Food Classifier Application

The app is found here:

https://singapore-food-classifier-app.herokuapp.com/

Preview of the web application

Table of contents

• Objective
• Dataset
• Data Preprocessing
• Modelling
• Project Usage
• Deployment

Objective

This is an end-to-end project, where the goal is to create a classification model to predict if a picture is one of 12 Singaporean food and serve it through a web application. A trained model is used to predict 12 different food classes, namely 🦀chilli crab, 🥟curry puff, 🍤dim sum, 🍧ice kacang, 🥪kaya toast, 🍚nasi ayam, 🌯popiah, 🥞roti prata, 🐠sambal stingray, 🍢satay, 🍵tau huay or 🍜wanton noodle. 😋

---

Dataset

This dataset consist of 12 classes with the breakdown shown in the table below:

Food	Number of Images
dim_sum	171
curry_puff	105
sambal_stingray	103
chilli_crab	102
satay	102
popiah	101
roti_prata	101
kaya_toast	100
wanton_noodle	100
ice_kacang	90
nasi_ayam	85
tau_huay	64

---

Data Preprocessing

The images were split into train, test, validation:

• 80% of our dataset to the training set
• held out 10% of our dataset for the validation set
• allocated 10% of our dataset to the test set

The images were loaded using tensorflow.keras.preprocessing.image.ImageDataGenerator and performed the following:

# As the parameters indicate, random rotations, 
# zooms, shifts, shears, and flips will be performed during in-place/on-the-fly data augmentation.

# ImageDataGenerator
train_datagen = ImageDataGenerator(rescale=1./255,
                                   zoom_range=0.2, # zoom
                                   rotation_range=20, # rotation
                                   width_shift_range=0.2, # horizontal shift
                                   height_shift_range=0.2, # vertical shift
                                   shear_range=0.15, # shearing
                                   horizontal_flip=True, # horizontal flip
                                   fill_mode="nearest")


valid_datagen = ImageDataGenerator(rescale=1./255)

test_datagen = ImageDataGenerator(rescale=1./255)

Note: Data augmentation is only done on the training set.

Input parameters used for .flow_from_directory(directory):

# set the target size as 256x256 for ImageNet-scale networks 
target_size = (256, 256)
batch_size = 32

---

Modelling

Model Architecture

The base model is created from the MobileNet V2 model developed at Google. This is pre-trained on the ImageNet dataset, a large dataset consisting of 1.4M images and 1000 classes. The convolutional base is freezed and used as a feature extractor. Additional classifier head is added on top of the convolutional base and trained on a training set of about 975 images belonging to 12 classes. Optimizer used was Adam with a learning rate of 0.001.

Summary of the model architecture:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
mobilenetv2_1.00_224 (Model) (None, 8, 8, 1280)        2257984   
_________________________________________________________________
global_average_pooling2d (Gl (None, 1280)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 12)                15372     
=================================================================
Total params: 2,273,356
Trainable params: 15,372
Non-trainable params: 2,257,984

Model Performance

The test accuracy achieved was 91%.

---

Project Usage

Prerequisites

Flask is used in this project as our web application framework because it is a lightweight WSGI web application framework. It is designed to make getting started quick and easy, with the ability to scale up to complex applications. It began as a simple wrapper around Werkzeug and Jinja and has become one of the most popular Python web application frameworks. We will also require Tensorflow for our machine learning framework.

The conda.yml file has all the Conda packages that this project depends on.

Usage

Model inference (frontend)

Select an image to upload on the main web page. The model only accepts .jpeg, .jpg and .png files.

Response

Displays the image uploaded and returns the result of the prediction

An example of the result:

There's a 72.15% probability that this is a ice_kacang.

---

Deployment

The web application is deployed on a Docker container and hosted on Heroku.

The source folder structure is as such:

<project_dir>
 ├── src                                # Contains the necessary source files to run the flask app
 │   ├── app.py				            # Python file of the flask app	
 │   ├── inference.py		            # Python file for model inference
 │   ├── train.py		            # Model training source code
 │   ├── static                         # Contains static files required for the flask app
 |   |   ├── uploads                    # Folder to contain the images uploaded for model inference
 |   |   ├── css
 |   |   |   └── template.css		    # Template CSS File to add some style formatting to the HTML structure
 │   ├── templates                      # Contains the HTML templates used in the flask app

Before we can deploy, we should ensure that the Docker image can be built and the container with the app can run locally.

Run docker image on local

In the base folder, run the following commands.

Build a docker image:

docker build -t tensorfood-app .

Verify that the docker image was built:

docker images

Running the flask app in docker container:

docker run -p 5000:5000 tensorfood-app

The web application should be up and running on: http://localhost:5000/ .