Author: Tianxiang Gao

Pytorch implementation of seq2seq French-to-English translation.

Run French-to-English translation on GPU:0, the hyperparameters are defined in the ./experiments/attention/f2e/config.json'

python main.py --gpu 0 --model_dir experiment/attention/f2e/

After the training, the best checkpoint will be stored in ./experiment/attention/f2e/ckpts/best.pth.tar. To evaluate the model on GPU:0, use the best checkpoint, beam search with width 3, and save attention weights heat map, simply run the following command:

python main.py --gpu 0 --model_dir experiment/attention/f2e/ --mode test --heatmap --beam_size 3

python main.py --gpu [gpu_id] --model_dir [model_dir]

In the experiments/attention/ folder, there are two configuration json files for French-English and English-French translation. It is easy to set different hyper-parameters and play around the model.

Optional arguments for training:

• --resume whether to resume training from check point

python main.py --gpu [gpu_id] --model_dir [model_dir] --mode test

Optional arguments:

• --beam_size [value] or --bs [value] whether use beam search. Greedy search if the value is 0. Otherwise, the value is the width of beam search.
• --heatmap or --hm whether to generate and store the attention weight heat map.

• Pytorch official seq2seq tutorial
• Stanford CS230 code examples