Estimating predictions quality and expose that to the user

This is an MR regarding the issue #67 (closed) .

Regarding the estimation predictions changes :

• Data are splited to train and test dataset (80%-20%).
• Only the train dataset is passed to the fit method.
• We make predictions.
• We compute the quality using RMSE.
• The results are displayed during the process using the LOGGER.

Regarding the exposition to the user:

• I added the mlflow package to the solution.
• I used the logging functions to expose the results of the predictions.
  1. The log_param is used for logging the context of the flow (used model, hyper parameters, ...)
  2. The log_metric is used for the results.
  3. The set_experiment is used to create an experiment with the data source name (sat name for the json file and file name if the data are from a csv file)
  4. The start_run is used to name the actual run to cross_correlate because it s the ML process running.

WIP or Not WIP ?

Well it depends if i have to do more configuration on mlflow in this MR. Because yes Mlflow need to be fine tuned, i'm thinking for example :

• Making a good use of the experiment feature.
• The place where the mlflow's db is created need to be well configured, at this time, the folder is created from where Polaris is running. I think the best way to do that is using the configuration file created by @saintaardvark .
• We need to find a way to configure mlflow also to use a external server (deepchaos) for making results available online for testing, demo, ...

I would be really happy to do that, but does it has to be covered by this MR ? in my opinion, it's a little bit out of topic but, it's a team call :p