Git of the project for the Food recipe generation with computational creativity

• Create recipes (list of instructions) based on a list of ingredients, and ingredient pairings (from the KitcheNette dataset)
• Make an extensive analysis of the datasets, and predict the cuisine country origin (French, Japanese, ...) based on the ingredients
• Make a comparison of the different ML models used for the two goals above

Apart from the packages in the requirements.txt file, also need NLTK sudo apt-get install python-nltk on a unix environment

(Using conda env, but installing packages with pip, tested with python 3.6)

http://pic2recipe.csail.mit.edu/ We only need the textual info from the dataset: the det_ingrs.json (Ingredient detections file on the website) and the layer1 json.

{
  id: String,  // unique 10-digit hex string
  title: String,
  instructions: [ { text: String } ],
  ingredients: [ { text: String } ],
  partition: ('train'|'test'|'val'),
  url: String
}

{
  valid: [ bool ] // True if the ingredient in the same order is really an ingredient ? #TODO: check
  id: String,  // unique 10-digit hex string
  ingredients: [ { text: String } ],
}

`` You can use the run_data_processing.bat script that calls data_processing.py to process the Recipe1M dataset. It :

1. Cleans and counts the words, including the ingredients in the dataset
2. Tokenizes words in the recipes The ingredients are those detected in the det_ingrs.json file, crossed with the main layer1.json

Only --recipe1m_path is necessary --threshold_ingrs : minimum ingr count threshold --minnuminstrs : min number of instructions (sentences) --maxnuminstr : max number of instructions (sentences) --maxnumingrs : max number of ingredients --minnumingrs : min number of ingredients --minnumwords : minimum number of words in recipe (sum of the words in all instructions) --forcegen : Force the generation of the vocabulary files, even if they exist in the target folder TODO: --forcegen-all : seems redundant with forcegen

TODO: générer automatiquement la docu

• allingrs_count.pkl: Counter of the ingredients of the preprocessed dataset (filtered out the recipes with too few/many ingredients/instructions)

• allwords_count : Counter of the words in the instructions of the preprocessed dataset (filtered out the recipes with too few/many ingredients/instructions). The instructions usually refer to the ingredients, so the words vocabulary should include the ingredients vocabulary, but that may not always be the case

• recipe1m_vocab_ingrs.pkl : Vocabulary (class defined in '''utils.py''') of ingredients in the dataset, created from the ingredients counter

• recipe1m_vocab_words.pkl : Vocabulary of words in the recipe instructions in the dataset, created from the words counter

• recipe1m_[train/val/test].pkl: Datasets for the food recipe generation models Structure : [ {'instructions': [cleaned instruction (remove instruction number, replace words using the "replace_dict_instrs" dict)], 'tokenized': toks, 'ingredients': ingrs_list, 'title': title} ]

Use the KitcheNette model to predict the ingredient pairing score of all ingredient combinations in the dataset generated at the previous step

The main.py file is the main file for training the models. It can resumes training from a saved model. The different kinds of model are:

• Seq2seq: baseline Seq2Seq with an RNN encoder and decoder. The optimizer is Adam with the Cross entropy loss
• Seq2seqAtt: Seq2Seq with a classic attention mechanism
• Seq2seqIngrAtt: Seq2Seq with the Ingredient attention mechanism
• Seq2seqIngrPairingAtt: Seq2Seq with the Ingredient attention mechanism which includes the score of ingredient pairing
• Seq2seqTitlePairing: Ingredient pairing attention mechanism and includes the title of the recipe as input
• Seq2seqCuisinePairing: Includes the cuisine of the recipe as input
• Seq2seqTrans: Uses a Transformer decoder model
• HierarchicalSeq2seq: Hierarchical Seq2seq baseline
• HierarchicalAtt
• HierarchicalIngrAtt
• HierarchicalIngrPairingAtt
• HierarchicalTitlePairing
• HierarchicalCuisinePairing