This folder contains some examples and tutorials showing how the library works.

WARNING: Almost all variables and parameters in tutorials representing model hyperparameters have been intentionally hardcoded, aiming to facilitate readability. On a real execution, these values are taken from the config.py file.

There are 2 colab notebooks:

• An introduction to a complete NMT experiment:
• A dissected NMT model:

These tutorials basically are a split version of the execution pipeline of the library.

The translation task is EuTrans (Amengual et al.), a toy-task mainly used for debugging purposes. Since it is an easy task, the model should converge in few (~6-8) epochs, yielding a BLEU-4 score close to 1 in the case of the Es->En language direction.

The configs folder contains two config files, which declare larger, "real-life" models. They are shown as a reference for more complex tasks, but they should be tuned and modified for each task.

In the documentation folder you'll some docuementation files:

• neural_machine_translation.pdf contains a (theoretical) overview of an attentional NMT system.
• staged_keras_wrapper.pdf contains the autogenerated documentation for the staged_keras_wrapper package (at 30/Nov/2016).
• typical_output.md describes the files and output generated during a typical training run.
• ensembling_tutorial.md shows the usage of the sample_ensemble script.