Contributers: Ang Li, Fenghao Yang, Yikai Mao, Jinghao Miao

First run get_data.sh to get Food 101 dataset

To run the code, run the main.py file and there are arguments that user can input for hyperparmeters, which includes:

• device_id (The id of the gpu to use; type=int; default=0)
• model (Model being used including baseline, custom, resnet18, and vgg16; type=str; default='custom')
• pt_ft (Whether model is for partial fine-tune model; type=int; default=1)
• bz (Batch size; type=int; default=32)
• shuffle_data (Whether shuffle the data; type=int; default=1)
• normalization_mean (Mean value of z-scoring normalization for each channel in image; type=tuple; default=(0.485, 0.456, 0.406))
• normalization_std (Standard deviation of z-scoring normalization for each channel in image; type=tuple; default=(0.229, 0.224, 0.225))
• epoch (Number of epoch; type=int; default=30)
• criterion (Which loss function to use; type=str; default='cross_entropy')
• optimizer (Which optimizer to use; type=str; default='adam')
• lr (Learning rate; type=float; default=1e-4)
• weight_decay (weight decay; type=float; default=1e-4)
• lr_scheduling (Whether enable learning rate scheduling; type=int; default=0)
• lr_scheduler (Learning rate scheduler; type=str; default='steplr')
• step_size (Period of learning rate decay; type=int; default=7)
• gamma (Multiplicative factor of learning rate decay; type=float; default=0.1)

Directly run the code with

python3 main.py

will train the and test the performance of baseline model. The results will be saved in results.pkl after training is finished, and can be used by visualization.ipynb to visualize the loss/accuracy graph.

• main.py: file for the entire code
• prepare_data.py: file to load the dataset
• data.py: file to pre-processe data, split train, validation and test set, create dataloader
• model.py: file with implementation of baseline, custom, resnet-18, vgg-16 model
• engine.py: file to prepare, train, test model, and save the results to results.pkl
• visualization.ipynb: notebook to plot graphs, visualize weight maps and feature maps