Neural Poetry Generation with Visual Inspiration

Final project for CMPT419: Neural Poetry Generation with Visual Inspiration

Poster & report

Report

Download dataset

python download_image.py

Dataset from link

Extract poem & img feature

python extract_feature -s poem
python extract_feature -s img

Output will be a dictionary of (id, feature) pairs where feature is a 512 dimension numpy vector.

Output will be saved to data/poem_features.plk or data/img_features.plk

Training

Before training, build vocabulary file, the vocabulary file will be saved to data/vocab.pkl

python vocab_builder.py 

Pretrain on Unim

python train_adversarial.py

Tune on multim

python train_adversarial.py --source multim

Use -h flag to see more options

Reference

Beyond Narrative Description: Generating Poetry from Images by Multi-Adversarial Training