TL;DR

I've trained a character-level LSTM (Long short-term memory) RNN (Recurrent Neural Network) on ~100k recipes dataset using TensorFlow, and it suggested me to cook "Cream Soda with Onions", "Puff Pastry Strawberry Soup", "Zucchini flavor Tea" and "Salmon Mousse of Beef and Stilton Salad with Jalapenos" .

Here you may find more examples of what I ended up with:

• 🎨 Cooking recipes generator demo - to try the model interactively right in your browser.
• 🏋🏻‍ LSTM model training process - to see how the model was trained.
• 🤖 Interactive Machine Learning Experiments repository - to see more experiments with "Objects detection", "Sketch Recognition", "Image Classification" etc.

This article contains details of how the LSTM model was actually trained on Python using TensorFlow 2 with Keras API.

What our model will eventually learn

For a couple of hours of training our character-level RNN model will learn basic concepts of English grammar and punctuation (I wish I could learn English that fast!). It will also learn how to generate different parts of recipes such as 📗 [RECIPE NAME], 🥕 [RECIPE INGREDIENTS] and 📝 [RECIPE INSTRUCTIONS]. Sometimes recipe name, ingredients and instructions will be pretty interesting, sometimes stupid, sometimes fun.

Here are couple of generated recipes examples:

📗 [NAME]

Orange Club Tea Sandwich Cookies

🥕 [INGREDIENTS]

• 1 cup (2 sticks) unsalted butter, softened
• 1 cup confectioners' sugar
• 1/2 cup flaxseed meal
• 1/2 cup shelled pumpkin seeds (pecans, blanched and sliced)
• 2 teaspoons vanilla extract

📝 [INSTRUCTIONS]

▪︎ Preheat oven to 350 degrees F.
▪︎ Combine cake mix, milk, egg and sugar in a large bowl. Stir until combined and smooth but not sticky. Using a spatula, sprinkle the dough biscuits over the bottom of the pan. Sprinkle with sugar, and spread evenly. Bake for 20 minutes. Remove from the oven and cool on a rack. To serve, add the chocolate.

Or another one:

📗 [NAME]

Mushrooms with Lentil Stewed Shallots and Tomatoes

🥕 [INGREDIENTS]

• 1 tablespoon olive oil
• 3 cloves garlic, smashed
• Kosher salt
• 1 1/2 pounds lean ground turkey
• 1 cup coarsely peeled tart apples
• 2 tablespoons chopped garlic
• 1 teaspoon ground cumin
• 1/2 teaspoon cayenne pepper
• 1 teaspoon chopped fresh thyme
• 3/4 cup chopped fresh basil
• 1/2 small carrot, halved lengthwise and cut into 1/2-inch pieces
• 1 roasted red pepper, halved and sliced vertically diced and separated into rough chops
• 3 tablespoons unsalted butter
• 2 cups shredded mozzarella
• 1/4 cup grated parmesan cheese
• 1/4 cup prepared basil pesto

📝 [INSTRUCTIONS]

▪︎ Stir the olive oil, garlic, thyme and 1 teaspoon salt in a saucepan; bring to a simmer over medium heat. Remove from the heat. Add the basil and toast the soup for 2 minutes.
▪︎ Meanwhile, heat 4 to 4 inches vegetable oil in the skillet over medium-high heat. Add the olive oil, garlic, 1/2 teaspoon salt and 1/2 teaspoon pepper and cook, stirring often, until cooked through, a

⚠️ The recipes in this article are generated just for fun and for learning purposes. The recipes are not for actual cooking! If you want some real recipes you may check 🥦 home_full_of_recipes Instagram channel.

Prior knowledge

It is assumed that you're already familiar with concepts of Recurrent Neural Networks (RNNs) and with Long short-term memory (LSTM) architecture in particular.

ℹ️ In case if these concepts are new to you I would highly recommend taking a Deep Learning Specialization on Coursera by Andrew Ng. It also might be beneficial to go through the Unreasonable Effectiveness of Recurrent Neural Networks article by Andrej Karpathy.

On a high level, Recurrent Neural Network (RNN) is a class of deep neural networks, most commonly applied to sequence-based data like speech, voice, text or music. They are used for machine translation, speech recognition, voice synthesis etc. The key feature of RNNs is that they are stateful, and they have an internal memory in which some context for the sequence may be stored. For example if the first word of the sequence was He the RNN might suggest the next word to speaks instead of just speak (to form a He speaks phrase) because the prior knowledge about the first word He is already inside the internal memory.

Image source: Wikipedia

Image source: Towards Data Science

Exciting part is that RNN (and LSTM in particular) could memorize not only word-to-word dependencies but also character-to-character dependencies! It doesn't really matter what sequence consists of: it might be words it might be characters. What is important is that they form a time-distributed sequence. For example, we have a sequence of characters ['H', 'e']. If we ask LSTM what may go next it may suggest a <stop_word> (meaning, that the sequence that forms word He is already complete, and we may stop), or it may also suggest a character l (meaning, that it tries to build a Hello sequence for us). This type of RNNs are called character-level RNNs (as opposed to word-level RNNs).

In this tutorial we will rely on this memorization feature of RNN networks, and we will use a character-level version of LSTM to generate cooking recipes.

Exploring the datasets

Let's go through several available datasets and explore their pros and cons. One of the requirements I want the dataset to meet is that it should have not only a list of ingredients but also a cooking instruction. I also want it to have a measures and quantities for each ingredient.

Here are several cooking recipes datasets I've found:

• 🤷 Recipe Ingredients Dataset (doesn't have ingredients proportions)
• 🤷 Recipe1M+ (a lot of recipes but requires registration to download)
• 🤷 Epicurious - Recipes with Rating and Nutrition (~20k recipes only, it would be nice to find more)
• 👍🏻 Recipe box (~125,000 recipes with ingredients proportions, good)

Let's try to use the "Recipe box" dataset. The number of recipes looks big enough, also it contains both ingredients and cooking instructions. It is interesting to see if RNN will be able to learn a connection between ingredients and instructions.

Setting TensorFlow/Python sandbox for training

There are several options you may follow to experiment with the code in this tutorial:

1. You may experiment by using GoogleColab right in your browser (no local setup is needed).
2. You may experiment by using Jupyter notebook in Binder right in your browser (no local setup is needed).
3. You may setup a Jupyter notebook locally.

I would suggest going with GoogleColab option since it doesn't require any local setup for you (you may experiment right in your browser), and it also provides a powerful GPU support for training that will make the model to train faster. You will be able to experiment with training parameters as well.

Importing dependencies

Let's start with importing some packages that we will use afterwards.

# Packages for training the model and working with the dataset.
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
import json

# Utility/helper packages.
import platform
import time
import pathlib
import os

First, let's make sure our environment is properly set up and that we're using a 2nd version of Tensorflow.

print('Python version:', platform.python_version())
print('Tensorflow version:', tf.__version__)
print('Keras version:', tf.keras.__version__)

➔ output:

Python version: 3.7.6
Tensorflow version: 2.1.0
Keras version: 2.2.4-tf

Loading the dataset

Let's load the dataset using tf.keras.utils.get_file. Using get_file() utility is convenient because it handles caching for you out of the box. It means that you will download the dataset files only once and then even if you launch the same code block in the notebook once again it will use cache, and the code block will be executed faster.

Create cache folder if it not exists:

CACHE_DIR = './tmp'
pathlib.Path(CACHE_DIR).mkdir(exist_ok=True)

Download and unpack the dataset:

dataset_file_name = 'recipes_raw.zip'
dataset_file_origin = 'https://storage.googleapis.com/recipe-box/recipes_raw.zip'

dataset_file_path = tf.keras.utils.get_file(
    fname=dataset_file_name,
    origin=dataset_file_origin,
    cache_dir=CACHE_DIR,
    extract=True,
    archive_format='zip'
)

print(dataset_file_path)

Here is a path to dataset file after it has been downloaded:

➔ output:

./tmp/datasets/recipes_raw.zip

Let's print the cache folder and see what exactly has been downloaded:

!ls -la ./tmp/datasets/

➔ output:

total 521128
drwxr-xr-x  7        224 May 13 18:10 .
drwxr-xr-x  4        128 May 18 18:00 ..
-rw-r--r--  1      20437 May 20 06:46 LICENSE
-rw-r--r--  1   53355492 May 13 18:10 recipes_raw.zip
-rw-r--r--  1   49784325 May 20 06:46 recipes_raw_nosource_ar.json
-rw-r--r--  1   61133971 May 20 06:46 recipes_raw_nosource_epi.json
-rw-r--r--  1   93702755 May 20 06:46 recipes_raw_nosource_fn.json

As you may see, the dataset consists of 3 files. We need to merge information from those 3 files into one dataset later.

Let's load datasets data from json files and preview examples from them.

def load_dataset(silent=False):
    # List of dataset files we want to merge.
    dataset_file_names = [
        'recipes_raw_nosource_ar.json',
        'recipes_raw_nosource_epi.json',
        'recipes_raw_nosource_fn.json',
    ]

    dataset = []

    for dataset_file_name in dataset_file_names:
        dataset_file_path = f'{CACHE_DIR}/datasets/{dataset_file_name}'

        with open(dataset_file_path) as dataset_file:
            json_data_dict = json.load(dataset_file)
            json_data_list = list(json_data_dict.values())
            dict_keys = [key for key in json_data_list[0]]
            dict_keys.sort()
            dataset += json_data_list

            # This code block outputs the summary for each dataset.
            if silent == False:
                print(dataset_file_path)
                print('===========================================')
                print('Number of examples: ', len(json_data_list), '\n')
                print('Example object keys:\n', dict_keys, '\n')
                print('Example object:\n', json_data_list[0], '\n')
                print('Required keys:\n')
                print('  title: ', json_data_list[0]['title'], '\n')
                print('  ingredients: ', json_data_list[0]['ingredients'], '\n')
                print('  instructions: ', json_data_list[0]['instructions'])
                print('\n\n')

    return dataset  

dataset_raw = load_dataset() 

➔ output:

./tmp/datasets/recipes_raw_nosource_ar.json
===========================================
Number of examples:  39802 

Example object keys:
 ['ingredients', 'instructions', 'picture_link', 'title'] 

Example object:
 {'title': 'Slow Cooker Chicken and Dumplings', 'ingredients': ['4 skinless, boneless chicken breast halves ADVERTISEMENT', '2 tablespoons butter ADVERTISEMENT', '2 (10.75 ounce) cans condensed cream of chicken soup ADVERTISEMENT', '1 onion, finely diced ADVERTISEMENT', '2 (10 ounce) packages refrigerated biscuit dough, torn into pieces ADVERTISEMENT', 'ADVERTISEMENT'], 'instructions': 'Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.\nCover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.\n', 'picture_link': '55lznCYBbs2mT8BTx6BTkLhynGHzM.S'} 

Required keys:

  title:  Slow Cooker Chicken and Dumplings 

  ingredients:  ['4 skinless, boneless chicken breast halves ADVERTISEMENT', '2 tablespoons butter ADVERTISEMENT', '2 (10.75 ounce) cans condensed cream of chicken soup ADVERTISEMENT', '1 onion, finely diced ADVERTISEMENT', '2 (10 ounce) packages refrigerated biscuit dough, torn into pieces ADVERTISEMENT', 'ADVERTISEMENT'] 

  instructions:  Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.
Cover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.




./tmp/datasets/recipes_raw_nosource_epi.json
===========================================
Number of examples:  25323 

Example object keys:
 ['ingredients', 'instructions', 'picture_link', 'title'] 

Example object:
 {'ingredients': ['12 egg whites', '12 egg yolks', '1 1/2 cups sugar', '3/4 cup rye whiskey', '12 egg whites', '3/4 cup brandy', '1/2 cup rum', '1 to 2 cups heavy cream, lightly whipped', 'Garnish: ground nutmeg'], 'picture_link': None, 'instructions': 'Beat the egg whites until stiff, gradually adding in 3/4 cup sugar. Set aside. Beat the egg yolks until they are thick and pale and add the other 3/4 cup sugar and stir in rye whiskey. Blend well. Fold the egg white mixture into the yolk mixture and add the brandy and the rum. Beat the mixture well. To serve, fold the lightly whipped heavy cream into the eggnog. (If a thinner mixture is desired, add the heavy cream unwhipped.) Sprinkle the top of the eggnog with the nutmeg to taste.\nBeat the egg whites until stiff, gradually adding in 3/4 cup sugar. Set aside. Beat the egg yolks until they are thick and pale and add the other 3/4 cup sugar and stir in rye whiskey. Blend well. Fold the egg white mixture into the yolk mixture and add the brandy and the rum. Beat the mixture well. To serve, fold the lightly whipped heavy cream into the eggnog. (If a thinner mixture is desired, add the heavy cream unwhipped.) Sprinkle the top of the eggnog with the nutmeg to taste.', 'title': 'Christmas Eggnog '} 

Required keys:

  title:  Christmas Eggnog  

  ingredients:  ['12 egg whites', '12 egg yolks', '1 1/2 cups sugar', '3/4 cup rye whiskey', '12 egg whites', '3/4 cup brandy', '1/2 cup rum', '1 to 2 cups heavy cream, lightly whipped', 'Garnish: ground nutmeg'] 

  instructions:  Beat the egg whites until stiff, gradually adding in 3/4 cup sugar. Set aside. Beat the egg yolks until they are thick and pale and add the other 3/4 cup sugar and stir in rye whiskey. Blend well. Fold the egg white mixture into the yolk mixture and add the brandy and the rum. Beat the mixture well. To serve, fold the lightly whipped heavy cream into the eggnog. (If a thinner mixture is desired, add the heavy cream unwhipped.) Sprinkle the top of the eggnog with the nutmeg to taste.
Beat the egg whites until stiff, gradually adding in 3/4 cup sugar. Set aside. Beat the egg yolks until they are thick and pale and add the other 3/4 cup sugar and stir in rye whiskey. Blend well. Fold the egg white mixture into the yolk mixture and add the brandy and the rum. Beat the mixture well. To serve, fold the lightly whipped heavy cream into the eggnog. (If a thinner mixture is desired, add the heavy cream unwhipped.) Sprinkle the top of the eggnog with the nutmeg to taste.



./tmp/datasets/recipes_raw_nosource_fn.json
===========================================
Number of examples:  60039 

Example object keys:
 ['ingredients', 'instructions', 'picture_link', 'title'] 

Example object:
 {'instructions': 'Toss ingredients lightly and spoon into a buttered baking dish. Top with additional crushed cracker crumbs, and brush with melted butter. Bake in a preheated at 350 degrees oven for 25 to 30 minutes or until delicately browned.', 'ingredients': ['1/2 cup celery, finely chopped', '1 small green pepper finely chopped', '1/2 cup finely sliced green onions', '1/4 cup chopped parsley', '1 pound crabmeat', '1 1/4 cups coarsely crushed cracker crumbs', '1/2 teaspoon salt', '3/4 teaspoons dry mustard', 'Dash hot sauce', '1/4 cup heavy cream', '1/2 cup melted butter'], 'title': "Grammie Hamblet's Deviled Crab", 'picture_link': None} 

Required keys:

  title:  Grammie Hamblet's Deviled Crab 

  ingredients:  ['1/2 cup celery, finely chopped', '1 small green pepper finely chopped', '1/2 cup finely sliced green onions', '1/4 cup chopped parsley', '1 pound crabmeat', '1 1/4 cups coarsely crushed cracker crumbs', '1/2 teaspoon salt', '3/4 teaspoons dry mustard', 'Dash hot sauce', '1/4 cup heavy cream', '1/2 cup melted butter'] 

  instructions:  Toss ingredients lightly and spoon into a buttered baking dish. Top with additional crushed cracker crumbs, and brush with melted butter. Bake in a preheated at 350 degrees oven for 25 to 30 minutes or until delicately browned.

Let's count the total number of examples after we merged the files:

print('Total number of raw examples: ', len(dataset_raw))

➔ output:

Total number of raw examples:  125164

Preprocessing the dataset

Filtering out incomplete examples

It is possible that some recipes don't have some required fields (name, ingredients or instructions). We need to clean our dataset from those incomplete examples.

The following function will help us filter out recipes which don't have either title or ingredients or instructions:

def recipe_validate_required_fields(recipe):
    required_keys = ['title', 'ingredients', 'instructions']

    if not recipe:
        return False

    for required_key in required_keys:
        if not recipe[required_key]:
            return False

        if type(recipe[required_key]) == list and len(recipe[required_key]) == 0:
            return False

    return True

Let's do the filtering now using recipe_validate_required_fields() function:

dataset_validated = [recipe for recipe in dataset_raw if recipe_validate_required_fields(recipe)]

print('Dataset size BEFORE validation', len(dataset_raw))
print('Dataset size AFTER validation', len(dataset_validated))
print('Number of incomplete recipes', len(dataset_raw) - len(dataset_validated))

➔ output:

Dataset size BEFORE validation 125164
Dataset size AFTER validation 122938
Number of incomplete recipes 2226

As you may see among 125164 recipes we had 2226 somehow incomplete.

Converting recipes objects into strings

RNN doesn't understand objects. Therefore, we need to convert recipes objects to string and then to numbers (indices). Let's start with converting recipes objects to strings.

To help our RNN learn the structure of the text faster let's add 3 "landmarks" to it. We will use these unique "title", "ingredients" and "instruction" landmarks to separate the logic sections of each recipe.

STOP_WORD_TITLE = '📗 '
STOP_WORD_INGREDIENTS = '\n🥕\n\n'
STOP_WORD_INSTRUCTIONS = '\n📝\n\n'

The following function converts the recipe object to a string (sequence of characters) for later usage in RNN input.

def recipe_to_string(recipe):
    # This string is presented as a part of recipes so we need to clean it up.
    noize_string = 'ADVERTISEMENT'

    title = recipe['title']
    ingredients = recipe['ingredients']
    instructions = recipe['instructions'].split('\n')

    ingredients_string = ''
    for ingredient in ingredients:
        ingredient = ingredient.replace(noize_string, '')
        if ingredient:
            ingredients_string += f'• {ingredient}\n'

    instructions_string = ''
    for instruction in instructions:
        instruction = instruction.replace(noize_string, '')
        if instruction:
            instructions_string += f'▪︎ {instruction}\n'

    return f'{STOP_WORD_TITLE}{title}\n{STOP_WORD_INGREDIENTS}{ingredients_string}{STOP_WORD_INSTRUCTIONS}{instructions_string}'

Let's apply recipe_to_string() function to dataset_validated:

dataset_stringified = [recipe_to_string(recipe) for recipe in dataset_validated]

print('Stringified dataset size: ', len(dataset_stringified))

➔ output:

Stringified dataset size:  122938

Let's preview first several recipes:

for recipe_index, recipe_string in enumerate(dataset_stringified[:3]):
    print('Recipe #{}\n---------'.format(recipe_index + 1))
    print(recipe_string)
    print('\n')

➔ output:

Recipe #1
---------
📗 Slow Cooker Chicken and Dumplings

🥕

• 4 skinless, boneless chicken breast halves 
• 2 tablespoons butter 
• 2 (10.75 ounce) cans condensed cream of chicken soup 
• 1 onion, finely diced 
• 2 (10 ounce) packages refrigerated biscuit dough, torn into pieces 

📝

▪︎ Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.
▪︎ Cover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.



Recipe #2
---------
📗 Awesome Slow Cooker Pot Roast

🥕

• 2 (10.75 ounce) cans condensed cream of mushroom soup 
• 1 (1 ounce) package dry onion soup mix 
• 1 1/4 cups water 
• 5 1/2 pounds pot roast 

📝

▪︎ In a slow cooker, mix cream of mushroom soup, dry onion soup mix and water. Place pot roast in slow cooker and coat with soup mixture.
▪︎ Cook on High setting for 3 to 4 hours, or on Low setting for 8 to 9 hours.



Recipe #3
---------
📗 Brown Sugar Meatloaf

🥕

• 1/2 cup packed brown sugar 
• 1/2 cup ketchup 
• 1 1/2 pounds lean ground beef 
• 3/4 cup milk 
• 2 eggs 
• 1 1/2 teaspoons salt 
• 1/4 teaspoon ground black pepper 
• 1 small onion, chopped 
• 1/4 teaspoon ground ginger 
• 3/4 cup finely crushed saltine cracker crumbs 

📝

▪︎ Preheat oven to 350 degrees F (175 degrees C). Lightly grease a 5x9 inch loaf pan.
▪︎ Press the brown sugar in the bottom of the prepared loaf pan and spread the ketchup over the sugar.
▪︎ In a mixing bowl, mix thoroughly all remaining ingredients and shape into a loaf. Place on top of the ketchup.
▪︎ Bake in preheated oven for 1 hour or until juices are clear.

Just out of curiosity let's preview the recipe somewhere from the middle of the dataset to see that it has expected data structure:

print(dataset_stringified[50000])

➔ output:

📗 Herbed Bean Ragoût 

🥕

• 6 ounces haricots verts (French thin green beans), trimmed and halved crosswise
• 1 (1-pound) bag frozen edamame (soybeans in the pod) or 1 1/4 cups frozen shelled edamame, not thawed
• 2/3 cup finely chopped onion
• 2 garlic cloves, minced
• 1 Turkish bay leaf or 1/2 California bay leaf
• 2 (3-inch) fresh rosemary sprigs
• 1/2 teaspoon salt
• 1/4 teaspoon black pepper
• 1 tablespoon olive oil
• 1 medium carrot, cut into 1/8-inch dice
• 1 medium celery rib, cut into 1/8-inch dice
• 1 (15- to 16-ounces) can small white beans, rinsed and drained
• 1 1/2 cups chicken stock or low-sodium broth
• 2 tablespoons unsalted butter
• 2 tablespoons finely chopped fresh flat-leaf parsley
• 1 tablespoon finely chopped fresh chervil (optional)
• Garnish: fresh chervil sprigs

📝

▪︎ Cook haricots verts in a large pot of boiling salted water until just tender, 3 to 4 minutes. Transfer with a slotted spoon to a bowl of ice and cold water, then drain. Add edamame to boiling water and cook 4 minutes. Drain in a colander, then rinse under cold water. If using edamame in pods, shell them and discard pods. Cook onion, garlic, bay leaf, rosemary, salt, and pepper in oil in a 2- to 4-quart heavy saucepan over moderately low heat, stirring, until softened, about 3 minutes. Add carrot and celery and cook, stirring, until softened, about 3 minutes. Add white beans and stock and simmer, covered, stirring occasionally, 10 minutes. Add haricots verts and edamame and simmer, uncovered, until heated through, 2 to 3 minutes. Add butter, parsley, and chervil (if using) and stir gently until butter is melted. Discard bay leaf and rosemary sprigs.
▪︎ Cook haricots verts in a large pot of boiling salted water until just tender, 3 to 4 minutes. Transfer with a slotted spoon to a bowl of ice and cold water, then drain.
▪︎ Add edamame to boiling water and cook 4 minutes. Drain in a colander, then rinse under cold water. If using edamame in pods, shell them and discard pods.
▪︎ Cook onion, garlic, bay leaf, rosemary, salt, and pepper in oil in a 2- to 4-quart heavy saucepan over moderately low heat, stirring, until softened, about 3 minutes. Add carrot and celery and cook, stirring, until softened, about 3 minutes.
▪︎ Add white beans and stock and simmer, covered, stirring occasionally, 10 minutes. Add haricots verts and edamame and simmer, uncovered, until heated through, 2 to 3 minutes. Add butter, parsley, and chervil (if using) and stir gently until butter is melted. Discard bay leaf and rosemary sprigs.

Filtering out large recipes

Recipes have different lengths. We need to have one hard-coded sequence length limit before feeding recipe sequences to RNN. We need to find out what recipe length will cover most of the recipe use-cases and at the same time we want to keep it as small as possible to speed up the training process.

recipes_lengths = []
for recipe_text in dataset_stringified:
    recipes_lengths.append(len(recipe_text))

plt.hist(recipes_lengths, bins=50)
plt.show()

➔ output:

Most of the recipes have length less than 5000 characters. Let's zoom in to see more detailed picture:

plt.hist(recipes_lengths, range=(0, 8000), bins=50)
plt.show()

➔ output:

Looks like a limit of 2000 characters for the recipes will cover most of the cases. We may try to train RNN with this maximum recipe length limit.

MAX_RECIPE_LENGTH = 2000

Therefore, let's filter out all the recipes that are longer than MAX_RECIPE_LENGTH:

def filter_recipes_by_length(recipe_test):
    return len(recipe_test) <= MAX_RECIPE_LENGTH 

dataset_filtered = [recipe_text for recipe_text in dataset_stringified if filter_recipes_by_length(recipe_text)]

print('Dataset size BEFORE filtering: ', len(dataset_stringified))
print('Dataset size AFTER filtering: ', len(dataset_filtered))
print('Number of eliminated recipes: ', len(dataset_stringified) - len(dataset_filtered))

➔ output:

Dataset size BEFORE filtering:  122938
Dataset size AFTER filtering:  100212
Number of eliminated recipes:  22726

We lost 22726 recipes during this filtering but now recipes' data is more dense.

Summarizing dataset parameters

TOTAL_RECIPES_NUM = len(dataset_filtered)

print('MAX_RECIPE_LENGTH: ', MAX_RECIPE_LENGTH)
print('TOTAL_RECIPES_NUM: ', TOTAL_RECIPES_NUM)

➔ output:

MAX_RECIPE_LENGTH:  2000
TOTAL_RECIPES_NUM:  100212

Finally, we ended up with ~100k recipes. Each recipe has 2000 characters length.

Creating vocabulary

Recurrent neural network doesn't understand characters or words. It understands numbers instead. Therefore, we need to convert recipe texts to numbers.

In this experiment we're going to use a character-level language model based on multi-layer LSTM (Long Short-Term Memory) network (as opposed to word-level language model). It means that instead of creating unique indices for words we will create unique indices for characters. By doing that we let the network predict the next character instead of the next word in a sequence.

ℹ️ You may find more details about character-level RNNs explanation in the Unreasonable Effectiveness of Recurrent Neural Networks article by Andrej Karpathy:

To create a vocabulary out of recipes texts we will use tf.keras.preprocessing.text.Tokenizer.

We also need to come with some unique character that will be treated as a stop-character and will indicate the end of a recipe. We need it for recipe generation afterwards since without this stop-character we won't know where the end of a recipe that we're generating is.

STOP_SIGN = '␣'

tokenizer = tf.keras.preprocessing.text.Tokenizer(
    char_level=True,
    filters='',
    lower=False,
    split=''
)

# Stop word is not a part of recipes, but tokenizer must know about it as well.
tokenizer.fit_on_texts([STOP_SIGN])

tokenizer.fit_on_texts(dataset_filtered)

tokenizer.get_config()

➔ output:

{'num_words': None,
 'filters': '',
 'lower': False,
 'split': '',
 'char_level': True,
 'oov_token': None,
 'document_count': 100213,

 'word_counts': '{"\\u2423": 1, "\\ud83d\\udcd7": 100212, " ": 17527888, "S": 270259, "l": 3815150, "o": 5987496, "w": 964459, "C": 222831, "k": 890982, "e": 9296022, "r": 4760887, "h": 2922100, "i": 4911812, "c": 2883507, "n": 5304396, "a": 6067157, "d": 3099679, "D": 63999, "u": 2717050, "m": 1794411, "p": 2679164, "g": 1698670, "s": 4704222, "\\n": 1955281, "\\ud83e\\udd55": 100212, "\\u2022": 922813, "4": 232607, ",": 1130487, "b": 1394803, "t": 5997722, "v": 746785, "2": 493933, "(": 144985, "1": 853931, "0": 145119, ".": 1052548, "7": 31098, "5": 154071, ")": 144977, "f": 1042981, "y": 666553, "\\ud83d\\udcdd": 100212, "\\u25aa": 331058, "\\ufe0e": 331058, "P": 200597, "6": 51398, "H": 43936, "A": 134274, "3": 213519, "R": 101253, "x": 201286, "/": 345257, "I": 81591, "L": 46138, "8": 55352, "9": 17697, "B": 123813, "M": 78684, "F": 104359, "j": 110008, "-": 219160, "W": 61616, "\\u00ae": 10159, "N": 12808, "q": 69654, "T": 101371, ";": 72045, "\'": 26831, "Z": 2428, "z": 115883, "G": 52043, ":": 31318, "E": 18582, "K": 18421, "X": 385, "\\"": 6445, "O": 28971, "Y": 6064, "\\u2122": 538, "Q": 3904, "J": 10269, "!": 3014, "U": 14132, "V": 12172, "&": 1039, "+": 87, "=": 113, "%": 993, "*": 3243, "\\u00a9": 99, "[": 30, "]": 31, "\\u00e9": 6727, "<": 76, ">": 86, "\\u00bd": 166, "#": 168, "\\u00f1": 891, "?": 327, "\\u2019": 111, "\\u00b0": 6808, "\\u201d": 6, "$": 84, "@": 5, "{": 8, "}": 9, "\\u2013": 1228, "\\u0096": 7, "\\u00e0": 26, "\\u00e2": 106, "\\u00e8": 846, "\\u00e1": 74, "\\u2014": 215, "\\u2044": 16, "\\u00ee": 415, "\\u00e7": 171, "_": 26, "\\u00fa": 48, "\\u00ef": 43, "\\u201a": 20, "\\u00fb": 36, "\\u00f3": 74, "\\u00ed": 130, "\\u25ca": 4, "\\u00f9": 12, "\\u00d7": 6, "\\u00ec": 8, "\\u00fc": 29, "\\u2031": 4, "\\u00ba": 19, "\\u201c": 4, "\\u00ad": 25, "\\u00ea": 27, "\\u00f6": 9, "\\u0301": 11, "\\u00f4": 8, "\\u00c1": 2, "\\u00be": 23, "\\u00bc": 95, "\\u00eb": 2, "\\u0097": 2, "\\u215b": 3, "\\u2027": 4, "\\u00e4": 15, "\\u001a": 2, "\\u00f8": 2, "\\ufffd": 20, "\\u02da": 6, "\\u00bf": 264, "\\u2153": 2, "|": 2, "\\u00e5": 3, "\\u00a4": 1, "\\u201f": 1, "\\u00a7": 5, "\\ufb02": 3, "\\u00a0": 1, "\\u01b0": 2, "\\u01a1": 1, "\\u0103": 1, "\\u0300": 1, "\\u00bb": 6, "`": 3, "\\u0092": 2, "\\u215e": 1, "\\u202d": 4, "\\u00b4": 2, "\\u2012": 2, "\\u00c9": 40, "\\u00da": 14, "\\u20ac": 1, "\\\\": 5, "~": 1, "\\u0095": 1, "\\u00c2": 2}',

 'word_docs': '{"\\u2423": 1, "k": 97316, "0": 61954, "o": 100205, "r": 100207, "d": 100194, "u": 100161, "S": 89250, "\\u25aa": 100212, "D": 40870, "1": 99320, "g": 99975, "n": 100198, "b": 99702, "t": 100202, ".": 100163, " ": 100212, "7": 24377, "3": 79135, "\\ud83d\\udcd7": 100212, "i": 100207, "5": 65486, "f": 98331, "c": 100190, "4": 82453, "a": 100205, "2": 96743, "v": 97848, "C": 83328, "s": 100204, "\\n": 100212, "6": 35206, "\\ud83d\\udcdd": 100212, ",": 98524, "\\ufe0e": 100212, "l": 100206, "e": 100212, "y": 96387, ")": 67614, "p": 100046, "H": 31908, "\\ud83e\\udd55": 100212, "m": 99988, "w": 99227, "(": 67627, "A": 60900, "h": 100161, "\\u2022": 100212, "P": 79364, "R": 54040, "9": 14114, "8": 37000, "L": 32101, "x": 72133, "I": 46675, "/": 89051, "j": 47438, "F": 57940, "B": 64278, "M": 48332, "-": 74711, "T": 53758, "\\u00ae": 5819, "N": 9981, "W": 38981, "q": 36538, ";": 33863, "G": 35355, "\'": 18120, "z": 42430, "Z": 2184, ":": 18214, "E": 12161, "K": 14834, "X": 321, "\\"": 2617, "O": 20103, "Y": 5148, "\\u2122": 448, "Q": 3142, "J": 8225, "!": 2428, "U": 10621, "V": 9710, "&": 749, "+": 32, "=": 48, "%": 717, "*": 1780, "\\u00a9": 91, "]": 26, "[": 25, "\\u00e9": 2462, ">": 33, "<": 27, "\\u00bd": 81, "#": 139, "\\u00f1": 423, "?": 207, "\\u2019": 64, "\\u00b0": 3062, "\\u201d": 3, "@": 4, "$": 49, "{": 7, "}": 8, "\\u2013": 491, "\\u0096": 7, "\\u00e0": 22, "\\u00e2": 45, "\\u00e8": 335, "\\u00e1": 38, "\\u2014": 95, "\\u2044": 9, "\\u00ee": 122, "\\u00e7": 120, "_": 8, "\\u00fa": 25, "\\u00ef": 24, "\\u201a": 10, "\\u00fb": 29, "\\u00f3": 40, "\\u00ed": 52, "\\u25ca": 2, "\\u00f9": 6, "\\u00d7": 4, "\\u00ec": 4, "\\u00fc": 19, "\\u2031": 2, "\\u00ba": 9, "\\u201c": 2, "\\u00ad": 11, "\\u00ea": 4, "\\u00f6": 4, "\\u0301": 6, "\\u00f4": 5, "\\u00c1": 2, "\\u00be": 18, "\\u00bc": 55, "\\u00eb": 2, "\\u0097": 1, "\\u215b": 2, "\\u2027": 3, "\\u00e4": 8, "\\u001a": 1, "\\u00f8": 1, "\\ufffd": 4, "\\u02da": 3, "\\u00bf": 191, "\\u2153": 1, "|": 2, "\\u00e5": 1, "\\u00a4": 1, "\\u201f": 1, "\\u00a7": 3, "\\ufb02": 1, "\\u0300": 1, "\\u01a1": 1, "\\u00a0": 1, "\\u01b0": 1, "\\u0103": 1, "\\u00bb": 2, "`": 3, "\\u0092": 2, "\\u215e": 1, "\\u202d": 1, "\\u00b4": 1, "\\u2012": 1, "\\u00c9": 15, "\\u00da": 5, "\\u20ac": 1, "\\\\": 5, "~": 1, "\\u0095": 1, "\\u00c2": 1}',

 'index_docs': '{"1": 100212, "165": 1, "25": 97316, "41": 61954, "5": 100205, "8": 100207, "11": 100194, "14": 100161, "33": 89250, "31": 100212, "58": 40870, "26": 99320, "18": 99975, "6": 100198, "19": 99702, "4": 100202, "21": 100163, "66": 24377, "37": 79135, "51": 100212, "7": 100207, "40": 65486, "22": 98331, "13": 100190, "34": 82453, "3": 100205, "29": 96743, "27": 97848, "35": 83328, "9": 100204, "16": 100212, "62": 35206, "53": 100212, "20": 98524, "32": 100212, "10": 100206, "2": 100212, "28": 96387, "43": 67614, "15": 100046, "64": 31908, "52": 100212, "17": 99988, "23": 99227, "42": 67627, "44": 60900, "12": 100161, "24": 100212, "39": 79364, "50": 54040, "71": 14114, "60": 37000, "63": 32101, "38": 72133, "54": 46675, "30": 89051, "47": 47438, "48": 57940, "45": 64278, "55": 48332, "36": 74711, "49": 53758, "76": 5819, "73": 9981, "59": 38981, "57": 36538, "56": 33863, "61": 35355, "68": 18120, "46": 42430, "84": 2184, "65": 18214, "69": 12161, "70": 14834, "92": 321, "79": 2617, "67": 20103, "80": 5148, "90": 448, "81": 3142, "75": 8225, "83": 2428, "72": 10621, "74": 9710, "86": 749, "105": 32, "100": 48, "87": 717, "82": 1780, "103": 91, "115": 26, "116": 25, "78": 2462, "106": 33, "108": 27, "98": 81, "97": 139, "88": 423, "93": 207, "101": 64, "77": 3062, "137": 3, "141": 4, "107": 49, "133": 7, "131": 8, "85": 491, "136": 7, "119": 22, "102": 45, "89": 335, "109": 38, "95": 95, "126": 9, "91": 122, "96": 120, "120": 8, "111": 25, "112": 24, "123": 10, "114": 29, "110": 40, "99": 52, "144": 2, "129": 6, "138": 4, "134": 4, "117": 19, "145": 2, "125": 9, "146": 2, "121": 11, "118": 4, "132": 4, "130": 6, "135": 5, "153": 2, "122": 18, "104": 55, "154": 2, "155": 1, "149": 2, "147": 3, "127": 8, "156": 1, "157": 1, "124": 4, "139": 3, "94": 191, "158": 1, "159": 2, "150": 1, "166": 1, "167": 1, "142": 3, "151": 1, "171": 1, "169": 1, "168": 1, "160": 1, "170": 1, "140": 2, "152": 3, "161": 2, "172": 1, "148": 1, "162": 1, "163": 1, "113": 15, "128": 5, "173": 1, "143": 5, "174": 1, "175": 1, "164": 1}',

 'index_word': '{"1": " ", "2": "e", "3": "a", "4": "t", "5": "o", "6": "n", "7": "i", "8": "r", "9": "s", "10": "l", "11": "d", "12": "h", "13": "c", "14": "u", "15": "p", "16": "\\n", "17": "m", "18": "g", "19": "b", "20": ",", "21": ".", "22": "f", "23": "w", "24": "\\u2022", "25": "k", "26": "1", "27": "v", "28": "y", "29": "2", "30": "/", "31": "\\u25aa", "32": "\\ufe0e", "33": "S", "34": "4", "35": "C", "36": "-", "37": "3", "38": "x", "39": "P", "40": "5", "41": "0", "42": "(", "43": ")", "44": "A", "45": "B", "46": "z", "47": "j", "48": "F", "49": "T", "50": "R", "51": "\\ud83d\\udcd7", "52": "\\ud83e\\udd55", "53": "\\ud83d\\udcdd", "54": "I", "55": "M", "56": ";", "57": "q", "58": "D", "59": "W", "60": "8", "61": "G", "62": "6", "63": "L", "64": "H", "65": ":", "66": "7", "67": "O", "68": "\'", "69": "E", "70": "K", "71": "9", "72": "U", "73": "N", "74": "V", "75": "J", "76": "\\u00ae", "77": "\\u00b0", "78": "\\u00e9", "79": "\\"", "80": "Y", "81": "Q", "82": "*", "83": "!", "84": "Z", "85": "\\u2013", "86": "&", "87": "%", "88": "\\u00f1", "89": "\\u00e8", "90": "\\u2122", "91": "\\u00ee", "92": "X", "93": "?", "94": "\\u00bf", "95": "\\u2014", "96": "\\u00e7", "97": "#", "98": "\\u00bd", "99": "\\u00ed", "100": "=", "101": "\\u2019", "102": "\\u00e2", "103": "\\u00a9", "104": "\\u00bc", "105": "+", "106": ">", "107": "$", "108": "<", "109": "\\u00e1", "110": "\\u00f3", "111": "\\u00fa", "112": "\\u00ef", "113": "\\u00c9", "114": "\\u00fb", "115": "]", "116": "[", "117": "\\u00fc", "118": "\\u00ea", "119": "\\u00e0", "120": "_", "121": "\\u00ad", "122": "\\u00be", "123": "\\u201a", "124": "\\ufffd", "125": "\\u00ba", "126": "\\u2044", "127": "\\u00e4", "128": "\\u00da", "129": "\\u00f9", "130": "\\u0301", "131": "}", "132": "\\u00f6", "133": "{", "134": "\\u00ec", "135": "\\u00f4", "136": "\\u0096", "137": "\\u201d", "138": "\\u00d7", "139": "\\u02da", "140": "\\u00bb", "141": "@", "142": "\\u00a7", "143": "\\\\", "144": "\\u25ca", "145": "\\u2031", "146": "\\u201c", "147": "\\u2027", "148": "\\u202d", "149": "\\u215b", "150": "\\u00e5", "151": "\\ufb02", "152": "`", "153": "\\u00c1", "154": "\\u00eb", "155": "\\u0097", "156": "\\u001a", "157": "\\u00f8", "158": "\\u2153", "159": "|", "160": "\\u01b0", "161": "\\u0092", "162": "\\u00b4", "163": "\\u2012", "164": "\\u00c2", "165": "\\u2423", "166": "\\u00a4", "167": "\\u201f", "168": "\\u00a0", "169": "\\u01a1", "170": "\\u0103", "171": "\\u0300", "172": "\\u215e", "173": "\\u20ac", "174": "~", "175": "\\u0095"}',

 'word_index': '{" ": 1, "e": 2, "a": 3, "t": 4, "o": 5, "n": 6, "i": 7, "r": 8, "s": 9, "l": 10, "d": 11, "h": 12, "c": 13, "u": 14, "p": 15, "\\n": 16, "m": 17, "g": 18, "b": 19, ",": 20, ".": 21, "f": 22, "w": 23, "\\u2022": 24, "k": 25, "1": 26, "v": 27, "y": 28, "2": 29, "/": 30, "\\u25aa": 31, "\\ufe0e": 32, "S": 33, "4": 34, "C": 35, "-": 36, "3": 37, "x": 38, "P": 39, "5": 40, "0": 41, "(": 42, ")": 43, "A": 44, "B": 45, "z": 46, "j": 47, "F": 48, "T": 49, "R": 50, "\\ud83d\\udcd7": 51, "\\ud83e\\udd55": 52, "\\ud83d\\udcdd": 53, "I": 54, "M": 55, ";": 56, "q": 57, "D": 58, "W": 59, "8": 60, "G": 61, "6": 62, "L": 63, "H": 64, ":": 65, "7": 66, "O": 67, "\'": 68, "E": 69, "K": 70, "9": 71, "U": 72, "N": 73, "V": 74, "J": 75, "\\u00ae": 76, "\\u00b0": 77, "\\u00e9": 78, "\\"": 79, "Y": 80, "Q": 81, "*": 82, "!": 83, "Z": 84, "\\u2013": 85, "&": 86, "%": 87, "\\u00f1": 88, "\\u00e8": 89, "\\u2122": 90, "\\u00ee": 91, "X": 92, "?": 93, "\\u00bf": 94, "\\u2014": 95, "\\u00e7": 96, "#": 97, "\\u00bd": 98, "\\u00ed": 99, "=": 100, "\\u2019": 101, "\\u00e2": 102, "\\u00a9": 103, "\\u00bc": 104, "+": 105, ">": 106, "$": 107, "<": 108, "\\u00e1": 109, "\\u00f3": 110, "\\u00fa": 111, "\\u00ef": 112, "\\u00c9": 113, "\\u00fb": 114, "]": 115, "[": 116, "\\u00fc": 117, "\\u00ea": 118, "\\u00e0": 119, "_": 120, "\\u00ad": 121, "\\u00be": 122, "\\u201a": 123, "\\ufffd": 124, "\\u00ba": 125, "\\u2044": 126, "\\u00e4": 127, "\\u00da": 128, "\\u00f9": 129, "\\u0301": 130, "}": 131, "\\u00f6": 132, "{": 133, "\\u00ec": 134, "\\u00f4": 135, "\\u0096": 136, "\\u201d": 137, "\\u00d7": 138, "\\u02da": 139, "\\u00bb": 140, "@": 141, "\\u00a7": 142, "\\\\": 143, "\\u25ca": 144, "\\u2031": 145, "\\u201c": 146, "\\u2027": 147, "\\u202d": 148, "\\u215b": 149, "\\u00e5": 150, "\\ufb02": 151, "`": 152, "\\u00c1": 153, "\\u00eb": 154, "\\u0097": 155, "\\u001a": 156, "\\u00f8": 157, "\\u2153": 158, "|": 159, "\\u01b0": 160, "\\u0092": 161, "\\u00b4": 162, "\\u2012": 163, "\\u00c2": 164, "\\u2423": 165, "\\u00a4": 166, "\\u201f": 167, "\\u00a0": 168, "\\u01a1": 169, "\\u0103": 170, "\\u0300": 171, "\\u215e": 172, "\\u20ac": 173, "~": 174, "\\u0095": 175}'}

To get a full size of a vocabulary we need to add +1 to the number of already registered characters because index 0 is a reserved index that won't be assigned to any word.

VOCABULARY_SIZE = len(tokenizer.word_counts) + 1

print('VOCABULARY_SIZE: ', VOCABULARY_SIZE)

➔ output:

VOCABULARY_SIZE:  176

Let's play around with tokenizer dictionaries to see how we may convert characters to indices and vice-versa:

print(tokenizer.index_word[5])
print(tokenizer.index_word[20])

➔ output:

o
,

Let's try to convert character to index:

tokenizer.word_index['r']

➔ output:

8

To illustrate what kind of characters form all the recipes in our dataset we may print all of them as an array:

array_vocabulary = tokenizer.sequences_to_texts([[word_index] for word_index in range(VOCABULARY_SIZE)])
print([char for char in array_vocabulary])

➔ output:

['', ' ', 'e', 'a', 't', 'o', 'n', 'i', 'r', 's', 'l', 'd', 'h', 'c', 'u', 'p', '\n', 'm', 'g', 'b', ',', '.', 'f', 'w', '•', 'k', '1', 'v', 'y', '2', '/', '▪', '︎', 'S', '4', 'C', '-', '3', 'x', 'P', '5', '0', '(', ')', 'A', 'B', 'z', 'j', 'F', 'T', 'R', '📗', '🥕', '📝', 'I', 'M', ';', 'q', 'D', 'W', '8', 'G', '6', 'L', 'H', ':', '7', 'O', "'", 'E', 'K', '9', 'U', 'N', 'V', 'J', '®', '°', 'é', '"', 'Y', 'Q', '*', '!', 'Z', '–', '&', '%', 'ñ', 'è', '™', 'î', 'X', '?', '¿', '—', 'ç', '#', '½', 'í', '=', '’', 'â', '©', '¼', '+', '>', '$', '<', 'á', 'ó', 'ú', 'ï', 'É', 'û', ']', '[', 'ü', 'ê', 'à', '_', '\xad', '¾', '‚', '�', 'º', '⁄', 'ä', 'Ú', 'ù', '́', '}', 'ö', '{', 'ì', 'ô', '\x96', '”', '×', '˚', '»', '@', '§', '\\', '◊', '‱', '“', '‧', '\u202d', '⅛', 'å', 'fl', '`', 'Á', 'ë', '\x97', '\x1a', 'ø', '⅓', '|', 'ư', '\x92', '´', '‒', 'Â', '␣', '¤', '‟', '\xa0', 'ơ', 'ă', '̀', '⅞', '€', '~', '\x95']

These are all the characters our RNN model will work with. It will try to learn how to assemble these characters into sequences that will look like recipes.

Let's see how we may use tokenizer functions to convert text to indices:

tokenizer.texts_to_sequences(['📗 yes'])

➔ output:

[[51, 1, 28, 2, 9]]

Vectorizing the dataset

Now, once we have a vocabulary (character --> code and code --> character relations) we may convert the set of recipes from text to numbers (RNN works with numbers as an input and not with the texts).

dataset_vectorized = tokenizer.texts_to_sequences(dataset_filtered)

print('Vectorized dataset size', len(dataset_vectorized))

➔ output:

Vectorized dataset size 100212

This is how the beginning of the first vectorized recipe looks like:

print(dataset_vectorized[0][:10], '...')

➔ output:

[51, 1, 33, 10, 5, 23, 1, 35, 5, 5] ...

Let's see how can we convert vectorized recipe back to text representation:

def recipe_sequence_to_string(recipe_sequence):
    recipe_stringified = tokenizer.sequences_to_texts([recipe_sequence])[0]
    print(recipe_stringified)

recipe_sequence_to_string(dataset_vectorized[0])

➔ output:

📗 Slow Cooker Chicken and Dumplings

🥕

• 4 skinless, boneless chicken breast halves 
• 2 tablespoons butter 
• 2 (10.75 ounce) cans condensed cream of chicken soup 
• 1 onion, finely diced 
• 2 (10 ounce) packages refrigerated biscuit dough, torn into pieces 

📝

▪︎ Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.
▪︎ Cover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.

Add padding to sequences

We need all recipes to have the same length for training. To do that we'll use tf.keras.preprocessing.sequence.pad_sequences utility to add a stop word to the end of each recipe and to make them have the same length.

Let's check the recipes lengths:

for recipe_index, recipe in enumerate(dataset_vectorized[:10]):
    print('Recipe #{} length: {}'.format(recipe_index + 1, len(recipe)))

➔ output:

Recipe #1 length: 546
Recipe #2 length: 401
Recipe #3 length: 671
Recipe #4 length: 736
Recipe #5 length: 1518
Recipe #6 length: 740
Recipe #7 length: 839
Recipe #8 length: 667
Recipe #9 length: 1264
Recipe #10 length: 854

Let's pad all recipes with a STOP_SIGN:

dataset_vectorized_padded_without_stops = tf.keras.preprocessing.sequence.pad_sequences(
    dataset_vectorized,
    padding='post',
    truncating='post',
    # We use -1 here and +1 in the next step to make sure
    # that all recipes will have at least 1 stops sign at the end,
    # since each sequence will be shifted and truncated afterwards
    # (to generate X and Y sequences).
    maxlen=MAX_RECIPE_LENGTH-1,
    value=tokenizer.texts_to_sequences([STOP_SIGN])[0]
)

dataset_vectorized_padded = tf.keras.preprocessing.sequence.pad_sequences(
    dataset_vectorized_padded_without_stops,
    padding='post',
    truncating='post',
    maxlen=MAX_RECIPE_LENGTH+1,
    value=tokenizer.texts_to_sequences([STOP_SIGN])[0]
)

for recipe_index, recipe in enumerate(dataset_vectorized_padded[:10]):
    print('Recipe #{} length: {}'.format(recipe_index, len(recipe)))

➔ output:

Recipe #0 length: 2001
Recipe #1 length: 2001
Recipe #2 length: 2001
Recipe #3 length: 2001
Recipe #4 length: 2001
Recipe #5 length: 2001
Recipe #6 length: 2001
Recipe #7 length: 2001
Recipe #8 length: 2001
Recipe #9 length: 2001

After the padding all recipes in the dataset now have the same length and RNN will also be able to learn where each recipe stops (by observing the presence of a STOP_SIGN).

Here is an example of how a first recipe looks like after the padding.

recipe_sequence_to_string(dataset_vectorized_padded[0])

➔ output:

📗 Slow Cooker Chicken and Dumplings

🥕

• 4 skinless, boneless chicken breast halves 
• 2 tablespoons butter 
• 2 (10.75 ounce) cans condensed cream of chicken soup 
• 1 onion, finely diced 
• 2 (10 ounce) packages refrigerated biscuit dough, torn into pieces 

📝

▪︎ Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.
▪︎ Cover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣

All recipes now end with one or many ␣ signs. We expect our LSTM model to learn that whenever it sees the ␣ stop-character it means that the recipe is ended. Once the network will learn this concept it will put stop-character at the end of every newly generated recipe.

Create TensorFlow dataset

Up until now we were working with the dataset as with the NumPy array. It will be more convenient during the training process if we will convert a dataset NumPy array to a TensorFlow dataset. It will give us an ability to use such helpers functions as batch(), shuffle(), repeat(), prefecth() etc.:

dataset = tf.data.Dataset.from_tensor_slices(dataset_vectorized_padded)

print(dataset)

➔ output:

<TensorSliceDataset shapes: (2001,), types: tf.int32>

Let's see what the first recipe in the dataset looks like by using a TensorFlow dataset API this time:

for recipe in dataset.take(1):
    print('Raw recipe:\n', recipe.numpy(), '\n\n\n')
    print('Stringified recipe:\n')
    recipe_sequence_to_string(recipe.numpy())

➔ output:

Raw recipe:
 [ 51   1  33 ... 165 165 165] 

Stringified recipe:

📗 Slow Cooker Chicken and Dumplings

🥕

• 4 skinless, boneless chicken breast halves 
• 2 tablespoons butter 
• 2 (10.75 ounce) cans condensed cream of chicken soup 
• 1 onion, finely diced 
• 2 (10 ounce) packages refrigerated biscuit dough, torn into pieces 

📝

▪︎ Place the chicken, butter, soup, and onion in a slow cooker, and fill with enough water to cover.
▪︎ Cover, and cook for 5 to 6 hours on High. About 30 minutes before serving, place the torn biscuit dough in the slow cooker. Cook until the dough is no longer raw in the center.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣

Split examples on input and target texts

For each sequence we need to duplicate and shift it to form the input and target texts. For example, say the sequence_length is 4 and our text is Hello. The input sequence would be Hell, and the target sequence ello.

def split_input_target(recipe):
    input_text = recipe[:-1]
    target_text = recipe[1:]

    return input_text, target_text

dataset_targeted = dataset.map(split_input_target)

print(dataset_targeted)

➔ output:

<MapDataset shapes: ((2000,), (2000,)), types: (tf.int32, tf.int32)>

You may notice from the line above, that now each example in the dataset consists of two tuples: input and target. Let's print an example:

for input_example, target_example in dataset_targeted.take(1):
    print('Input sequence size:', repr(len(input_example.numpy())))
    print('Target sequence size:', repr(len(target_example.numpy())))
    print()

    input_stringified = tokenizer.sequences_to_texts([input_example.numpy()[:50]])[0]
    target_stringified = tokenizer.sequences_to_texts([target_example.numpy()[:50]])[0]

    print('Input:  ', repr(''.join(input_stringified)))
    print('Target: ', repr(''.join(target_stringified)))

➔ output:

Input sequence size: 2000
Target sequence size: 2000

Input:   '📗   S l o w   C o o k e r   C h i c k e n   a n d   D u m p l i n g s \n \n 🥕 \n \n •   4   s k i n l e'
Target:  '  S l o w   C o o k e r   C h i c k e n   a n d   D u m p l i n g s \n \n 🥕 \n \n •   4   s k i n l e s'

Each index of these vectors is processed as one time step by RNN. For the input at time step 0, the model receives the index for 📗 and tries to predict the index for (a space character) as the next character. At the next time-step, it does the same thing, but the RNN considers the previous step context in addition to the current input character.

for i, (input_idx, target_idx) in enumerate(zip(input_example[:10], target_example[:10])):
    print('Step {:2d}'.format(i + 1))
    print('  input: {} ({:s})'.format(input_idx, repr(tokenizer.sequences_to_texts([[input_idx.numpy()]])[0])))
    print('  expected output: {} ({:s})'.format(target_idx, repr(tokenizer.sequences_to_texts([[target_idx.numpy()]])[0])))

➔ output:

Step  1
  input: 51 ('📗')
  expected output: 1 (' ')
Step  2
  input: 1 (' ')
  expected output: 33 ('S')
Step  3
  input: 33 ('S')
  expected output: 10 ('l')
Step  4
  input: 10 ('l')
  expected output: 5 ('o')
Step  5
  input: 5 ('o')
  expected output: 23 ('w')
Step  6
  input: 23 ('w')
  expected output: 1 (' ')
Step  7
  input: 1 (' ')
  expected output: 35 ('C')
Step  8
  input: 35 ('C')
  expected output: 5 ('o')
Step  9
  input: 5 ('o')
  expected output: 5 ('o')
Step 10
  input: 5 ('o')
  expected output: 25 ('k')

Split up the dataset into batches

We have ~100k recipes in the dataset, and each recipe has two tuples of 2000 characters.

print(dataset_targeted)

➔ output:

<MapDataset shapes: ((2000,), (2000,)), types: (tf.int32, tf.int32)>

Let's print constants values:

print('TOTAL_RECIPES_NUM: ', TOTAL_RECIPES_NUM)
print('MAX_RECIPE_LENGTH: ', MAX_RECIPE_LENGTH)
print('VOCABULARY_SIZE: ', VOCABULARY_SIZE)

➔ output:

TOTAL_RECIPES_NUM:  100212
MAX_RECIPE_LENGTH:  2000
VOCABULARY_SIZE:  176

If we will feed the complete dataset during the training process to the model and then will try to do a back-propagation for all examples at once we might run out of memory and each training epoch may take too long to execute. To avoid the situation like this we need to split our dataset into batches.

# Batch size.
BATCH_SIZE = 64

# Buffer size to shuffle the dataset (TF data is designed to work
# with possibly infinite sequences, so it doesn't attempt to shuffle
# the entire sequence in memory. Instead, it maintains a buffer in
# which it shuffles elements).
SHUFFLE_BUFFER_SIZE = 1000

dataset_train = dataset_targeted \
  # Shuffling examples first.
  .shuffle(SHUFFLE_BUFFER_SIZE) \
  # Splitting examples on batches.
  .batch(BATCH_SIZE, drop_remainder=True) \
  # Making a dataset to be repeatable (it will never ends). 
  .repeat()

print(dataset_train)

➔ output:

<RepeatDataset shapes: ((64, 2000), (64, 2000)), types: (tf.int32, tf.int32)>

From the line above you may notice that our dataset now consists of the same two tuples of 2000 characters but now they are grouped in the batches by 64.

for input_text, target_text in dataset_train.take(1):
    print('1st batch: input_text:', input_text)
    print()
    print('1st batch: target_text:', target_text)

➔ output:

1st batch: input_text: tf.Tensor(
[[ 51   1  54 ... 165 165 165]
 [ 51   1  64 ... 165 165 165]
 [ 51   1  44 ... 165 165 165]
 ...
 [ 51   1  69 ... 165 165 165]
 [ 51   1  55 ... 165 165 165]
 [ 51   1  70 ... 165 165 165]], shape=(64, 2000), dtype=int32)

1st batch: target_text: tf.Tensor(
[[  1  54   4 ... 165 165 165]
 [  1  64   5 ... 165 165 165]
 [  1  44   6 ... 165 165 165]
 ...
 [  1  69   3 ... 165 165 165]
 [  1  55   3 ... 165 165 165]
 [  1  70   2 ... 165 165 165]], shape=(64, 2000), dtype=int32)

Build the model

We will use tf.keras.Sequential to define the model. For this experiment we will use the following layer types:

• tf.keras.layers.Embedding - the input layer (a trainable lookup table that will map the numbers of each character to a vector with embedding_dim dimensions),
• tf.keras.layers.LSTM - a type of RNN with size units=rnn_units (you can also use a GRU layer here),
• tf.keras.layers.Dense - the output layer, with VOCABULARY_SIZE outputs.

Figuring out how the Embedding Layer works

Let's do a quick detour and see how Embedding Layer works. It takes several char indices sequences (batch) as an input. It encodes every character of every sequence to a vector of tmp_embedding_size length.

tmp_vocab_size = 10
tmp_embedding_size = 5
tmp_input_length = 8
tmp_batch_size = 2

tmp_model = tf.keras.models.Sequential()
tmp_model.add(tf.keras.layers.Embedding(
  input_dim=tmp_vocab_size,
  output_dim=tmp_embedding_size,
  input_length=tmp_input_length
))
# The model will take as input an integer matrix of size (batch, input_length).
# The largest integer (i.e. word index) in the input should be no larger than 9 (tmp_vocab_size).
# Now model.output_shape == (None, 10, 64), where None is the batch dimension.
tmp_input_array = np.random.randint(
  low=0,
  high=tmp_vocab_size,
  size=(tmp_batch_size, tmp_input_length)
)
tmp_model.compile('rmsprop', 'mse')
tmp_output_array = tmp_model.predict(tmp_input_array)

print('tmp_input_array shape:', tmp_input_array.shape)
print('tmp_input_array:')
print(tmp_input_array)
print()
print('tmp_output_array shape:', tmp_output_array.shape)
print('tmp_output_array:')
print(tmp_output_array)

➔ output:

tmp_input_array shape: (2, 8)
tmp_input_array:
[[2 4 7 5 1 6 9 7]
 [3 6 8 1 4 0 1 2]]

tmp_output_array shape: (2, 8, 5)
tmp_output_array:
[[[-0.02229502 -0.02800617 -0.0120693  -0.01681594 -0.00650246]
  [-0.03046973 -0.03920818  0.04956308  0.04417323 -0.00446874]
  [-0.0215276   0.01532575 -0.02229529  0.02834387  0.02725342]
  [ 0.04567988  0.0141306   0.00877035 -0.02601192  0.00380837]
  [ 0.02969306  0.02994296 -0.00233263  0.00716375 -0.00847433]
  [ 0.04598364 -0.00704358 -0.01386416  0.01195388 -0.00309662]
  [-0.00137572  0.01275543 -0.02348721 -0.04825885  0.00527108]
  [-0.0215276   0.01532575 -0.02229529  0.02834387  0.02725342]]

 [[ 0.01082945  0.03824175 -0.00450991 -0.02865709  0.02502238]
  [ 0.04598364 -0.00704358 -0.01386416  0.01195388 -0.00309662]
  [ 0.02275398  0.03806095 -0.03491788  0.04705564  0.00167596]
  [ 0.02969306  0.02994296 -0.00233263  0.00716375 -0.00847433]
  [-0.03046973 -0.03920818  0.04956308  0.04417323 -0.00446874]
  [-0.02909902  0.04426369  0.00150937  0.04579213  0.02559013]
  [ 0.02969306  0.02994296 -0.00233263  0.00716375 -0.00847433]
  [-0.02229502 -0.02800617 -0.0120693  -0.01681594 -0.00650246]]]

LSTM Model

Let's assemble the model.

ℹ️ You may check Text generation with an RNN notebook from TensorFlow documentation for more details on model components.

def build_model(vocab_size, embedding_dim, rnn_units, batch_size):
    model = tf.keras.models.Sequential()

    model.add(tf.keras.layers.Embedding(
        input_dim=vocab_size,
        output_dim=embedding_dim,
        batch_input_shape=[batch_size, None]
    ))

    model.add(tf.keras.layers.LSTM(
        units=rnn_units,
        return_sequences=True,
        stateful=True,
        recurrent_initializer=tf.keras.initializers.GlorotNormal()
    ))

    model.add(tf.keras.layers.Dense(vocab_size))

    return model

model = build_model(
  vocab_size=VOCABULARY_SIZE,
  embedding_dim=256,
  rnn_units=1024,
  batch_size=BATCH_SIZE
)

model.summary()

➔ output:

Model: "sequential_13"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_13 (Embedding)     (64, None, 256)           45056     
_________________________________________________________________
lstm_9 (LSTM)                (64, None, 1024)          5246976   
_________________________________________________________________
dense_8 (Dense)              (64, None, 176)           180400    
=================================================================
Total params: 5,472,432
Trainable params: 5,472,432
Non-trainable params: 0
_________________________________________________________________

Let's visualize the model:

tf.keras.utils.plot_model(
    model,
    show_shapes=True,
    show_layer_names=True,
    to_file='model.png'
)

➔ output:

For each character the model looks up the embedding, runs the LSTM one time-step with the embedding as input, and applies the dense layer to generate logits predicting the log-likelihood of the next character:

Image source: Text generation with an RNN notebook.

The picture above illustrates GRU network, but you may easily replace GRU with LSTM.

Trying the model before training

Let's play around with un-trained model to see its interface (what input do we need and what output will we have) and let's see what model predicts before the training:

for input_example_batch, target_example_batch in dataset_train.take(1):
    example_batch_predictions = model(input_example_batch)
    print(example_batch_predictions.shape, "# (batch_size, sequence_length, vocab_size)")

➔ output:

(64, 2000, 176) # (batch_size, sequence_length, vocab_size)

To get actual predictions from the model we need to sample from the output distribution, to get actual character indices. This distribution is defined by the logits over the character vocabulary.

print('Prediction for the 1st letter of the batch 1st sequense:')
print(example_batch_predictions[0, 0])

➔ output:

Prediction for the 1st letter of the batch 1st sequense:
tf.Tensor(
[-9.0643829e-03 -1.9503604e-03  9.3381782e-04  3.7442446e-03
 -2.0541784e-03 -7.4054599e-03 -7.1884273e-03  2.6014952e-03
  4.8721582e-03  3.0045470e-04  2.6016519e-04 -4.1374690e-03
  5.3856964e-03  2.6284808e-03 -5.6002503e-03  2.6019611e-03
 -1.9491187e-03 -3.1097094e-04  6.3465843e-03  1.4640498e-03
  2.4560774e-03 -3.1256995e-03  1.4104056e-03  2.5478401e-04
  5.4266443e-03 -4.1188141e-03  3.6904984e-03 -5.8337618e-03
  3.6372752e-03 -3.1899021e-05  3.2178329e-03  1.5033322e-04
  5.2770867e-04 -8.1920059e-04 -2.2364906e-03 -2.3271297e-03
  4.4109682e-03  4.2381673e-04  1.0532180e-03 -1.4208974e-03
 -3.2446394e-03 -4.5869066e-03  4.3250201e-04 -4.3490473e-03
  3.7889536e-03 -9.2122913e-04  7.8936084e-04 -9.7079907e-04
  1.7070504e-03 -2.5260956e-03  6.7904620e-03  1.5470090e-03
 -9.4337866e-04 -1.5072266e-03  6.8939931e-04 -1.0795534e-03
 -3.1912089e-03  2.3665284e-03  1.7737487e-03 -2.3504677e-03
 -6.8649277e-04  9.6421910e-04 -4.1204207e-03 -3.8750230e-03
  1.9077851e-03  4.7145790e-05 -2.9846188e-03  5.8050319e-03
 -5.6210475e-04 -2.5910907e-04  5.2890396e-03 -5.8653783e-03
 -6.0040038e-06  2.3905798e-03 -2.9405006e-03  2.0132761e-03
 -3.5594390e-03  4.0282350e-04  4.7719614e-03 -2.4438011e-03
 -1.1028582e-03  2.0007135e-03 -1.6961874e-03 -4.2196750e-03
 -3.5689408e-03 -4.1934610e-03 -8.5307617e-04  1.5773368e-04
 -1.4612130e-03  9.5826073e-04  4.0543079e-04 -2.3562380e-04
 -1.5394683e-03  3.6650903e-03  3.5997448e-03  2.2390878e-03
 -6.8982318e-04  1.4068574e-03 -2.0531749e-03 -1.5443334e-03
 -1.8235333e-03 -3.2099178e-03  1.6660831e-03  1.2230751e-03
  3.8084832e-03  6.9559496e-03  5.7684043e-03  3.1751506e-03
  7.4234616e-04  1.1971325e-04 -2.7798198e-03  2.1485630e-03
  4.0362971e-03  6.4410735e-05  1.7432809e-03  3.2334479e-03
 -6.1469898e-03 -2.2205685e-03 -1.0864032e-03 -2.0876178e-07
  2.3065242e-03 -1.5816523e-03 -2.1492387e-03 -4.4033155e-03
  1.1003019e-03 -9.7132073e-04 -6.3941808e-04  3.0277157e-03
  2.9096641e-03 -2.4778468e-03 -2.9532036e-03  7.7463314e-04
  2.7473709e-03 -7.6333171e-04 -8.1811845e-03 -1.3959130e-03
  3.2840301e-03  6.0461317e-03 -1.3022404e-04 -9.4000692e-04
 -2.0096730e-04  3.3895797e-03  2.9710699e-03  1.9046264e-03
  2.5092331e-03 -2.0799250e-04 -2.2211851e-04 -3.4621451e-05
  1.9962704e-03 -2.3159904e-03  2.9832027e-03  3.3852295e-03
  3.4411502e-04 -1.9019389e-03 -3.6734296e-04 -1.4232489e-03
  2.6938838e-03 -2.8015859e-03 -5.7366290e-03  8.0239226e-04
 -6.2909431e-04  1.1508183e-03 -1.5899434e-04 -5.9326587e-04
 -4.1618512e-04  5.2454891e-03  1.2823739e-03 -1.7550631e-03
 -3.0120560e-03 -3.8433261e-03 -9.6873334e-04  1.9963509e-03
  1.8154597e-03  4.7434499e-03  1.7146189e-03  1.1544267e-03], shape=(176,), dtype=float32)

For each input character the example_batch_predictions array contains a vector of probabilities of what the next character might be. If probability at position 15 in that vector is, lets say, 0.3 and the probability at position 25 is 1.1 it means that we should better pick the character with the index 25 as next following character.

Since we want our network to generate different recipes (even for the same input), we can't just pick the maximum probability value. In this case we will end up with the same recipe being predicted by the network over and over again. What we will do instead is drawing samples from predictions (like the one printed above) by using tf.random.categorical() function. It will bring some fuzziness to the network. For example, let's say we have character H as an input, then, by sampling from categorical distribution, our network may predict not only the word He, but also words Hello, and Hi etc.

Understanding how tf.random.categorical works

# logits is 2-D Tensor with shape [batch_size, num_classes].
# Each slice [i, :] represents the unnormalized log-probabilities for all classes.
# In the example below we say that the probability for class "0"
# (element with index 0) is low but the probability for class "2" is much higher.
tmp_logits = [
  [-0.95, 0, 0.95],
];

# Let's generate 5 samples. Each sample is a class index. Class probabilities 
# are being taken into account (we expect to see more samples of class "2").
tmp_samples = tf.random.categorical(
    logits=tmp_logits,
    num_samples=5
)

print(tmp_samples)

➔ output:

tf.Tensor([[2 1 2 2 1]], shape=(1, 5), dtype=int64)

Sampling from LSTM predictions

sampled_indices = tf.random.categorical(
    logits=example_batch_predictions[0],
    num_samples=1
)

sampled_indices = tf.squeeze(
    input=sampled_indices,
    axis=-1
).numpy()

sampled_indices.shape

➔ output:

(2000,)

Let's see some sampled predictions for the first 100 chars of the recipe:

sampled_indices[:100]

➔ output:

array([ 64,  21,  91, 126, 170,  42, 146,  54, 125, 164,  60, 171,   9,
        87, 129,  28, 146, 103,  41, 101, 147,   3, 134, 171,   8, 170,
       105,   5,  44, 173,   5, 105,  17, 138, 165,  32,  88,  96, 145,
        83,  33,  65, 172, 162,   8,  29, 147,  58,  81, 153, 150,  56,
       156,  38, 144, 134,  13,  40,  17,  50,  27,  35,  39, 112,  63,
       139, 151, 133,  68,  29,  91,   2,  70, 112, 135,  31,  26, 156,
       118,  71,  49, 104,  75,  27, 164,  41, 117, 124,  18, 137,  59,
       160, 158, 119, 173,  50,  78,  45, 121, 118])

We may see now what our untrained model actually predicts:

print('Input:\n', repr(''.join(tokenizer.sequences_to_texts([input_example_batch[0].numpy()[:50]]))))
print()
print('Next char prediction:\n', repr(''.join(tokenizer.sequences_to_texts([sampled_indices[:50]]))))

➔ output:

Input:
 '📗   R e s t a u r a n t - S t y l e   C o l e s l a w   I \n \n 🥕 \n \n •   1   ( 1 6   o u n c e )   p'

Next char prediction:
 'H . î ⁄ ă ( “ I º Â 8 ̀ s % ù y “ © 0 ’ ‧ a ì ̀ r ă + o A € o + m × ␣ ︎ ñ ç ‱ ! S : ⅞ ´ r 2 ‧ D Q Á'

As you may see, the model suggests some meaningless predictions, but this is because it wasn't trained yet.

Training the model

We want to train our model to generate recipes as similar to the real ones as possible. We will use all data from dataset for training. There is not need to extract test or validation sub-sets in this case.

Attach an optimizer, and a loss function

We're going to use tf.keras.optimizers.Adam optimizer with tf.keras.losses.sparse_categorical_crossentropy() loss function to train the model:

# An objective function.
# The function is any callable with the signature scalar_loss = fn(y_true, y_pred).
def loss(labels, logits):
    entropy = tf.keras.losses.sparse_categorical_crossentropy(
      y_true=labels,
      y_pred=logits,
      from_logits=True
    )

    return entropy

example_batch_loss = loss(target_example_batch, example_batch_predictions)

print("Prediction shape: ", example_batch_predictions.shape, " # (batch_size, sequence_length, vocab_size)")
print("scalar_loss.shape:      ", example_batch_loss.shape)
print("scalar_loss:      ", example_batch_loss.numpy().mean())

➔ output:

Prediction shape:    (64, 2000, 176)  # (batch_size, sequence_length, vocab_size)
scalar_loss.shape:   (64, 2000)
scalar_loss:         5.1618285

Let's finally compile the model:

adam_optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)

model.compile(
    optimizer=adam_optimizer,
    loss=loss
)

Configuring callbacks

Early stopping callback

For model training process we may configure a tf.keras.callbacks.EarlyStopping callback. It will stop the training automatically in case if model is not improving for several epochs anymore:

early_stopping_callback = tf.keras.callbacks.EarlyStopping(
    patience=5,
    monitor='loss',
    restore_best_weights=True,
    verbose=1
)

Model checkpoints callback

Let's also configure a tf.keras.callbacks.ModelCheckpoint checkpoint that will allow us to periodically save trained weights to the file so that we could restore the model from weights afterwards.

# Create a checkpoints directory.
checkpoint_dir = 'tmp/checkpoints'
os.makedirs(checkpoint_dir, exist_ok=True)

checkpoint_prefix = os.path.join(checkpoint_dir, 'ckpt_{epoch}')
checkpoint_callback=tf.keras.callbacks.ModelCheckpoint(
    filepath=checkpoint_prefix,
    save_weights_only=True
)

Execute the training

Let's train our model for 500 epochs with 1500 steps per each epoch. For each epoch step the batch of 64 recipes will be fetched and gradient descent will be executed for those 64 recipes of length 2000 step by step.

If you're experimenting with training parameters it might make sense to reduce the number of epochs to, let's say 20 along with the number of steps per epoch and then see how the model performs under that conditions. If the model improves its performance you may add more data (steps and epochs) to the training process. It might save you some time while you adjust parameters.

EPOCHS = 500
INITIAL_EPOCH = 1
STEPS_PER_EPOCH = 1500

print('EPOCHS:          ', EPOCHS)
print('INITIAL_EPOCH:   ', INITIAL_EPOCH)
print('STEPS_PER_EPOCH: ', STEPS_PER_EPOCH)

➔ output:

EPOCHS:           500
INITIAL_EPOCH:    1
STEPS_PER_EPOCH:  1500

Let's launch the training:

history = model.fit(
    x=dataset_train,
    epochs=EPOCHS,
    steps_per_epoch=STEPS_PER_EPOCH,
    initial_epoch=INITIAL_EPOCH,
    callbacks=[
        checkpoint_callback,
        early_stopping_callback
    ]
)

# Saving the trained model to file (to be able to re-use it later).
model_name = 'recipe_generation_rnn_raw.h5'
model.save(model_name, save_format='h5')

Visualizing training progress

def render_training_history(training_history):
    loss = training_history.history['loss']

    plt.title('Loss')
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.plot(loss, label='Training set')
    plt.legend()
    plt.grid(linestyle='--', linewidth=1, alpha=0.5)
    plt.show()

render_training_history(history)

➔ output:

ℹ️ On the chart above only first 10 epochs are presented.

We can see from the chart that model performance is getting better during the training. It means that the model learns to predict next characters in a way that the final sequence looks similar to some real recipe texts.

Generating recipes

Restore the model from the latest checkpoint

To keep this prediction step simple, we will restore the saved model and rebuild it with a batch size of 1. Because of the way the RNN state is passed from time-step to time-step, the model only accepts a fixed batch size once built. To run the model with a different batch_size, we need to rebuild the model and restore the weights from the checkpoint.

tf.train.latest_checkpoint(checkpoint_dir)

➔ output:

'tmp/checkpoints/ckpt_1'

Let's rebuild the model with batch size of 1 and load trained weights to it:

simplified_batch_size = 1

model_simplified = build_model(vocab_size, embedding_dim, rnn_units, simplified_batch_size)
model_simplified.load_weights(tf.train.latest_checkpoint(checkpoint_dir))
model_simplified.build(tf.TensorShape([simplified_batch_size, None]))

model_simplified.summary()

➔ output:

Model: "sequential_6"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_6 (Embedding)      (1, None, 256)            45056     
_________________________________________________________________
lstm_5 (LSTM)                (1, None, 1024)           5246976   
_________________________________________________________________
dense_5 (Dense)              (1, None, 176)            180400    
=================================================================
Total params: 5,472,432
Trainable params: 5,472,432
Non-trainable params: 0
_________________________________________________________________

Let's double check that input shape is simplified:

model_simplified.input_shape

➔ output:

(1, None)

The prediction loop

To use our trained model for recipe generation we need to implement a so-called prediction loop. The following code block generates the text using the loop:

• It starts by choosing a start string, initializing the RNN state and setting the number of characters to generate.
• It gets the prediction distribution of the next character using the start string, and the RNN state.
• Then, it uses a categorical distribution to calculate the index of the predicted character. It uses this predicted character as the next input to the model.
• The RNN state returned by the model is fed back into the model so that it now has more context, instead of only one character. After predicting the next character, the modified RNN states are again fed back into the model, which is how it learns as it gets more context from the previously predicted characters.

Image source: Text generation with an RNN notebook.

The temperature parameter here defines how fuzzy or how unexpected the generated recipe is going to be. Low temperatures results in more predictable text. Higher temperatures result in more surprising text. You need to experiment to find the best setting. We will do some experimentation with different temperatures below.

def generate_text(model, start_string, num_generate = 1000, temperature=1.0):
    # Evaluation step (generating text using the learned model)

    padded_start_string = STOP_WORD_TITLE + start_string

    # Converting our start string to numbers (vectorizing).
    input_indices = np.array(tokenizer.texts_to_sequences([padded_start_string]))

    # Empty string to store our results.
    text_generated = []

    # Here batch size == 1.
    model.reset_states()
    for char_index in range(num_generate):
        predictions = model(input_indices)
        # remove the batch dimension
        predictions = tf.squeeze(predictions, 0)

        # Using a categorical distribution to predict the character returned by the model.
        predictions = predictions / temperature
        predicted_id = tf.random.categorical(
            predictions,
            num_samples=1
        )[-1, 0].numpy()

        # We pass the predicted character as the next input to the model
        # along with the previous hidden state.
        input_indices = tf.expand_dims([predicted_id], 0)

        next_character = tokenizer.sequences_to_texts(input_indices.numpy())[0]

        text_generated.append(next_character)

    return (padded_start_string + ''.join(text_generated))

Figuring out proper temperature for prediction loop

Now, let's use generate_text() to actually generate some new recipes. The generate_combinations() function goes through all possible combinations of the first recipe letters and temperatures. It generates 56 different combinations to help us figure out how the model performs and what temperature is better to use.

def generate_combinations(model):
    recipe_length = 1000
    try_letters = ['', '\n', 'A', 'B', 'C', 'O', 'L', 'Mushroom', 'Apple', 'Slow', 'Christmass', 'The', 'Banana', 'Homemade']
    try_temperature = [1.0, 0.8, 0.4, 0.2]

    for letter in try_letters:
        for temperature in try_temperature:
            generated_text = generate_text(
                model,
                start_string=letter,
                num_generate = recipe_length,
                temperature=temperature
            )
            print(f'Attempt: "{letter}" + {temperature}')
            print('-----------------------------------')
            print(generated_text)
            print('\n\n')

To avoid making this article too long only some of those 56 combinations will be printed below.

generate_combinations(model_simplified)

➔ output:

Attempt: "A" + 1.0
-----------------------------------
📗 Azzeric Sweet Potato Puree

🥕

• 24 large baking potatoes, such as Carn or Marinara or 1 (14-ounce) can pot wine
• 1/4 pound unsalted butter, cut into small pieces
• 1/2 cup coarsely chopped scallions

📝

▪︎ Bring a large pot of water to a boil, place a large nonstick skillet over medium-high heat, add All Naucocal Volves. Reduce heat to medium and cook the potatoes until just cooked through, bubbles before adding the next layer, about 10 to 12 minutes. Remove ground beans and reserve. Reserve the crumb mixture for about 6 greased. Let cool 2 minutes. Strain soak into a glass pitcher. Let cool in ice. Add short-goodfish to the batter and stir to dissolve. Pour in the cheese mixture and whisk until smooth. Set aside for 20 seconds more. Remove dumplings and cheese curds. Spread 1/3 cup of the mixture on each circle for seal ballo. Transfer mixture into a greased 9-by-11-inch baking dish and chill for 20 minutes.
▪︎ Bake, covered, for 30 minutes. Serve warm.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣



Attempt: "A" + 0.4
-----------------------------------
📗 Apricot "Cookie" Cakes

🥕

• 1 cup all-purpose flour
• 1 cup corn flour
• 1 cup sugar
• 1 tablespoon baking powder
• 1 teaspoon salt
• 1 teaspoon ground cinnamon
• 1 cup grated Parmesan
• 1 cup pecans, chopped
• 1/2 cup chopped pecans
• 1/2 cup raisins

📝

▪︎ Preheat oven to 350 degrees F.
▪︎ Butter and flour a 9 by 13-inch baking dish. In a medium bowl, whisk together the flour, sugar, baking powder, baking soda and salt. In a small bowl, whisk together the eggs, sugar, and eggs. Add the flour mixture to the butter mixture and mix until just combined. Stir in the raisins and pecans and transfer to the prepared pan. Spread the batter over the top of the crust. Bake for 15 minutes. Reduce the oven temperature to 350 degrees F, and bake until the cupcakes are set and the top is golden brown, about 20 minutes more. Transfer the cake to a wire rack to cool to room temperature. Refrigerate until ready to serve.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣



Attempt: "A" + 0.2
-----------------------------------
📗 Alternative to the Fondant

🥕

• 1 cup sugar
• 1 cup water
• 1 cup heavy cream
• 1 teaspoon vanilla extract
• 1/2 cup heavy cream
• 1/2 cup heavy cream
• 1 teaspoon vanilla extract
• 1/2 cup chopped pecans

📝

▪︎ In a saucepan over medium heat, combine the sugar, sugar, and corn syrup. Cook over medium heat until the sugar is dissolved. Remove from the heat and stir in the vanilla. Refrigerate until cold. Stir in the chocolate chips and the chocolate chips. Serve immediately.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣



Attempt: "B" + 0.4
-----------------------------------
📗 Battered French Toast with Bacon, Bacon, and Caramelized Onions and Pecorino

🥕

• 1/2 pound squid (shredded carrots)
• 1 small onion, diced
• 1 small green pepper, seeded and cut into strips
• 1 red bell pepper, stemmed, seeded and cut into 1/4-inch dice
• 1 small onion, chopped
• 1 green bell pepper, chopped
• 1 cup chicken stock
• 1 cup heavy cream
• 1/2 cup shredded sharp Cheddar
• 1 teaspoon ground cumin
• 1 teaspoon salt
• 1 teaspoon freshly ground black pepper

📝

▪︎ Preheat the oven to 350 degrees F.
▪︎ For the bacon mixture: In a large bowl, combine the cheese, sour cream, mustard, salt, pepper, and hot sauce. Stir together and mix well. Fold in the milk and set aside.
▪︎ For the filling: In a large bowl, mix the flour and salt and pepper, to taste. Add the beaten eggs and mix to combine. Set aside.
▪︎ For the topping: Mix the cream cheese with the mayonnaise, salt and pepper in a medium bowl. Add the chicken and toss to coat the other side. Transfer the mixture to the prepared





Attempt: "C" + 1.0
-----------------------------------
📗 Crema battered Salmon

🥕

• 1 cup fresh cranberries (from 4 tablespoons left of 4 egg whites)
• 3 teaspoons sugar
• 1 tablespoon unsalted butter
• 2 tablespoons truffle oil
• Coarse salt
• Freshly ground black pepper

📝

▪︎ Place cornmeal in a small serving bowl, and combine it. Drizzle milk over the plums and season with salt and pepper. Let stand for about 5 minutes, until firm. Serve immediately.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣



Attempt: "C" + 0.8
-----------------------------------
📗 Classic Iseasteroles

🥕

• 3 cups milk
• 3/4 cup coconut milk
• 1/2 cup malted maple syrup
• 1/2 teaspoon salt
• 3 cups sugar
• 4 1-inch strawberries, sliced into 1/4-inch pieces
• 1/2 teaspoon ground cinnamon

📝

▪︎ Place the cherries in a small saucepan; sprinkle with the sugar. Bring to a simmer over medium-low heat, then remove from the heat. Let stand until the coconut fluffy, about 15 to 20 minutes. Drain the coconut oil in a stream, whisking until combined. Add the cream, espresso and cocoa powder and stir to combine. Cover and refrigerate until ready to serve. Makes 10 to 12 small springs in the same fat from the surface of the bowl, which using paper colors, and freeze overnight.
▪︎ Meanwhile, combine the cream, sugar, vanilla and salt in a medium saucepan. Cook over medium heat until the sugar dissolves and the sugar melts and begins to boil, about 5 minutes. Remove from the heat and stir in the vanilla.
▪︎ To serve, carefully remove the pops from the casserole and put them in



Attempt: "C" + 0.4
-----------------------------------
📗 Cinnamon Corn Cakes with Coconut Flour and Saffron Sauce

🥕

• 3 cups shredded sharp Cheddar
• 1 cup grated Parmesan
• 2 cups shredded sharp Cheddar
• 1 cup grated Parmesan
• 1 cup shredded part-skim mozzarella cheese
• 1 cup grated Parmesan
• 1 cup grated Parmesan
• 1 cup grated Parmesan
• 1 teaspoon kosher salt
• 1/2 teaspoon freshly ground black pepper

📝

▪︎ Preheat the oven to 400 degrees F. Line a baking sheet with a silpat and preheat the oven to 350 degrees F.
▪︎ In a large bowl, combine the masa harina, cumin, cayenne, and salt and pepper. Dredge the pasta in the flour and then dip in the egg mixture, then dip in the eggs, then dip in the egg mixture and then dredge in the breadcrumbs. Place the breaded cheese on a sheet tray. Bake until the crust is golden brown and the filling is bubbling, about 25 to 30 minutes. Remove from the oven and serve hot.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣




Attempt: "L" + 0.4
-----------------------------------
📗 Lighted Flan with Chocolate and Pecans

🥕

• 2 cups milk
• 1 cup sugar
• 1 teaspoon vanilla extract
• 1 cup heavy cream
• 1/2 cup heavy cream
• 1 tablespoon powdered sugar
• 1 teaspoon vanilla extract
• 1/2 cup heavy cream
• 1/2 teaspoon ground cinnamon
• 1/2 teaspoon ground nutmeg
• 1/2 cup chopped pecans

📝

▪︎ Watch how to make this recipe.
▪︎ In a small saucepan, combine the sugar, salt, and a pinch of salt. Cook over medium heat, stirring occasionally, until the sugar has dissolved. Remove from the heat and set aside to cool. Remove the cherries from the refrigerator and place in the freezer for 1 hour.
▪︎ In a blender, combine the milk, sugar, vanilla, salt and water. Blend until smooth. Pour the mixture into a 9-by-13-inch glass baking dish and set aside.
▪︎ In a small saucepan, combine the remaining 2 cups sugar, the vanilla, and 2 cups water. Bring the mixture to a boil, and then reduce the heat to low. Cook until the sugar is dissolved, about 5 minutes. Remove from the heat an



Attempt: "L" + 0.2
-----------------------------------
📗 Lighted Fondanta with Chocolate and Cream Cheese Frosting

🥕

• 1 cup heavy cream
• 1 tablespoon sugar
• 1 tablespoon vanilla extract
• 1 teaspoon vanilla extract
• 1 cup heavy cream
• 1 cup heavy cream
• 1/2 cup sugar
• 1 teaspoon vanilla extract
• 1 teaspoon vanilla extract
• 1/2 cup chopped pistachios

📝

▪︎ Preheat the oven to 350 degrees F.
▪︎ In a large bowl, combine the cream cheese, sugar, eggs, vanilla, and salt. Stir until smooth. Pour the mixture into the prepared baking dish. Sprinkle with the remaining 1/2 cup sugar and bake for 15 minutes. Reduce the heat to 350 degrees F and bake until the crust is golden brown, about 15 minutes more. Remove from the oven and let cool completely. Spread the chocolate chips on the parchment paper and bake until the chocolate is melted and the top is golden brown, about 10 minutes. Set aside to cool.
▪︎ In a medium bowl, whisk together the egg yolks, sugar, and vanilla until smooth. Stir in the cream and continue to beat until the chocolate



Attempt: "Mushroom" + 1.0
-----------------------------------
📗 Mushroom and Bacon Soup with Jumbo Sugar Coating

🥕

• 2 tablespoons vegetable oil
• 1 2/3 pounds red cabbage, shredded, about 4 cups of excess pasted dark ends of fat, and pocked or firm
• 2 red bell peppers, cored, seeded and diced
• 1 poblano pepper, chopped
• 3 medium carrots, finely chopped
• 1/2 medium pinch saffron
• 4 cups water
• 2 cups mushrooms or 1/2 cup frozen Sojo Bean red
• Salt and freshly ground black pepper
• 1 pound andouille sausage
• 1 gallon vegetable broth
• Chopped fresh parsley, cilantro leaves, for garnish

📝

▪︎ In a large Dutch oven for gas burner, heat oil over moderate heat. Add the leeks to the pot, scraping the bottom of the skillet. Add the beans and sausage and sprinkle the reserved potatoes with some orange juice cooked sausage (such as The Sauce.) Add roasted vegetables and pinto beans, mozzarella, basil and bamboo shoots. Simmer rice until soup is absorbed, 15 to 20 minutes.
▪︎ Bring another pan of water to a boil and cook shrimp for 5 minutes. While onions



Attempt: "Mushroom" + 0.8
-----------------------------------
📗 Mushrooms with Lentil Stewed Shallots and Tomatoes

🥕

• 1 tablespoon olive oil
• 3 cloves garlic, smashed
• Kosher salt
• 1 1/2 pounds lean ground turkey
• 1 cup coarsely peeled tart apples
• 2 tablespoons chopped garlic
• 1 teaspoon ground cumin
• 1/2 teaspoon cayenne pepper
• 1 teaspoon chopped fresh thyme
• 3/4 cup chopped fresh basil
• 1/2 small carrot, halved lengthwise and cut into 1/2-inch pieces
• 1 roasted red pepper, halved and sliced vertically diced and separated into rough chops
• 3 tablespoons unsalted butter
• 2 cups shredded mozzarella
• 1/4 cup grated parmesan cheese
• 1/4 cup prepared basil pesto

📝

▪︎ Stir the olive oil, garlic, thyme and 1 teaspoon salt in a saucepan; bring to a simmer over medium heat. Remove from the heat. Add the basil and toast the soup for 2 minutes.
▪︎ Meanwhile, heat 4 to 4 inches vegetable oil in the skillet over medium-high heat. Add the olive oil, garlic, 1/2 teaspoon salt and 1/2 teaspoon pepper and cook, stirring often, until cooked through, a



Attempt: "Mushroom" + 0.4
-----------------------------------
📗 Mushroom Ravioli with Chickpeas and Shiitake Mushrooms and Sun-Dried Tomatoes

🥕

• 1 pound zucchini
• 1 cup chicken broth
• 1 cup fresh basil leaves
• 1/2 cup chopped fresh basil leaves
• 1/2 cup grated Parmesan
• 1 teaspoon salt
• 1/2 teaspoon freshly ground black pepper
• 1 teaspoon chopped fresh thyme
• 1 teaspoon fresh lemon juice
• 2 cups chicken broth
• 1/2 cup grated Parmesan
• 1/2 cup grated Parmigiano-Reggiano

📝

▪︎ Preheat oven to 450 degrees F.
▪︎ Place the bread cubes in a large bowl. Add the basil, parsley, olive oil, parsley, thyme, basil, salt and pepper and toss to coat. Spread the mixture out on a baking sheet and bake until the sausages are cooked through, about 20 minutes. Serve immediately.
▪︎ In a small saucepan, bring the chicken stock to a boil. Reduce the heat to low and cook the soup until the liquid is absorbed. Remove from the heat and stir in the parsley, shallots and season with salt and pepper. Serve immediately.
␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣␣



Attempt: "Mushroom" + 0.2
-----------------------------------
📗 Mushroom and Spicy Sausage Stuffing

🥕

• 1 tablespoon olive oil
• 1 medium onion, chopped
• 2 cloves garlic, minced
• 1 cup frozen peas
• 1 cup frozen peas
• 1/2 cup chopped fresh parsley
• 1/2 cup grated Parmesan
• 1/2 cup grated Parmesan
• 1 teaspoon salt
• 1/2 teaspoon freshly ground black pepper
• 1 cup shredded mozzarella
• 1/2 cup grated Parmesan
• 1 cup shredded mozzarella
• 1 cup shredded mozzarella cheese

📝

▪︎ Preheat the oven to 350 degrees F.
▪︎ Bring a large pot of salted water to a boil. Add the pasta and cook until al dente, about 6 minutes. Drain and reserve.
▪︎ Meanwhile, heat the olive oil in a large skillet over medium-high heat. Add the shallots and saute until tender, about 3 minutes. Add the garlic and cook for 1 minute. Add the sausage and cook until the shallots are tender, about 3 minutes. Add the sausage and cook until tender, about 2 minutes. Add the garlic and cook, stirring, until the garlic is lightly browned, about 1 minute. Add the sausage and cook until the s

Interactive model demo

You may use 🎨 Cooking recipes generator demo to play around with this model, input text, and temperature parameters to generate some random recipes right in your browser.

Things to improve

This out of scope for this article but model still has the following issues that need to be addressed:

• We need to get rid of duplicates on the ingredients section.
• Recipe sections (name, ingredients and cooking steps) are disconnected most of the time, meaning that we may see, let's say, mushrooms in the ingredients section, but they are not mentioned in the name of the recipe or in the cooking steps.