shreyrastogi / ratatouille-3.0-novel-generation-tool Goto Github PK

Due to the abundance of cooking recipes available online, there is growing interest in harnessing this information to develop new recipes. In order to solve the "Novel Recipe Generation" problem, our primary goal is to create new, realistic cooking recipes that are so similar to the real ones that even culinary experts can't tell the difference. We fed a vast amount of structured recipe data into a deep learning model like GPT-2 in order to create these innovative recipes.

We used the GPT2 model to generate novel recipes. GPT2 is essentially a machine learning language model that can anticipate the following word from a sentence's context (single token). As part of their investigation, the Open 11 researchers searched the internet to gather a sizable 40GB dataset known as Web Text, which was used to train the GPT-2. The trained GPT-2's smallest variant requires 500 MB of storage to keep all of its parameters. Given that the largest GPT-2 model is 13 times larger, it may use more than 6.5 GB of storage. Transformer decoding blocks are used to construct GPT2. GPT2 outputs one token at a time. GPT2 is auto-regressive in nature; each token in the sentence has the context of the previous words. After each token is produced, that token is added to the sequence of inputs. And that new sequence becomes the input to the model in its next step. This is an idea called "auto-regression". The GPT-2 model can take into account sampling words other than the top word by using a parameter called top-k, which is the case when top-k = 1. While the GPT 3 model is trained on 175 billion parameters, the GPT2 model uses about 1.5 billion parameters.

Overview of feeding in text and generating a single token in GPT2:

1. Tokenization - Take some words, break them up into their common pieces. Take those common pieces and replace them with a number. Tokenization is necessary because computers only work with numbers. This also represents words efficiently.
2. Embedding with time signal - Take that 1 string of numbers and convert each number to a vector. This captures the position of words relative to one another and allows words to take value from other words associated with them, e.g. "The boy ran through the woods, and he surely had not stolen the cherry pie for which they were chasing him." In this sentence "he" should clearly tie a lot of its meaning to "boy".
3. Decoder Block - The pieces are self-attention blocks, feedforward neural nets, and Normalization. Self-attention blocks identify which words to focus on. In the sentence, "Jimmy played with the burning bush, and then went around to the next bush," the words "Jimmy," "played," "burning," and "bush" capture a high proportion of the meaning in that sentence. This idea that certain words and phases capture more meaning and thus should be given more "attention" is the intuition of self-attention blocks.
4. Linear Layer - Prior to the tokenization process, a vocabulary size will be decided upon and a vocabulary will be set up.
5. The vocabulary is just a list of all the possible tokens (numbers) that can be produced and which letter or group of letters the tokens are equal to. The linear layer takes the output of the last decoder block and converts it to a vector whose dimensions are vocabulary size by 1. In short, it takes a lot of inputs and produces a list where each spot represents a token. The higher the number in the spot the better the chance that that token is the best pick.
6. Softmax - Converts the output of the linear layer to a probability distribution. The output of the linear layer tells you information about which tokens are the best picks, but it is hard to use. The values range from very small values(huge negative values) to very large values and their meaning is in relation to all the other values. To make them easier to use apply the Softmax function which converts the vector to a probability distribution. This means each number represents the probability that that token is the correct one.
7. Pick a token - Choose the method to pick the next token from the probability distribution of tokens, and use that method to pick the token. There are various methods to do so including, greedy, temperature sampling, nucleus sampling, and top-k sampling. Convert Token to a word piece using the vocabulary.

We used the RecipeDB_v1 dataset which is an organized collection of recipes, ingredients, and nutrition profiles related to flavor profiles and health associations. The repertoire consists of the laborious fusion of over 1,18,000 recipes from cuisines around the world (six continents, 26 geo-cultural regions, and 74 countries), cooked with over 23,500 ingredients from various categories using 268 different processes (heat, cook, boil, simmer, bake, etc.), and further linked to their flavor molecules (FlavorDB), nutritional profiles (USDA), and empirical records of disease associations obtained from Medline (DietRx). We used the file “Recipe_correct_ndb_updated_v1.csv” in the RecipeDB dataset to go ahead with our problem. The RecipeDB dataset consists of 1154404 records.

Below are the 12 fields in the dataset:

Out of all the fields, recipe_no, ingredient fields are used further to calculate the popularity of the ingredients across the recipes.

Following steps were performed for the data preprocessing:

1. Removal of recipes: Recipes which contain only a single ingredient are removed from the recipe database because they are not of much importance.
2. Merging of useful Information corresponding to recipes: Recipes are mapped with their corresponding instructions, region, continent, sub-region.
3. Lemmatization: Performed lemmatization on the ingredients set of all the recipes to convert all forms of each ingredient to their lemmatized form. Before performing this, all ingredient words are converted to lowercase so that the model should be able to analyze the same ingredients present in multiple recipes because we want that “potato” in one recipe and “Potato” in another recipe should be considered the same by the model.
4. Fixing Punctuations: This operation is also performed on the ingredients set of all recipes.
5. Removing Useless Symbols: Some symbols are present in recipe instructions that are not of any use, therefore we have removed all of these symbols from the instructions of all the recipes.

Along with the default tokens used by the GPT2 model, we have used the following tokens to differentiate different parts of the recipe data.

We have used the 118083 recipes data. Initially, we splitted the data into train and test sets where the train set contains 94.6 % and the test set contains 5.4% of the data. Then we attach the tokens defined above to each of the sets, as a result, train.txt and test.text files are generated. Then the model is trained using the recipes of the train data.

Following values of functional parameters of GPT 2 model are used to train the model:

Model training had taken about 1 to 1.5 hours.

We used the RecipeDB data to train the GPT2 model, which we then used to produce new recipes. The list of ingredients was submitted to the GPT2 model. The method used to create the list of ingredients is listed below.

1. A random number from 2 and 8 was generated, and that many ingredients were chosen at random in proportion to how frequently they appeared across the recipes. The list of those ingredients was fed into the GPT2 model.

Using the trained GPT2 model, we produced 1,20,000 recipes in total. The RecipeDB data was first processed, then special tokens were added, and finally the train set and test set were separated. The GPT2 model, which had been trained on the train set, was fed a list of random ingredients, and it then developed a recipe using those items plus a few extras.

Ratatouille 3.0 implementation is based on popularity constraint. More constraints can be used in the future versions of Ratatouille like: cuisine, dish type, nutrition, cost, ingredient availability, allergy, health condition, dietary preferences.

Ratatouille: A tool for Novel Recipe Generation, IEEE 38th International Conference on Data Engineering Workshops (ICDEW) (2022).