Plug-and-Play Recipe Generation with Content Planning

This repo contains code and model checkpoints of our EMNLP 2023 GEM Workshop paper:
Yinhong Liu, Yixuan Su, Ehsan Shareghi and Nigel Collier "Plug-and-Play Recipe Generation with Content Planning".
If you hae any questions, please contact Yinhong via yl535 at cam.ac.uk. If you find our paper and resources useful, please kindly leave a star and cite our papers. Thanks!

@article{liu2022plug,
  title={Plug-and-Play Recipe Generation with Content Planning},
  author={Liu, Yinhong and Su, Yixuan and Shareghi, Ehsan and Collier, Nigel},
  journal={arXiv preprint arXiv:2212.05093},
  year={2022}
}

Generator, Classifier and Planner

Our trained model checkpoints could be downloaded by:

from transformers import BartForConditionalGeneration, GPT2LMHeadModel, AutoModelForSequenceClassification
from transformers import BartTokenizer, GPT2Tokenizer, AutoTokenizer

planner = BartForConditionalGeneration.from_pretrained('yinhongliu/recipe_with_plan_bart_planner')
planner_tokenizer = BartTokenizer.from_pretrained('yinhongliu/recipe_with_plan_bart_planner')

classifier = AutoModelForSequenceClassification.from_pretrained('yinhongliu/recipe_with_plan_distilbert_classifier')
classifier_tokenizer = AutoTokenizer.from_pretrained('yinhongliu/recipe_with_plan_distilbert_classifier')

generator = GPT2LMHeadModel.from_pretrained('yinhongliu/recipe_with_plan_gpt2_generator')
generator_tokenizer = GPT2Tokenizer.from_pretrained('yinhongliu/recipe_with_plan_gpt2_generator')

Or you can also train them from scratch:

Fine-tuning Generator

The GPT2 generator could be finetuned with following shell command. (Adjust the batch_size and data/model path according to your situations.)

python3 train_generator.py --model_name='gpt2'\
                            --data_path='PATH_TO_PREPROCESSED_RECIPE1M_DATA/'\
                            --save_path='PATH_TO_SAVE_MODEL_CHECKPOINT/' \
                            --max_len=512\
                            --batch_size=16\
                            --epoch_number=4\
                            --warmup_steps=200\
                            --print_steps=1000\
                            --save_steps=30000\
                            --eval_steps=10000\
                            --lr=8e-5\

Training Classifier

python3 stage_classifier.py --preprocessed_data_path='PATH_TO_PREPROCESSED_RECIPE1M_DATA/' \
                            --model_saving_path='PATH_TO_SAVE_MODEL_CHECKPOINT/' \
                            --lr=8e-5 \
                            --l2_decay=0.01 \
                            --epoch=3 \
                            --batch_size=300 \
                            --save_steps=15000 \
                            --eval_steps=10000 \
                            --warmup_steps=500

Training Planner

python3 stage_planner.py --preprocessed_data_path='PATH_TO_PREPROCESSED_RECIPE1M_DATA/' \
                        --model_saving_path='PATH_TO_SAVE_MODEL_CHECKPOINT/' \
                        --train=True \
                        --lr=8e-5 \
                        --l2_decay=0.01 \
                        --epoch=4 \
                        --batch_size=256 \
                        --save_steps=2000 \
                        --eval_steps=2000 \
                        --warmup_steps=200

Generate Content Plan with a trained planner

python3 stage_planner.py --preprocessed_data_path='PATH_TO_PREPROCESSED_RECIPE1M_DATA/' \
                        --model_saving_path='PATH_TO_SAVE_MODEL_CHECKPOINT/' \
                        --trained_model_path='PATH_TO_LOAD_MODEL_CHECKPOINT_FOR_PREDICTION/' \
                        --predict=True \
                        --batch_size=256 \

Evaluation Examples

The evaluation results are shown in evaluation.ipynb

RecipeWithPlans

Commits

ee2

e2

evaluation

generator

update generator