Updates after FPL2020

[thesps]

Two main updates after FPL2020:

• Update to 'part 2'. The flow of this part is now improved, I think: the profiling and trace functionality are introduced first, then the config is update to reduce the precision while maintaining accuracy. Previously, the config change made performance a lot worse
• Include the Dockerfile and associated scripts to build a docker image. For the tutorial itself, I built an image like this, and uploaded that to my JupterHub instance. This is not so much for use by people wanting to take the tutorial, but for people setting up tutorials to give in the future. Although users can of course create an image if they want an even more self-contained environment.

I also made a few minor improvements:

• Set the random seed to one found to give good model performance. Previously, without setting the random seed, sometimes the hls4ml model would lose some performance wrt the Keras model. This happened when the default ap_fixed<16,6> config was not suitable for the particular trained model. Fixing the seed guarantees reproducibility.
• Save the dataset locally. fetch_openml is supposed to cache things, but it never seems to work. This saves a bit of time.

I'm setting this to 'WIP' for now, because developing this exposed a bug in hls4ml, which is fixed by PR hls-fpga-machine-learning/hls4ml#219. So, when that is merged, the environment.yml here will need to update.

To test the PR notebooks on Binder:

[husnainmubarik]

I've the following error - when I run the first tutorial (part1_getting_started) without making any changes

myproject/input1_V' has no fanin or fanout and is left dangling

And my hls accuracy is 0.20 where the keras-model accuracy is 0.75.

Any idea what could be the issue here?