This is a pytorch implementation of the Transformer model like tensorflow/tensor2tensor.
I tested it with PyTorch 1.0.0 and Python 3.6.8.
It's using SpaCy to tokenize languages for wmt32k
dataset. So, if you want to run wmt32k problem which is a de/en translation
dataset, you should download language models first with the following command.
$ pip install -U -r requirements.txt
$ python -m spacy download en_core_web_sm
$ python -m spacy download ja_core_news_sm
- Train a model. If you choose either wmt32k or lm1b for your problem set, comment it out if you do not want to use it. After commenting out, execute the following command
source train.sh
If you want to try fast_transformer, give a model argument after installing
tcop-pytorch.
$ python train.py --problem lm1b --output_dir ./output --data_dir ./lm1b_data --model fast_transformer
- You can translate a single sentence with the trained model.
$ python decoder.py --translate --data_dir ./wmt32k_data --model_dir ./output/last/models