## Key takeaways 
 - I have experimented with different model architectures and found this network best
 - Don't blindly put dropouts in every layers
 - Learning rate scheduler is very good if your training loss is not converging
 - Don't train for too many epochs in small dataset like this