Skip to content

Latest commit

 

History

History
129 lines (88 loc) · 6.61 KB

README.md

File metadata and controls

129 lines (88 loc) · 6.61 KB

Update 12/20/2020

  1. We have uploaded an evaluation.sh script for your convenience to reproduce prediction and evaluation shown in Table 4. Please update your local repo, and please copy and paste e2e-metrics folder into the same directory as this script. This script will run Transformer Baseline, PAG, CNE_Enc, CNE_Dec prediction and evaluate the generated context automatically.
  2. Please run the following commands to get the prediction and scores.
$ git clone https://github.com/tuetschek/e2e-metrics
$ ./evaluation.sh
  1. Notice: Since for the transformer baseline model, it was an quite early experiment, so we deleted it. The transformer baseline model weights are trained again today, so the evaluation results is slightly different from Table 4. Besides, there is a typo in Table 4 in the report, the actual ROUGE-L score for the best CNE-Dec experiments is 0.6850, instead of 0.6500.

CS269 NLG Course Project: Sentence Controllable DA-based E2E Generation in Restaurant Domain

The code of our course project is modified based on https://github.com/UKPLab/e2e-nlg-challenge-2017

Updates:

  1. We found a critical error in the original code, which could possibly make the training wrong. (components/model/modules/attention/attn_bahd.py#68) We have fixed it in this project.

  2. Two more data pre-processing strategies are implemented. (A General Model for Neural Text Generation from Structured Data, End-to-End Content and Plan Selection for Data-to-Text Generation)

  3. Transformer model

  4. Sentence Control stratigies:

    • Predict and Generate (PAG)
    • Controllable NOS Embedding (CNE)

Background info

Project structure

We provide basic scripts and their utilities in this repository, along with some output files' content:

  • run_experiment.py: main script to run (please freeze this script since experiment configurations are declared in YAML file).
  • config/train_XXXX.yaml and config/predict_XXXX.yaml: configuration files to use with the script above.
  • components/: data_preprocessing (HAV, HIT, UKP), model (MLP, GRU, Transformer...), trainer, evaluator, and necessary utils.

Requirements

  • 64-bit Linux versions
  • Python 3 and dependencies:
    • PyTorch v1.2.0
    • Progressbar2 v3.18.1

Installation

  • Install Python3 dependencies:
$ conda install pytorch torchvision cuda80 -c soumith 
$ conda install progressbar2
  • Python2 dependencies are needed only to run the official evaluation scripts.

Running the experiments

Preparation

  • Step 1

For your convenience, we have setup multiple configuration files in config/.

  1. train_transformer.YAML and predict_transformer.YAML run experiments using standard Transformer model. Please freeze nos_option to be 0.
  2. train_transformer_CNE_Enc.YAML and predict_transformer_CNE_Enc.YAML run experiments using Transformer model with CNE NOS embedding in encoder input. Please freeze nos_option and nos_position to maintain the CNE_Enc structure.
  3. train_transformer_CNE_Dec.YAML and predict_transformer_CNE_Dec.YAML run experiments using Transformer model with CNE NOS embedding in decoder input. Please freeze nos_option and nos_position to maintain the CNE_Dec structure.
  4. train_transformer_PAG.YAML and predict_transformer_PAG.YAML run experiments using Transformer model with PAG NOS. Please freeze nos_option to maintain the PAG structure. (nos_position is not used in this method).
  5. train_XXXX.YAML and predict_XXXX.YAML files are required to be used in pairs.
  6. train_MLP.YAML and predict_MLP.YAML run experiments using standard MLP model.
  7. train_gru.YAML and predict_gru.YAML run experiments using standard gru model.
  8. train_tfenc.YAML and predict_tf.YAML run experiments using transformer as encoder model.
  9. data_module can be chosen from e2e_data_hav, e2e_data_hit, and e2e_data_MLP, do not forget to modify corresponding parameters in configuration files following the instruction within them.
  10. Constraint: NOS related terms are specifically designed for Transformer model.
  11. For other model configuration parameters including embedding size, hidden size, input size, etc., if your setting confirms the instruction within the configuration file, the script can execute correctly.
  • Step2

Please first download e2e-metrics toolkit from 3. We are going to use the measure_score.py script to evaluate the quality of the generated sentence.

Training models

  1. Adjust data paths and choose a configuration file or use your own defined YAML (train_transformer.yaml, as a running example).

  2. Run the following command:

    $ python run_experiment.py config/train_transformer.yaml

  3. After the experiment, a folder will be created in the directory specified by the experiments_dir field of train_transformer.yaml file. This folder should contain the following files:

    • model weights and development set predictions for each training epoch (weights.epochX, predictions.epochX)
    • a csv file with scores and train/dev losses for each epoch (scores.csv)
    • configuration dictionary in json format (config.json)
    • pdf files with learning curves (optional)
    • experiment log (log.txt)

  4. If you use a model for prediction (setting "predict" as the value for the mode field in the config file and specifying model path in model_fn), use corresponding predict_transformer.yaml or set the configuration parameters to be the same as your training configuration if you want to use the self-defined configuration. the predictions done by the loaded model will be stored in:
    * $model_fn.devset.predictions.txt
    * $model_fn.testset.predictions.txt

Evaluation

Evaluate the generated context using following command:

```
$ cd e2e-metrics
$ python measure_scores.py <ref> <generated>
$ #Eg.
$ python measure_scores.py e2e-dataset/testset_w_refs.csv.multi-ref exp/e2e_model_transformer_seed1-emb256-hid512-drop0.05-bs128-lr0.0002_2020-Dec-17_23.53.25/weights.epoch1.testset.predictions.txt.
```

This script will calculate BLEU, NIST, CIDEr, ROUGE-L, and METEOR.