Dropout not used correctly in machine_translation example

[aalmah]

In this example, dropout is effectively ignored because:

• The cost used in GradientDescent should be cg.outputs[0] rather than 'cost':
  https://github.com/mila-udem/blocks-examples/blob/12013e6c737171ce285b27381f0f5aacfdff7cab/machine_translation/__init__.py#L160
• Also the cost used in TrainingDataMonitoring:
  https://github.com/mila-udem/blocks-examples/blob/12013e6c737171ce285b27381f0f5aacfdff7cab/machine_translation/__init__.py#L112
  is not the dropped out one, so instead one should use here cg.outputs[0].

[dwf]

@orhanf thoughts on this?

[orhanf]

The first item should definitely be fixed, but for the second one, there are some differences between implementations. What we usually monitor is the bleu score or log-prob of the validation set or a subset of the training set. And for both, we use models without any regularization. So, I think, the first item should be fixed, the second can be optional.