jbdatascience/recipegpt-exp

RecipeGPT: Generative Pre-training Based Cooking Recipe Generation and Evaluation System (TheWebConf'2020; WWW'20)

Overview

This is the github repo for

Helena H. Lee, Ke Shu, Palakorn Achananuparp, Philips Kokoh Prasetyo, Yue Liu, Ee-Peng Lim, and Lav R. Varshney. 2020. RecipeGPT: Generative Pre-training Based Cooking Recipe Generation and Evaluation System. In Companion Proceedings of the Web Conference 2020 (WWW 20 Companion), April 20-24, 2020, Taipei, Taiwan. ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/3366424.3383536

Our paper can be downloaded at arxiv

Our poster can be downloaded at here

We also provide an online website that allows the users to generate cooking recipes at https://recipegpt.org/

Please contact [email protected] if you have any questions or problems.

We forked the this implementation in order to train and fine-tune gpt-2.

Google colab

We create a notebook google-colab-notebook to assist you re-produce our work. We show examples of model training, model inference, and model evaluation. It's a simplified version of our notebook 6, 7, 9

Project Structure

By default, the project assumes the following directory structure:

+-- data                                    # Files that are within GitHub's file size limit
¦   +-- vocab.bin                           # A word embedding model, will be used in utils.tree
¦   
¦   +-- recipe1M_example                    # Examples of testing data, n=5
¦       +-- test/X                          # Inputs in testing set (directory of .txt files)
¦       +-- test/y                          # Human-written outputs in testing set (directory of .txt files)
¦   
¦   
+-- big_data                                # Files that exceeds GitHub's file size limit
¦   +-- layer1.pickle                       # Recipe1M from http://pic2recipe.csail.mit.edu/  (Please click Layers (381 MiB))
¦   +-- recipe1M_ny.pickle                  # Recipe1M processed with ny-times-parser        
¦   +-- data.pickle                         # The data after our own data pre-processing
¦ 
¦   +-- food_taxonomy.txt                   # Ingredients database
¦   +-- database.pickle                     # Ingredients root noun database
¦   
¦   
+-- recipe1M_1218                           # files created by notebook 3 using data.pickle
¦   +-- chunk.train                         # Essential training data
¦   +-- chunk.val                           # Essential validation data
¦   +-- test/X                              # Inputs in testing set (directory of .txt files)
¦   +-- test/y                              # Human-written outputs in testing set (directory of .txt files)
¦   
¦   
+-- analysis                                
¦   +-- notebook 1-1, 1-2, 2, 3             # Useful for data pre-processing
¦   +-- notebook 4, 5                       # Useful for analyzing the generated texts
¦   +-- notebook 9                          # Compare the generated texts with human-written texts
¦   +-- notebook 10                         # Explain how we convert the users' inputs to model input
¦   +-- notebook 11                         # Explain the 'compare' feature on the website
¦   
¦ 
+-- training                                
¦   +-- gpt-2                               # The source code modified from OpenAI GPT-2
¦       +--src/load_dataset_pad.py              # Padding and fields shuffing
¦       +--src/conditional_gen_web.py           # Input .txt files and receive the output in .txt files
¦       +--train_ppl_pickle.py                  # The main script for fine-tuning with recipe data
¦       +--train_ppl_scratch.py                 # The main script for training from scratch with recipe data
¦       ...
¦       
¦                                           # Details of our experiments as described in the paper
¦   +-- notebook 6                          # Commands of fine-tuning/training the model
¦   +-- notebook 7                          # Ask the model to generate the title/ingredients/instructions 
¦   +-- notebook 8                          # Evaluate the model perplexity
¦ 
+-- common                                  # Import numpy, pickle, ... etc common packages
+-- utils                                   # Some modules related to model evaluation

Download related files

We share the dataset, model, and related files at google-drive

Please download and put them under the correct directory.

RecipeGPT-exp/big_data/
RecipeGPT-exp/training/gpt-2/models/
RecipeGPT-exp/recipe1M_1218/

• We utilize food_taxonomy.txt from Here to create an ingredient database . (Download your own or from our google-drive)
• We utilize vocab.bin, the word embedding trained by Salvador (Already in github)
• We utilize layer1.json , which contains the textual content of Recipe1M(Please click Layers (381 MiB) and download your own)

You may need these files if you want to reproduce notebook 9.

Environment

conda create -n recipegpt python=3.5 anaconda
pip install -r requirements.txt

tensorflow version:

pip install tensorflow-gpu==1.12.0
pip install tensorflow==1.12.0
or use the default version in colab