Fine-Grained Thai Food Image Classification

Project description

Our project involves implementing a network and training it on a dataset of Thai food images. We will then compare its performance with that of three pre-trained models in PyTorch: DenseNet, ResNet50, and Vision Transformer (base-sized model). We will modify the pre-trained models to work with our dataset and visualize some of their layers to understand how they classify Thai dishes.

Team

• Thean Cheat Lim
• Wenlin Fang

Links to project materials:

• Dataset
• Model
• Presentation Video
• Paper

Required Packages

• datasets
• transformers
• evaluate

Python files

• image_preprocessing.py: run the main function to calculate the mean and the standard deviations of the training images
• custom.py: run the main function to train the model and see the accuracies on the validation and test dataset
• resnet_densenet.py:
  • Fine-tune ResNet and DenseNet models
  • Usage: python resnet_densenet.py resnet50 model_finetuned_outname 20 /train_data_dir /val_data_dir /test_data_dir
  • Usage: python resnet_densenet.py densenet161 model_finetuned_outname 20 /train_data_dir /val_data_dir /test_data_dir
• swinv2.py:
  • Fine-Tune the SwinV2 (tiny) model using the THFOOD-50 dataset.
  • Usage: python swinv2.py 20 /output_dir 20 means training for 20 epoch
• visualization.py: run the main function to visualize the attention maps of the four models
• pytorch_model_utils.py: Utility functions for creating/training/testing Pytorch models.