SMAC-integration

[ravinkohli]

Major changes include

1. Added autopytorch/evaluation where 1.1. tae.py contains the ExecuteTAFuncWithQueue. 1.2. train_evaluater.py contains eval_function closure to be used in ExecuteTAFuncWithQueue. Currently, only tested with holdout and cross validation on tabular tasks.
2. Changed way of creating splits in BaseDataset. now it will initialise splits attribute of the BaseDataset.
3. Made dataset_properties as a key in fit dictionary which contains output of datamanager.get_dataset_properties
4. Added predict_proba function to the pipeline

[franchuterivera]

I recommend the following test to be added:

• Test that we can fit a dummy classifier

• Test that we can fit a random forest with this and that the result is 1 (when evaluated with the same data used for training)

• Test that we can capture an error. Make a fit that crash so we see what happens when you crash a fit configuration

• We have to make sure the predictions saved to disk are for all the data. check the size of the predictions saved by finish_up

[ravinkohli]

I recommend the following test to be added:

* Test that we can fit a dummy classifier

* Test that we can fit a random forest with this and that the result is 1 (when evaluated with the same data used for training)

* Test that we can capture an error. Make a fit that crash so we see what happens when you crash a fit configuration

* We have to make sure the predictions saved to disk are for all the data.  check the size of the predictions saved by finish_up

Your second point means that we can fit a random forest using the ExecuteTaeFuncWithQueue, without having smac or the evaluator, right?

[franchuterivera]

Your second point means that we can fit a random forest using the ExecuteTaeFuncWithQueue, without having smac or the evaluator, right?

Yes, so that we make sure that your code in isolation works no matter what.