New feature: offline training analysis

[atruskie]

We're interested in incorporating learned models into our workflow, which we intend to use for a new class of recognizers. Here we use the term offline to denote that each instance of an offline training analysis will sample data from many files to build a model; it will not produce any classification results.

Conversely, online is the term we will use when processing one audio file to get predictions.

Expected behaviour

AP.exe will have a new entry point named "model-learning".

This entry point will accept:

• a list of samples to process (remotely downloaded from Ecosounds)
• A config file (which will determine the type of training to do)
• and an output directory to put the trained model and the model metadata

This analysis should have common options enabled that define:

• whether or not to save the features used to train the model
• whether or not to save the audio recording samples used to train the model

This analysis will implement the IAnalyzer2 interface. The interfaces stages will map as such:

Analyze is the map stage and will download each audio segment (using analysis coordinator), and do feature extraction.

SummarizeResults is the reduce stage and will be responsible for training and then producing the model.

This will be used by @mkholghi and @towsey - comments welcome.

[towsey]

This sounds like a good approach to me, although I am not across the finer points.