Bail out of training immediately if no training data batches are provided

[shuttle1987]

I propose we bail out immediately if no training data is found, currently we have this:

                    batch_gen = self.corpus_reader.train_batch_gen()
    
                    train_ler_total = 0
                    print("\tBatch...", end="")
                    for batch_i, batch in enumerate(batch_gen):
                        print("%d..." % batch_i, end="")
                        sys.stdout.flush()
                        batch_x, batch_x_lens, batch_y = batch
    
                        feed_dict = {self.batch_x: batch_x,
                                    self.batch_x_lens: batch_x_lens,
                                    self.batch_y: batch_y}
    
                        _, ler, = sess.run([self.optimizer, self.ler],
                                        feed_dict=feed_dict)
    
                        train_ler_total += ler

But if batch_gen provides nothing we have an error state where the code will break at some future unspecified point.

Instead I'm thinking we can do something like this:

                    batch_gen = self.corpus_reader.train_batch_gen()
    
                    train_ler_total = 0
                    print("\tBatch...", end="")
                    for batch_i, batch in enumerate(batch_gen):
                        print("%d..." % batch_i, end="")
                        sys.stdout.flush()
                        batch_x, batch_x_lens, batch_y = batch
    
                        feed_dict = {self.batch_x: batch_x,
                                    self.batch_x_lens: batch_x_lens,
                                    self.batch_y: batch_y}
    
                        _, ler, = sess.run([self.optimizer, self.ler],
                                        feed_dict=feed_dict)
    
                        train_ler_total += ler
                    else:
                        raise PersephoneError("No training data was provided, check your batch generation")

This would make it really clear that something has gone wrong and would immediately report the failure.

[shuttle1987]

This is related closely to #184 and #185

[shuttle1987]

There's actually a regression here because previously working test cases now fail such as this one: https://travis-ci.org/persephone-tools/persephone-web-API/jobs/441199123

[alexis-michaud]

"There's actually a regression here"

Shush! The world is watching. Instead of just 'regression', could you kindly refer to this as, let's say, deep regression, or (maybe even better) vanilla deep regression? In the profuse terminology of deep learning, there are plenty of terms that could come in handy as euphemisms 🙊

[oadams]

Sorry about that, I've changed it back. I shouldn't have been working on master directly.