This is a toy Japanese to English translater. It's based on the paper Attention is All You Need. It's very early days and is not suitable for anything but experimentation.
Just open translation.ipynb and run the steps. You will almost certainly need a GPU to train at all efficiently. There's a trained model stored in model.pt you can load if you don't want to full trained.
input="私は一日に100ユーロ稼ぎます。"
translated='BOSI am to earn to earnest under a day.EOS'
expected='I make 100 euros per day.'
input="すぐに諦めて昼寝をするかも知れない。"
translated='BOSWe want to keep to sleep on your day after lunch.EOS'
expected='I may give up soon and just nap instead.'
input="そんなことは起きないでしょう。"
translated="BOSDon't get it up so that.EOS"
expected="That sort of thing won't happen."
As you can see there are significant errors and issues with the translations. We probably need:
- More training. This was only 100000 training steps with a batch size of 16. You'd need a lot more to truly understand the structure of the two languages.
- More optimized architecture. This is absically the dumbest least optimized way to run this model. Using more optimized transformer pieces (like those provided by pytorch) would be better
- More training data. The data came from https://tatoeba.org/en/ which is of unknown quality (and my Japanese is beginner so I can't even verify the quality).