Logo for the website. A blue P-like symbol and a red K-like symbol. Pushkar Kurhekar 
Back to projects
•
3 min read

Food Image Classifier

Table of Contents

Food Image Classification with CNNs

Installation

The following packages are required:

  • PyTorch
  • Pandas
  • scikit-learn
  • matplotlib
  • tqdm
  • numpy
  • kaggle

For PyTorch, please follow the instructions here (for pip or conda): https://pytorch.org/get-started/locally/

pip install numpy pandas matplotlib scikit-learn tqdm kaggle torch torchvision torchaudio

Downloading the dataset

There are a few ways to obtain the Food-101 dataset. One is from the author’s website here. Another option is to download it from Kaggle here.

Alternatively, if you decide to run the Notebook, it already contains code to download the zip from either the dataset website or from kaggle. For the kaggle option via the Notebook, you will need a kaggle.json file in the same directory as the notebook. More info here.

Usage

The repository contains both Jupyter notebooks and Python scripts. The Jupyter notebooks have commands to execute any additional libraries that may be required for installation. Included is also the experiment code that was run for training the Food-11 dataset as well in food11-training.ipynb. All the notebooks can also be run in Google Colab as-is.

The python scripts have the same code as the notebooks, just more cleaned up and better organized. You will get the same results regardless of which option you use.

Training and Evaluation

This script will execute the entire training pipeline for all 5 models, both pretrained and non-pretrained. Two CSV files and pickle files will be saved at the end with the results on the test set and model metrics respectively. All the trained models will be stored in the Trained_Models folder in the same directory as the script.

python train_classifier.py

Inference

You will need to set the input image path in the main() function. The script will print out the predicted class.

python predict_image.py

Plotting

The jupyter notebook plots.ipynb contains the plotting code to generate plots for training and testing loss and accuracy for both pretrained and non-pretrained models. You will need to specify the path to the pickle files for the pretrained and non-pretrained model metrics in order to generate the figures.

You can download the pickle files obtained from the training experiment here: https://drive.google.com/file/d/1cp-0ytNgz7hh4NbRV1jCWYsJ46QIf9YE/view?usp=sharing