ML.NET command line tool

[Zruty0]

I think we should consider making ML.NET command-line tool an actual first-class citizen.

A bit of insider knowledge: we currently already have a commandline tool, that is a port of maml.exe. You can launch it as follows (this shows the 'help' command, or the ? command).

dotnet .\bin\AnyCPU.Debug\Microsoft.ML.Console\netcoreapp2.1\MML.dll ?

Generally, the syntax is dotnet MML.dll <command> <arguments>, and the current list of commands include things like train, CV, showdata etc. (full list is available via dotnet MML.dll ? kind=command).

This command-line tool is actually very powerful (although the language is clunky).

We could easily expand the command-line tool to handle common programming sub-tasks, like:

• Generate prediction code from model (actually, we already have codegen, but it needs some retouch to work with the new API);
• Inspect the contents of the model (which transformers are there and how they are set up);
• Introspect into the TensorFlow or other external model format (Microsoft.ML.Tensorflow could have a commandline interface to list the nodes of the graph;
• etc.

[justinormont]

I wholeheartedly agree with this proposal.

What would it take to bring the command-line to be a first-class citizen?

Work topics I see:

• Documentation of the command-line syntax & components
• Provide example code (training, scoring, feature importance, etc)
• Describe the interaction & interoperability with other APIs
• Expand functionality (as noted above: TensorFlow/ML.NET model inspectors, improve codegen, etc)

Additional background information: This is the language used in many of our unit tests and benchmarks.

[Zruty0]

The biggest problems with the command-line tool are that:

• It's so involved that we essentially invented a new language for it, that has to be learned;
• This language is not documented anywhere;
• Some of the 'features' of the language are dubious at best.

As far as I'm concerned, in order to make it a first-class citizen, we need to shrink the functionality, not expand :)

For instance, our kool kustom subcomponents should become JSONs (and ideally reside in a file, not in the command line).

[justinormont]

What kind of kool kustom sub-components would you convert to JSONs?

I'd cut down the lesser used functionality. This could be larger full concepts/components, down to even small syntax options, like the various ways to specify input/output columns.

Designed in Windows, we hit oddities in OSX/Linux like brace expansion. MAML has a braces+commas syntax. Bash's brace expansion rewrites xf=Concat{col=Features:A,B,C} to xf=Concatcol=Features:A xf=ConcatB xf=ConcatC before our syntax parser, which errors. This can be avoided by using double quotes around the whole command, or putting the command in a .rsp file.

Thoughts on the command-line from a data science perspective

Likes:

• self-contained pipelines/commands
• portable (copy/paste a pipeline between machines)
• terse
• power
• quickly editable / scriptable
• runner is pre-compiled & portable (no visual studio per model)
• output is verbose and helpful
• simple tasks are easy; harder tasks scale linearly

Dislikes:

• complexity (background knowledge to be productive)
• discovery (CL, unlike a GUI, doesn't lend itself to exploration of pipelines by suggesting similar components/methods to try next, and a trivial way to iterate over options)
• duplication (e.g. FastTreeClassification, FastTree, ft, ftc, FastRankBinaryClassification, FastRankBinaryClassificationWrapper, FastRankClassification, fr, btc, frc, fastrank, and fastrankwrapper are all the same component)
• constraints (e.g. can't represent a DAG of dataflows to send 10% of data to a tree based featurizer that's memory bound, and stack its output in to a final linear learner using all data; there are hacks for these but not clean methods)
• failure modes (e.g. errors are found when running)

[eerhardt]

I totally agree a command line tool is super useful/valuable.

Do you consider this a "must have" in v1? Or is it something that could be productized in a version after the initial release?

[artidoro]

Adding @GalOshri to the conversation.

[jwood803]

I would love to help with this. 😄

@Zruty0 @justinormont @eerhardt Is this already in the repository or is it something that will be released later?

[eerhardt]

@jwood803 - That would be great.

The command line tool is already in this repository. https://github.com/dotnet/machinelearning/tree/master/src/Microsoft.ML.Maml

One place that could use some help/attention is to create a .NET Global Tool out of the console application.

[Zruty0]

@jwood803 it's already in the repository.

I think doing what @eerhardt suggested would be a great first step.
After that, I would really love to have these things happen:

1. update codegen command to correctly handle model files trained by ML.NET (as opposed to those trained via train command), and generate the scoring code that is in line with the latest API.
2. Take a stab at documenting our command line language. @justinormont can give some really bizarre 20-line command lines to demonstrate the power of it :)

[justinormont]

Here's one example for classification of GitHub issues:

dotnet MML.dll TrainTest test=corefx_issues_TEST.tsv eval=MultiClassClassifierEvaluator{topkacc=3 nccf=40 opcs=+} data=corefx_issues_TRAIN.tsv loader=TextLoader{quote=- sparse=- col=ID:TX:0 col=Label:TX:1 col=Title:TX:2 col=Description:TX:3 header=+} xf=Term{col=Label} xf=CopyColumns{col=Name:ID} xf=TextTransform{col=FeaturesTextTitle:Title tokens=+ wordExtractor=NGramExtractorTransform{ngram=2} charExtractor=NGramExtractorTransform{ngram=4 all=-}} xf=TextTransform{col=FeaturesTextDescription:Description tokens=+ wordExtractor={}} xf=Concat{col=FeaturesText:FeaturesTextTitle,FeaturesTextDescription} xf=TrainScore{tr=LightGBMMulticlass feat=FeaturesText} xf=Concat{col=FeaturesTrainScoreLGBMOnNGrams:Score} xf=WordEmbeddingsTransform{col=FeaturesWordEmbOnDesc:FeaturesTextDescription_TransformedText model=GloVe50D} xf=TrainScore{tr=OVA{p=AveragedPerceptron{iter=10}} feat=FeaturesWordEmbOnDesc} xf=Concat{col=FeaturesTrainScoreAPOnWordEmbOnDesc:Score} xf=Concat{col=Features:FeaturesTrainScoreLGBMOnNGrams,FeaturesTrainScoreAPOnWordEmbOnDesc,FeaturesWordEmbOnDesc,FeaturesText} tr=OVA {p=AveragedPerceptron{iter=10}} out={c:\output\tutorial3\05-LearnerSweep-TrainScoreOnNGrams,_TrainScoreOnWordEmbGloVe50DOnDesc,_FeatsFromNGramsAndWordEmb\0.model.zip}

The DAG of components looks like:

@Zruty0 -- apologies, this one isn't even bizarre