This repo is for the second machine learning course of the TBDSG. This meetup session is being held jointly with the tampa bay python meetup. Thus, I've shuffled the material around a bit.
This session will focus on support tools for machine learning, plus a brief survey of how to use those tools.
Since this session has two target audiences, it will be difficult to address both. Try pairing up to share expertise.
note: the links in this readme are intended to be executed from an jupyter notebook
Perhaps more than other areas, machine learning environments have deep and often specific dependencies. These can often conflict, so as a rule you should start with a virtual environment. This session assumes you are doing so.
You'll need git to get this repo. Install that now if you don't already have it.
I'm using python3 for this session. (yay)
python3 has a built in support via pyvenv, so you should not need to download additional scripts. However, if you like other tools (eg mkvirtualenv) feel free to use them.
I've listed the dependencies in requirements.txt. The step below will install those dependencies. I'll talk more about them in the next few pages.
After you have the external dependecies installed, bootstrap the environment via these commands.
git clone http://github.com/ryuhonllc/mls2
cd mls2
pyvenv .
. bin/activate
pip install -r requirements.txt
pip install pymc
Don't worry too much about having multiple venvs. pip caches stuff so you won't have to redownload, and disk space is cheap.
This section deal with stuff you can do inside python.
numpy is essential. This session touches on:
-
To get actually fast code, you need numpy
-
many libraries depend on it
- pandas
- scipy
-
Many that don't, imitate it
- tensor flow
There isn't much to say about csv. Of course, there is the standard csv library. However, be aware that many technical libraries also support csv:
If you've outgrown csv, where should you go next? hdf5 offers some great features.
notebook comparing load speeds
Looking for speed? saving your data as python and then evaluating it to reload is actually pretty fast.
make your code easier to run. Start with begins and then add parameters as needed. Check out the low pass filter and animation example.
I don't find this to be great for development, but it is great for sharing. Because many of these examples are in notebooks, I'm not covering this directly.
This is the de facto visualization tool for python, but you might outgrow it quickly depending on your requirements. It's still a good place to start.
Even though it's for computer vision, opencv can be handy for visualizing your code, especially if you are working with images, of course.
The neural network video used opencv (along with ffmpeg) to generate visuals.
The greek alphabet is all over machine learning. α is everywhere. However, we usually write 'alpha' in our code. Unfortunately, this often makes your equations too long for pep8. Increase readability by using greek symbols directly!
example:
iota = a*(1 - alpha)^(beta*t) + aprime*(alpha)^(beta*t)
# vs
Φ = a*(1-α)^(β*t) + aprime*(α)^(β*t)
- this looks more like this equation in whatever paper you're using
- the code is shorter and easier to read
Most code editors make it easy to assign abbreviations. I suggest the LaTeX bindings (eg \alpha → α). Take a look at greek.py for some vim tricks.
Finally, machine learning. Making use of the APIs in this section builds on the previous section. You'll find this in your work as well. Often, a majority of your work in say, optimization, will involve feeding the right data to the optimization function.
labeled data
- watch this video of a project I did in tensor flow
Neural networks are known for being hard to understand. Decision trees "make sense". Random forest performance is often pretty good.
unlabeled data
- clustering
- pca
- feature discovery
K-means can be useful without being complicated. This example is in one dimension, but automated grouping speeds my work up.
Arguably supervised if you consider the target equation a "label"
- hill climbing
- simulated annealing
- curve fitting
Here's an example of optimizing a low pass filter
Not well known (and not a great fit for these types)
- genetic algorithms
- Q-Learning
- bonus: monte carlo simulation