Experiment

An experiment propagates DataSets (encapsulated by a DataSource) through a Model via method run. The specifics of such propagations are handled by Propagators. The propagation of a DataSet is called an epoch. At the end of each epoch, a monitoring step is performed where reports are generated for all Observers.

The Experiment keeps a log of the report of the experiment after every epoch. This is done by calling the report method of every contained object, except Observers. The report is a read-only table that is passed to Observers along with the Mediator through Channels for which they are Subscribers. The report is also passed to sub-objects during propagation in case they need to act upon data found in other branches of the experiment tree.

dp.Experiment{...}

An Experiment constructor which takes key-value arguments:

  • id is an ObjectID uniquely identifying the experiment. Defaults to using dp.uniqueID().
  • model is a Module instance shared by all Propagators.
  • optimizer is an Optimizer instance used for propagating the train set.
  • validator is an Evaluator instance used for propagating the valid set.
  • tester is an Evaluator instance used for propagating the test set.
  • observer is an Observer instance used for extending the subject Experiment.
  • random_seed is a number used to initialize the random number generator. Defouts to 7.
  • epoch at which to start the experiment. This is useful reusing trained Models in new Experiments. Defaults to 0.
  • mediator is a Mediator, a Singleton, used for inter-object communication. Defaults to Mediator().
  • overwrite is a boolead with a default value of false used to overwrite existing values. For example, if a datasource is provided, and optimizer is already nitialized with a dataset, and overwrite is true, then optimizer would be setup with datasource:trainSet().
  • max_epoch sets the maximum number of epochs allocated to the experiment. Defaults to 1000.
  • description a short description of the experiment.

The only mandatory arguments are model and at least one optimizer, validator or tester.

run(datasource)

This method loops through the propagators until a doneExperiment is received or experiment reaches max_epoch epochs. The datasource is provided as an argument here, as opposed to being passed to the constructor to keep it from being serialized with the Experiment.

includeTarget(mode)

When mode is true (the default), the targets will be included in the input to every model. Such that the input will have the form :

input = {input, target}

The default behavior for the Experiment is to not include the targets in the input. This is useful for modules like SoftMaxTree which require the target be fed in as well as the input.

[report] report()

This method returns a report table using for monitoring the components of the Experiment. This report includes reports (in attributes of the same name) for the optimizer, validator, tester and model. It also includes attributes for the current epoch, random_seed and id.

verbose([on])

Toggle the verbosity of all objects in the experiment. When on is true (the default), objects will print messages.

silent()

Calls self:verbose(false).