet/deep-learning-models, I've grabbed the opportunity.
Porting https://github.com/fchollet/keras/blob/master/examples/deep_dream.py to work with resnet50 was trivial, but the results were quite boring. The Keras deep dream example does not implement the octaves functionality of the original deep dream algorithm, which goes through iterative re-scaling of the image, and it helps tremendously large-scale dreamed objects to appear. So I've implemented octaves, hoping that it would make deep dreaming on resnets more interesting. Unfortunately, it did not help much. But this is the first full deep dream implementation on Keras that I'm aware of, so that's something at least. I haven't yet given up on using deep dream to help understand resnets better, stay tuned, or even better, help me out.