This is a testing code base on pytorch framework. It is a seq2seq translation model and attention mechanism was applied. The model trains a neural network to translate from French to English.
- Load data from data/eng-fra.txt, which includes English-French language translation pairs.
- Pre-process language pairs, and train a Seq2seq model based on Recurrent Neural Network.
- Evaluate translation for random sentences.
Results of this model include:
- Training and evaluation result log.
E.g.
[KEY: > input, = target, < output]
> nous sommes armees .
= we re armed .
< we re ruined . <EOS>
> je suis gras .
= i m fat .
< i m fat . <EOS>
> il est tres courageux .
= he is very brave .
< he is very brave . <EOS>
> nous regrettons de ne pas pouvoir vous aider .
= we re sorry we can t help you .
< we re sorry we can t help you . <EOS>
> je suis un tel idiot .
= i m such a fool .
< i m a little fool . <EOS>
> il est tres deprime .
= he is very depressed .
< he is very depressed . <EOS>
> je garde ce siege pour tom .
= i m saving this seat for tom .
< i m doing this for tom . <EOS>
- The entire task-processing time on Orion platform takes around 20 minutes on single 1080Ti GPU.
- Nebula AI working node will save all results in the directory Result, users will be able to retrieve contents from result directory, together with system execution log.
- For more details on how-to-guides for task submission, please refer to instructions on Nebula AI developer portal.