Skip to content

Latest commit

 

History

History
60 lines (45 loc) · 2.55 KB

README.md

File metadata and controls

60 lines (45 loc) · 2.55 KB
Raw

This repository is for training and testing slot filling models from various tasks. Its goal is to provide you following advantages:

  • Dynamic Keras model creation by specifying your architecture in config.yaml file. See example/configs for sample configs.
  • Easy to add Pytorch and Tensorflow available models by providing basics like data generators.
  • Easy model evaluation.
  • Adding different tasks to train were never been easy before.

Creating Environments

For Keras/Tensorflow models

If you want to create, train and evaluate Keras/Tensorflow models, create this environment.

conda env create -f tensorflow_env.yml

For Pytorch models

If you want to create, train and evaluate Pytorch models, create this environment.

conda env create -f torch_env.yml

Training and Evaluating Models

sample scripts to train and evaluate models are available in ./scripts directory. Just prepare your task, specify your cuda version in LD_LIBRARY_PATH and GPUs to use in CUDA_VISIBLE_DEVICES in your script and the rest is the same as sample .sh files. Use command chmod 750 YourFile.sh to set required permissions, then execute it.

Adding Tasks

To add your task,

  • Put your dataset templates files in ./data dir.
  • Create a package for it in ./main/tasks/ package.
  • Create a YourTask class and inherit Task class. Implement required methods to load templates.
  • Implement your task util in the same format as BuyCharge and HotelReservation tasks.
    • SLOT: is a dictionary that maps slots in templates to the generator functions for them.
    • WORD_FILLER: is a dictionary that maps placeholders in templates to the generator functions for them.
    • ALL_WORDS: is a list of all words used in your generators and available words in your dataset.
  • Don't forget to also add your task in tasks/loader.py module functions.

Models

Currently only three type of models are available:

  • RNNEncoder
  • TransformerEncoder
  • ConvSeq2Seq from Convolutional Sequence to Sequence model paper.

Configs parameters

  • name: name of your model. (RNNEncoder, TransformerEncoder, ConvSeq2Seq)
  • task: name of your task. (BuyChargeTask, HotelReservationTask)
  • seed
  • model: your model architecture stays here.
  • model_params: your model training parameters stays here.
  • data.labels: all labels we are predicitng. Note: PAD label and O "not slot" label must be 0 and 1 respectively, and <sos> and <eos> labels are also must be in the list, even if we are not using them.

see one of sample config files for more comments.