More informative logs

[laoliu5280]

For our hypothesis generation pipeline, the current logger outputs the following information:

• Generated hypotheses for each update
• Training time t
• Hypothesis-based inference accuracies
• All hyperparameters

To improve this logger, we need to include other information, such as: (list is not exhaustive)

• When a training example at time t is labeled as "wrong example"
• Model input and outputs (preferably in DEBUG)