This part of the repo contains my paper implementation and training code for german to english translation using the RNNsearch model. If you find any mistakes or stuff that I've could build better, please let me know.
Before you start training you should make sure, that you have installed all the packages from the requirements.txt and downloaded the spacy tokenizers:
python -m spacy download en_core_web_sm
python -m spacy download de_core_news_sm
If you want to use this for other languages you have to download the tokenizers you need and maybe make some minor changes in the config.py and dataset.py.
You find the custom configurations in config.py. If you want to change some hyperparameters you can create your own custom_config.yaml containing all hyperparameters you want to override.
python train.py --config="your_custom_config.yaml"
Currently I haven't implemented mixed precision of multi-GPU training.
I've trained a small model for 36 epochs so you can use it to test my code. You can find the model and the corresponding config file here.
Here are some references I used: