@WGierke, sure. Unfortunately I wasn't joking when I said I needed a new harddisk before I can start to run the models and I'm out of town for the next 2 weeks without my DL box. So feel free to merge into master now. On a first glance I noticed you probably missed some dependencies, ```scikit-image``` and ```lxml```. Can you maybe take a look at my branch? I should have the complete list in the markdown file. Same location. Also, I have seen that in order to get it running on Linux, you will have to remove and replace the dependency on ```ntpath``` with ```os.path``` in the step3 file. If I am not mistaken Python will handle this dependency automatically, so it should still work on Windows. If I find anything else later I will create a new issue. Great writeups so far. Hope I can write one soon too.

I have been looking at the Julian de Wit model myself, and I was preparing to do a test run on my local machine -- it took a bit longer than expected since NSDB3 has a lot of training data and my harddisk is too small. My approach was to get the model running locally, so that I am sure that the current versions of libraries and CUDA/cuDNN versions work with this model, after which we can freeze the requirements for the documentation. For instance, doing this I have noticed one dependency on ```ntpath``` in Julian de Wit's model which has to be replaced with ```os.path``` before it can run on Linux. This seems to be even more important with the Daniel Hammack algorithm (a.o. a possible openCV issue and Keras amendments that can cause problems in reproducing the results). Is it important to have detailed versions of libraries for the documentation here or should be cross that bridge when we get there?