FoodImageClassifier

Did this for my Deep Learning Module Assignment 1. Based off the food-101 dataset on Kaggle, which was used in a DL competition.

Achieved an 'A' grade / 4.0 GPA for this assignment. (Models, Presentation which is not included and Final Report).

Assigned 10 types of food.

• caprese_salad
• hamburger
• creme_brulee
• baklava
• omelette
• paella
• peking_duck
• cheesecake
• sushi
• beef_carpaccio

Problem and Objective

The main objective of this assignment is to experiment with various Convolutional Neural Networks (CNNs) to determine which model is the best for classifying new images.

The problem is of a multi-class, single label type of classification as there are multiple food categories, with the condition that one image can only belong to one category.

Recommended Hardware

• At least 16GB RAM
• An i7 Intel Processor (or equivalent)
• A NVIDIA GPU with 2GB VRAM (or equivalent)
• NvMe SSD for unzipping the large (food-101) dataset

Software Packages

• Anaconda
  • Tensorflow (which includes Keras) and Tensorflow-Base
  • Tensorflow GPU (optional)

conda install tensorflow
conda install tensorflow-gpu

Preprocessing Steps

1. Download food-101 zip from this website. Make sure you download the .zip file and NOT the .zip.zip file. Warning: If you have a non-NVMe SSD, this could take up to 8 hours to unzip

2. Place Image_Preprocessing.ipynb, 40.txt (or any other specified .txt file from the Food_List.zip) and the extracted image file (food-101/food-101/food-101/images) into the same base directory of choice. An example is given below:

DL Assignment 1
|- Image_Preprocessing.ipynb
|- 40.txt
|- Assignment_1.ipynb
|- images (102 items)
   |- apple_pie (1000 images)
      |-
   |- baby_back_ribs
   |- baklava
   ...
   |- tiramisu
   |- tuna_tartare
   |- waffles
   |- .DS Store (can ignore)

3. Run Image_Preprocessing.ipynb with 40.txt and your path of the images folder specified. The folder structure should be modified slightly as shown below.

DL Assignment 1
|- Image_Preprocessing.ipynb
|- 40.txt
|- train
|- test
|- validate
|- Assignment_1.ipynb
|- images (102 items)
   |- apple_pie (1000 images)
      |-
   |- baby_back_ribs
   |- baklava
   ...
   |- tiramisu
   |- tuna_tartare
   |- waffles
   |- .DS Store (can ignore)

4. Create a new folder named Assigned_Images and move the train, test and validate folders to this folder. The reultant directory structure is shown below.

DL Assignment 1
|- Image_Preprocessing.ipynb
|- 40.txt
|- Assigned_Images
   |- train
   |- test
   |- validate
|- Assignment_1.ipynb
|- images (102 items)
   |- apple_pie (1000 images)
      |-
   |- baby_back_ribs
   |- baklava
   ...
   |- tiramisu
   |- tuna_tartare
   |- waffles
   |- .DS Store (can ignore)

Building the Model

For all my models, I chose an image size of 150 x 150 and used data augmentation for all models, which aided in increasing the diversity of the relatively small dataset of 1000 images per food category. For most models, I used a standard set of augmentation parameters, which can be found in the notebook (base models).

Results

Model Number:	Description/Type:	Test Accuracy (%):
0	Non-PT RMSProp Base	64.2
1	Non-PT RMSProp Lower LR	68
2	Non-PT RMSProp Change Steps Per Epoch	57
3	Non-PT RMSProp Higher LR	59.8
4	Non-PT Adam Lower LR	55.8
5	Non-PT Adam Base	71.2
6	Non-PT SGD+Momentum Base	67.4
7	Non-PT SGD+Momentum Lower Momentum	64.4
8	Non-PT Adam Regularization	69.6
9	PT VGG16 Fine Tuning 2	74.6
10	PT VGG16 Base*	67.8
11	PT VGG16 Fine Tuning 1	73.2
12	PT ResNet50 Base	75.4
13	PT ResNet50 Fine Tuning 1	79.6
14	PT ResNet50 Fine Tuning 2	78.8
15	PT ResNet50 Fine Tuning 3	83.4

*underfitted.

Best Model Evaluation

Criteria:

• Not overfit too quickly
• Have good test accuracy of above 70%
• Be able to classify MOST of the images I downloaded from the Internet

Best Model: Model 9 (see report for scoring)