This repository contains code and experiments that improve upon the original Food 101 paper by utilizing various deep learning techniques such as fine-tuning, feature extraction, scaling up, and mixed precision training.
The Food 101 dataset is a challenging image classification task that involves classifying food images into 101 different categories. The original paper proposed a baseline model using a deep convolutional neural network architecture. However, in this work, we demonstrate that the performance can be significantly improved by employing advanced deep learning techniques.
-
Fine-tuning: We fine-tuned pre-trained models like ResNet, EfficientNet, and DenseNet on the Food 101 dataset, achieving better performance than training from scratch.
-
Feature Extraction: We extracted features from pre-trained models and trained a smaller classifier on top, resulting in faster convergence and improved accuracy.
-
Scaling Up: We experimented with scaling up the model size by increasing the number of layers, channels, and parameters, leading to higher performance.
-
Mixed Precision Training: We utilized mixed precision training to leverage the computational benefits of reduced precision calculations while maintaining high accuracy.
Our best model achieves a top-1 accuracy of 91.1% on the Food 101 test set, surpassing the original paper's performance by a significant margin. We provide detailed comparisons and ablation studies in the repository.
To replicate our results or build upon our work, follow these steps:
- Clone the repository:
git clone https://github.com/your-username/food101-improvements.git - Install the required dependencies:
pip install -r requirements.txt - Download the Food 101 dataset and update the data paths in the configuration files.
- Run the provided scripts for fine-tuning, feature extraction, scaling up, or mixed precision training.
Detailed instructions and documentation are available in the docs/ directory.
We welcome contributions to this project! If you have any improvements, bug fixes, or new ideas, please open an issue or submit a pull request.
This project is licensed under the MIT License.
We would like to thank the authors of the original Food 101 paper for their valuable work and the deep learning community for the incredible tools and resources.