Updates

[pavlin-policar]

Description of changes

Range of minor updates.

• Fix initialization std from 1e-2 to 1e-4
• Change the default number of iterations to 750. Fixes Change default learning rate for large data sets #104
• Add auto option for learning rate
• Replace the ErrorLogger with a verbose flag. Fixes Add verbose input parameter #112
• Remove min_grad_norm parameter. I feel like a deprecation warning here is unnecessary since I have never in my life seen that stopping condition hit, or used anywhere ever.
• Fix KL divergence sometimes being reported as NaN. Fixes Non-consistent nan value for KL divergence #108

Includes

• Code changes
• Tests
• Documentation

[dkobak]

George found what looks like a problem with the learning rate = n/12 heuristic: KlugerLab/FIt-SNE#93