ENH: Add datasets

[sinhrks]

Closes #115. Added Dataset struct which has data() and target() impl (intended for supervised learning).

Adding more data once API looks OK.

[AtheMathmo]

Just need to add a feature flag - or at least discuss this.

Otherwise this looks ready to go.

[AtheMathmo]

Minor: It might be easier to use the matrix! macro here. My thinking is that if we need to add a row there's a little less work.

[AtheMathmo]

I think this makes sense for now. We might want to be more strict in future if we want to be generic over DataSets. However, this is something that I don't think we will ever want to do.

[AtheMathmo]

We should feature gate this. My thinking is that if we have a few datasets users will not want to download all of this data by default.

To do this:

• Add a new feature to Cargo.toml
• In lib.rs add a feature flag

[sinhrks]

Added feature gates. Is it ok to be included by default ATM (as it is likely to be used in most tests)?

[AtheMathmo]

Thanks for the update.

It looks good but I'm a little cautious about having the datasets flag included by default. I wanted it feature flagged specifically so that it had to be opted-in. I can see that we will probably want to use it in some tests but I'd try a few ways around this first.

• Can we fit all tests in the datasets module?
• Can we somehow put the tests in the tests folder behind the feature flag too?
• As a last resort, do as you have done and keep the "datasets" flag for the "test" feature.

Finally note that we will need to modify the travis CI matrix to include the "datasets" flag.

[sinhrks]

OK, made datasets as optional, and fixed travis tests.

[AtheMathmo]

This looks good to me but before merging I'd like to check out the branch and play around with it a little.

Thanks!

[AtheMathmo]

I checked out the code and I have a few thoughts. I am happy to merge this in without any further changes but we should at least write up a tracking issue for improvements.

I think the description of load_iris should include more information on the dataset. Or a link directly to the iris dataset: http://archive.ics.uci.edu/ml/datasets/Iris (instead of just the /ml root). We should say what the attributes and classes are. From the link:

Attribute Information:

1. sepal length in cm
2. sepal width in cm
3. petal length in cm
4. petal width in cm
5. class:
   -- Iris Setosa
   -- Iris Versicolour
   -- Iris Virginica

Also I think it might be a good idea to organize the datasets module a little differently. If we add more datasets the module is going to get large quickly and difficult to manage. I think we should move the iris data into a new iris module. This iris module should have a pub fn load() -> DataSet. As a user I would be quite happy to do:

use rusty_machine::datasets:iris;

let (inputs, targets) = iris::load();

But we could also use the current format and have a load_iris function in the datasets module that calls iris::load. In this case we might want to keep the iris module private?

Let me know what you think. If you don't want to make any of these changes now I'll merge and move this information to a separate ticket.

[sinhrks]

Thx for the comment. I've did a change requested. Pls take a look when u have a time.

[AtheMathmo]

Looks good to me now!

I have a minor nitpick for the features section but so far as I can tell it makes no real difference.

[AtheMathmo]

We don't need to include the test or default features. These already exist as defined here.

[sinhrks]

Correct, removed.

[AtheMathmo]

This description is great!

[AtheMathmo]

Thank you! Merging now.