Implementation of models in paper Fast and Effective Neural Conversational Banking, if you use this code, please cite this work.
With CUDA 8.0:
$ conda create --name conv-bank python=3.6.1
$ source activate conv-bank
$ conda install pytorch torchvision cuda80 -c soumith
<!-- $ pip install ipdb -->
Without CUDA:
$ conda create --name conv-bank python=3.6.1
$ source activate conv-bank
$ conda install pytorch torchvision -c soumith
<!-- $ pip install ipdb -->
$ cd data
$ tar -zxvf samantha.tar.gz
$ cd ..
$ source data_handler/create_dataset.sh
- fasp
- redfasp
- bytenet
- rnn
- attn-rnn
- redmulnet-1
- redmulnet-2
- fasp --multi-linear
- redfasp --multi-linear
- bytenet --multi-linear
Training command:
$ python train.py --model <model> <--multi-linear> --epochs <max-epochs> --save-dir <save-path>
To train a single model:
$ python train.py --model fasp --epochs 20 --save-dir checkpoint/samantha/fasp
To train all models with default parameters:
$ source train_models.sh <max-epochs>
Evaluating command:
$ source eval_models.sh <path-to-models> <path-to-data>
To eval a single model:
$ source eval_models.sh checkpoint/samantha/fasp data/samantha/dataset.pckl
To eval all models:
$ source eval_models.sh checkpoint/samantha/ data/samantha/dataset.pckl