Foodie API for healthy foodies! Restaurant recommender API and dish calories calculations in real time
This API aims to be both a restaurant recommender and a calories calculator. This last one is achieved by means of machine learning: neural nets trainning on 5 different dishes.
The final deployment can be found in the following link:
https://apifoodie.herokuapp.com/
The script is developed in Python and deployed as a Docker image on Heroku.
This project is my final project of Ironhack's Data Analytics Fulltime Bootcamp.
For this project, the following APIs have been used. Specially thanks to their contributors for spreading so rich information:
- Food Images (Food-101)--> https://www.kaggle.com/kmader/food41 The dataset contains a number of different subsets of the full food-101 data. The idea is to make a more exciting simple training set for image analysis than CIFAR10 or MNIST. For this reason the data includes massively downscaled versions of the images to enable quick tests.
The data has been reformatted as HDF5 and specifically Keras HDF5Matrix which allows them to be easily read in. The file names indicate the contents of the file.
On behalf of data protection, the dataset has not been uploaded to Github.
- FourSquare API--> https://developer.foursquare.com/docs/api-reference
This dataset contains over 16000 movie an TV series and their appearance on different streamming platforms.
- Spoonacular API--> https://spoonacular.com/food-api/
The spoonacular Nutrition, Recipe, and Food API allows you to access over 365,000 recipes and 86,000 food products. It makes it possible to search for recipes using natural language queries, such as "gluten free brownies without sugar" or "low fat vegan cupcakes."
It also calculates the nutritional information for any recipe, which will be used for this project.
-
README.txt: As a resume for the content of the project an its development
-
server.py: It contains the main script for the project
-
src: It contains relevant files so that the code can be runned. a)Inception_V6.ipynb: The NN trainning on Google Colab
b)api_gen.py: Definition of endpoints along the API
c)data_extraction.py: Auxiliar functions for the API
d)image_prediction.ipynb: Auxiliar jupyter notebook where the Inception model is loaded, and a photo is analyzed
e)html-> It contains every html file for the endpoints' display
-
static: Where APIs images are saved; i.e: cover background, icon, etc.
-
templates: Where templates for the endpoints are saved
As this project has two main components, the restaurant recommender and the calories calculator, both tasks were built in paralell. The clue here was the use of Google Colab, as training the ML models in the cloud gave me the opportunity to develop the endpoints of my API whilst the NN were working "by their own".
So, as far as the restaurants recommender is concerned:
-
I decided to display a Folium map by making a request to Foursquare's API in real time. This way, I avoided saving a huge dataset in a database as MongoDB, and I opened the access to the recommender worldwide. Otherwise, it would have been immeasurable to do it.
-
For each request, a Folium map is created in the location specified by the user. In case a location is not defined, the IP address's location is considerated. Furthermore, 5 markers are drawn in the map, with up to the 5 nearest matches.
-
Defensive programming is also beared in mind, in case the location is not written properly, the query did not return any restaurant and so on.
In terms of calories calculator:
-The main task here is to train an image classification neural net for 5 labels (i.e. sushi, pizza, tacos, spaguetti bolognese and burger). Each label is trained with 1000 images. In this case, InceptionV3 model has been used with 14 epochs, adding afterwars a GlobalAveragePooling2D layer, a Dense layer and a Dropout layer of 70% to ensure dismission for non-representative photos.
The model training has been developed in Google Colab, so as to take advantage of their GPU possibilities.
An early stopping has been set for the case that the validation loss function increases during 5 epochs. The selected loss function is the categorical crossentropy.
As a result, the following values have been obtained for both loss function and accuracy:
Where the confusion matrix looks as follows:
As last, an image is created by means of Docker, and the final deployment is developed on Heroku.
In progress!!!