README

Implementation of Distinctive Slogan Generation with Reconstruction.

Installation

This repo requires python3 and pip

git clone https://akshaybhatia10@bitbucket.org/akshaybhatia10/enc-dec-baseline.git
cd enc-dec-baseline
mkdir outputs/

Install libraries

pip3 install -r requirements.txt

The datasets are individual CSV files for training, testing and validation. These must be in the datasets/ folder in the this repo. The trained model is saved in the outputs/ folder.

Arguments

Below are some of the arguments with their default values required by the training script

• enc_dec_hidden: Number of hidden units for encoder decoder GRUs - int, 512
• bs: Batch size - int, 8
• epochs: Number of epochs - int, 10
• dataset_path: Path to dataset - must have train.csv, test.csv, valid.csv - str, datasets
• vocab_size: Size of vocab - int, 30000
• embed_size: Word Embedding size - int, 200
• copy: Whether to enable copy mechanism or not - store_action i.e True if included else falses
• recons: Whether to enable reconstruction model or not - store_action i.e True if included else falses
• evaluate: Evaluate the model using the pretrained model - store_action, i.e True if included else falses

Usage

To train the encoder-decoder model only

python3 main.py --enc_dec_hidden 512 --bs 8 --epochs 10 --dataset_path datasets --vocab_size 30000 --embed_size 200

To train the encoder-decoder model with copy mechanism only

python3 main.py --enc_dec_hidden 512 --bs 8 --epochs 10 --dataset_path datasets --vocab_size 30000 --embed_size 200 --copy

To train the encoder-decoder model with copy mechanism and reconstruction model - 2020 paper

python3 main.py --enc_dec_hidden 512 --bs 8 --epochs 10 --dataset_path datasets --vocab_size 30000 --embed_size 200 --copy --recons

To evaluate the trained model:

python3 main.py --enc_dec_hidden 512 --bs 8 --epochs 10 --dataset_path datasets --vocab_size 30000 --embed_size 200 --copy --recons --evaluate

Attention Visualization

Below is a visualization of the generated slogan with its description. The below command will prompt an input in the terminal. Enter your text description and press q to exit.

python3 visualize_attention.py

Evaluation results

• Encoder-Decoder only 2015 paper

  • Test - R1: 16.74, R2: 4.67, RL: 15.77
  • Valid - R1: 16.45, R2: 4.89, RL: 15.58

• Encoder-Decoder with copy mechanism only

  • Test - R1: 17.22, R2: 4.48, RL: 16.31
  • Valid - R1: 17.27, R2: 4.65, RL: 16.37

• Encoder-Decoder with copy mechanism and reconstruction model - 2020 paper

  • Test - R1: 19.32, R2: 5.86, RL: 18.36
  • Valid - R1: 19.35, R2: 5.98, RL: 18.36