Update documentation to stop mentioning interfaces that no longer exist

[antoniovs1029]

So some weeks ago I noticed that the documentation under docs/code mention several interfaces that currently don't exist in ML.NET's codebase. They existed back in TLC... and perhaps some of them, if not all of them, actually made it into ML.NET before version 1.0 release... but the case is that they no longer exist, so I just wanted to update them.

After looking into the TLC code, I decided to make the following replacements in the docs:
ISchema -> DataViewSchema
IRandom -> System.Random
IRow -> DataViewRow
IRowCursor and ICursor -> DataViewRowCursor
IDataReader -> IDataLoader

Side notes

1. About IRandom: In TLC it seems that only SysRandom and TauswortheHybrid implemented IRandom. In ML.NET IRandom and SysRandom don't exist, but TauswortheHybrid inherits from System.Random. So I guess it's safe to replace IRandom for "System.Random" in the docs.

[antoniovs1029]

So, any ideas about what has replaced IDataReader? It's also mentioned in the docs, but I am not totally sure which would be the "equivalent" in ML.NET.

https://github.com/dotnet/machinelearning/blob/master/docs/code/MlNetHighLevelConcepts.md#data-reader #Resolved

[justinormont]

I see one usage of and various references to IRowCursor -- https://github.com/dotnet/machinelearning/search?utf8=%E2%9C%93&q=IRowCursor&type=

Example:

machinelearning/src/Microsoft.ML.Data/Evaluators/BinaryClassifierEvaluator.cs

Line 1540 in 610ffcb

var cursors = new IRowCursor[prList.Count];

#Resolved

[antoniovs1029]

Yes, it seems it is the only place in "actual code" where IRowCursor appears.

It is wrapped inside a #if !CORECLR (link). So when building ML.NET nothing inside that block is compiled. If I remove the CORECLR if directive, then I get some errors, not only corresponding to the IRowCursor (which doesn't exist currently in ML.NET).

I don't know if I should change it, then. #Resolved

[yaeldekel]

The code in BinaryClassifierEvaluator will not compile even if we change IRowCursor to DataViewRowCursor, since it uses a class that does not exist in our code: XYPlot. We had that class in an internal repo, and it used System.Windows.Forms which I believe is only for .Net Framework, so we didn't include it in ML.NET. I think that code can safely be deleted.

---

In reply to: 368201681 [](ancestors = 368201681)

[antoniovs1029]

I will open another PR to remove that piece of code then, since I prefer that this PR remains only as changes in the documentation. Thanks for the clarification, @yaeldekel !

---

In reply to: 368281931 [](ancestors = 368281931,368201681)

[antoniovs1029]

I've opened the PR in here: #4694

---

In reply to: 369266609 [](ancestors = 369266609,368281931,368201681)

[antoniovs1029]

Since OneToOneTransformBase currently doesn't inherit from DataViewSchema or DataViewRowCursor, I guess it's safe to remove this. #Resolved

[yaeldekel]

This comment is intended to explain that classes deriving from OneToOneTransformBase do not need to worry about creating a schema or a cursor, since everything necessary is already implemented in OneToOneTransformBase. This is still true even when we don't have the ISchema and IRowCursor interfaces:

• It has the class deriving from DataViewRowCursor that is used by all classes deriving from OneToOneTransformBase.
• It's true that it doesn't implement DataViewSchema (since there is only one schema class, it is not abstract), but it does have a Bindings class, that has an AsSchema property.

---

In reply to: 368201579 [](ancestors = 368201579)

[antoniovs1029]

I see. So I've added some more lines to this comment to explain what you've just said in here. Please let me know if it's clear enough.

---

In reply to: 368283641 [](ancestors = 368283641,368201579)

[yaeldekel]

Thanks. You can mention that by "don't need to worry about" we mean that the abstract property OutputSchema and the abstract method GetRowCursorCore of TransformBase are implemented in this class. Other than that looks clear enough.

---

In reply to: 369197668 [](ancestors = 369197668,368283641,368201579)

[yaeldekel]

IDataReader is the old name of IDataLoader.
Since the loaders don't actually load anything until you get a row cursor (not even when you call Load...), we initially thought we should make a distinction, but we ended up calling them loaders anyway.

---

In reply to: 575820928 [](ancestors = 575820928)

[antoniovs1029]

I've just seen that PR #2731 actually explicitly states that IDataReader became IDataLoader. Some docs were updated on that PR, but I guess that they've missed updating the MlNetHighLevelConcepts.md file (which I've just updated in here)

There are also some mentions of IDataReader on the following spec:
https://github.com/dotnet/machinelearning/blob/master/docs/specs/mlnet-database-loader/mlnet-database-loader-specs.md

But after discussing it with @harishsk, it seems that it might be better to remove that file (although I would still need to check it with other people that worked on the database loader).

---

In reply to: 575986788 [](ancestors = 575986788,575820928)

[antoniovs1029]

Ok, so after further discussion there's no need to remove the mlnet-database-loader-specs.md file.

Evenmore, it turns out that the IDataReader mentioned in there isn't ML.NET's old IDataReader, but actually System.Data.IDataReader, so I will only make a minor change in that file to make it clear they are referring to that interface.

---

In reply to: 576902800 [](ancestors = 576902800,575986788,575820928)

[yaeldekel]

[justinormont]

LGTM