Make example data downloader more robust

[rebeccabilbro]

Make the UCI data downloader from the examples more robust.

[bbengfort]

In the examples directory we have a simple download script that will download the example data sets for the blog posts we created on yellowbrick (and power the Jupyter notebook that is also in that directory).

However, we also need to make sure that this script can be used to download or unpack data for testing (e.g. to a tmp directory). So we have to do one of the following:

• Add compressed datasets and write a script in tests to load and decompress them in memory (or write them to a temporary directory).
• Upload the datasets to S3 (they're on dropbox now I think, which isn't exactly a great host) and then write a more robust download script.

Whoever takes this item, we'd be happy to discuss a path for either of the above.

[bbengfort]

@rebeccabilbro moving forward on this for my afternoon Yellowbrick session; just to take on something fairly easy so that I can get back to my proposal.

@ojedatony1616 I'm going to add the datasets to the DDL Data Bucket on S3; hope that's ok, let me know if you have other thoughts.

Tasks I'm going to do here:

• Create data set bundles as CSV with header and meta data and README from the three data sets we currently have.
• Zip the dataset bundles into a single file with an MD5 hash.
• Create a downloader that checks the hash against the file and stores it locally.
• Update the examples.ipynb and examples.py with downloaded data.
• Create a test fixture that downloads the data to a temporary file.
• Write a test or two that uses the temporary fixture.

[bbengfort]

@rebeccabilbro ok, so what I've done (particularly as it relates to #60) for each dataset is as follows:

1. Created a README.md from the UCI ML Repository information with correct citations (including bibtex) and other information (similar to the bundle methodology)

2. Created a meta.json with the feature names and target names if a classifier

3. Created a single csv dataset (e.g. combining test, training, etc.) with a header row that works with pd.read_csv.

4. Ensured that the feature names in the CSV file were easily understandable.

5. Ensure that the target is the last column in the dataset

6. Packaged the dataset into a directory with zip as follows:

   $ zip -r name.zip name/ 
   

   Where name is the single identifier of the directory (e.g. "occupancy")

7. Took the sha256 hash of the file and stored it in download.py to ensure that the latest version of the dataset is being downloaded and that it hasn't been corrupted.

8. Uploaded the .zip file to S3 -- namely to DDL's data-lake bucket and made it publicly available for download

9. Modified download.py to ensure that the new file is downloaded in download_all.

If possible, I'd like to ensure that all of our data sets that we produce for Yellowbrick are treated in a similar fashion.

[bbengfort]

Also note that this means the examples.ipynb will break if you haven't redownloaded the data files!

[bbengfort]

And now we cross our fingers that Travis passes ...