Proposal for mlflow integration about mlj.jl

[deyandyankov]

Thanks for putting this together, @ablaom.
A few notes and suggestions from my end.

Regarding logging=nothing and 1. Serializing machines, it might be useful to think of different types of arguments we might provide. As per MLFlowClient's reference, we have 3 main types that we can use as parameters regarding logging.
MLFlowClient.MLFlow - this is a type which we use to define an mlflow client, it is usually instantiated such as mlf = MLFlow("http://localhost:5000").
Then when we create an experiment and a run, it looks like this:

# Create MLFlow instance
mlf = MLFlow("http://localhost:5000")

# Initiate new experiment
experiment_id = createexperiment(mlf; name="experiment name, default is a uuid")

# Create a run in the new experiment
exprun = createrun(mlf, experiment_id)

I'll start with the simplest case described in the original post here:

Serializing machines: Calling MLJModelInterface.save(location, mach) whenever location is an mljflow experiment (instead of a path to file).

location could either be an MLFlow, MLFlowExperiment, or MLFlowRun.

The most obious case is when we provide an MLFlowRun. Runs belong to experiments and experiments belong to an mlflow instance. A single experiment may have 0 or more runs.

Thus, we could define:

• MLJModelInterface.save(location::MLFlowRun, mach) - save the machine as a serialized artifact in an existing run.
• MLJModelInterface.save(location::MLFlowExperiment, mach) - create a new run in an existing experiment and fall back to MLJModelInterface.save(location::MLFlowRun, mach)
• MLJModelInterface.save(location::MLFlow, mach) - create a new experiment in the provided location::MLFlow and fall back to MLJModelInterface.save(location::MLFlowExperiment, mach)

We can use similar logic when initiating logging from different places, such as performance evaluation, hyperparameter tuning, and controlled model iteration.

[pebeto]

@deyandyankov I bundled a MLFlow object inside a general MLFlowInstance type that allows us to save the most important project configurations: base_uri, experiment_name and artifact_location (we can expand them, it's just a draft). You can see more about that here. With that, there is no piece of code from MLFlowClient loaded at first glance. We need to import the library first to operate with the methods that logs our info; and if that's not the case, it's easy to send an error requesting things to user.

[ablaom]

Proposed behaviour

When should MLJ actions trigger mlflow logging?

It should be possible to request logging for these actions:

1. Serializing machines: Calling MLJModelInterface.save(location, mach) whenever location is an mljflow experiment (instead of a path to file).
2. Performance evaluation: Calling evaluate(mach, ...) or evaluate(model, data..., ...) for any mach/model (including composite models, such as pipelines)
3. Hyperparameter tuning: Calling MLJModelInterface.fit(TunedModel(model, ...), ...) for any model (and hence calling fit! on an associated machine)
4. Controlled model iteration: Calling MLJModelInterface.fit(IteratedModel(model, ...), ...) for any model(and hence calling fit! on an associated machine)

Moreover, it should be possible to arrange automatic logging, i.e., without explicitly requesting logging for each such action.

What should be logged?

1. Serializing machines

• the file ordinarily created by save(file, mach) should be instead saved as an mlflow artifact
• additionally, all hyperparameters (i.e., a suitably unpacked representation of model)

2. Performance evaluation

Compulsory

• all hyperparameters (i.e., a suitably unpacked representation of model)
• names measures (aka metrics) applied
• each corresponding aggregate measurement

And, if possible:

• the resampling strategy used (e.g., CV) and, if possible, it's parameters (e.g., nfolds)
• value of repeats (to signal possibility this is a Monte Carlo variation of resampling)

Optional (included by default)

• the explicit row indices for the each train/test fold pair

3. Hyperparameter tuning

Compulsory

For the optimal model:

• The same compulsory items in 2.

Optional (included by default)

For each model in the search (each hyperparameter-set):

• the same compulsory items as in 1, although it might suffice to only log hyperparameters
  that change during training

3. Controlled model iteration

Compulsory

For the final trained model (different from the last "evaluated" model, if retrain=true;
see
here)

• The same compulsory items in 2, plus a final training error, if available (not all
  iterative MLJ models report a training loss)

Optional (included by default)

For the partially trained model at every "break point" in the iteration:

• The same compulsory items in 2, plus a final training error, if available

• Serialization of the corresponding "training machine" (see
  docs),
  as an artifact.

How should logging be structured

I'm less clear about details here, but here are some comments:

• In tuning, each model evaluated should be a separate run within same experiment as the optimal model run

• Iteration would be similarly structured

• Since a model (hyperparameter set) can be nested (e.g, pipelines and wrappers), I
  suggest that a flattened version of the full tree of parameters be computed for purposes
  of logging, and suggestive composite names created for the nested
  hyperparameters. Possibly, we may want to additionally log model as a julia-serialized
  artifact??

User interface points

Some suggestions:

How does user request logging?

In serialization, one just replaces location in MLJBase.save(location, mach) with the
(wrapped?) mljflow experiment.

In performance evaluation, we add a new kwarg logger=nothing to evaluate/evaluate!
which user can set to an (wrapped?) mljflow experiment.

Cases 3 and 4 are similar but logger=nothing becomes new field of the wrapper
(TunedModel or IteratedModel structs).

How does user request auto logger?

Add global variable DEFAULT_LOGGER accessed/set by user with new methods
logger()/logger(default_logger) initialized to nothing in __init__, and change the
above defaults from logger=nothing to logger=DEFAULT_LOGGER.

How does user suppress optional logging?

We could eigher extra kwargs/fields to control level of verbosity, or if we are wrapping
experiments anyway, include the verbosity level in the experiment wrapper. I'm leaning
towards the latter (or just making everying compulsory).

[ablaom]

Some miscellaneous thoughts on implementation

1. A proof of concept already exists for performance evaluation. This shows how to add the new functionality using an extension
   module, which also forces us to keep the extension as disentangled from current
   functionality as much as possible, for better maintenance.

2. When a TunedModel is fit, it "essentially" calls evaluate! on each model in the
   tuning range, so we can get some functionality in that case by simply passing the
   logger parameter on. What actually happens is that fit wraps the model as
   Resampler(model, ...),
   which has fields for each kwarg of evaluate, this resampler gets wrapped as a
   machine, trained, and then a special evaluate method is called on this machine to get
   the actual evaluation object. So we also need to add logger to the Resampler struct
   (which is not public)

3. Some hints about how to flatten models appear
   here
   and here.

4. In IteratedModel we already have the
   Save
control. Currently the default filename is "machine.jls", but if !isnothing(logger)
   we could instead pass logger as the default. Then, we change the default for
   controls to include Save() if !isnothing(logger). I imagine something similar
   could be worked out for WithEvaluationDo and WithTrainingLossesDo to get the other
   information we want logged.

[ablaom]

cc @pebeto @deyandyankov @tlienart @darenasc