Benchmarking ConvNeXt models on FGVC datasets

This repository adds data-loader supports for 3 fine-grained visual classification datasets for training ConvNeXt. You can train ConvNeXt using:

• CUB dataset
• FoodX-251 dataset
• CUB + Stanford Dogs dataset (a custom dataset containing categories from both CUB and Dogs datasets providing 320 classes)

This repo was mainly designed to make a strong ConvNeXt teacher model trained on FGVC dataset and use it to distill knowledge into DeiT model. Please refer here to our DeiT Knowledge distillation repository for complete project details.

Requirements and Installation

We have tested this code on Ubuntu 20.04 LTS with Python 3.8.

Please check INSTALL.md for installation instructions.

Datasets

CUB Dataset

Download CUB dataset from here. Extract the file. Our code expects the dataset folder to have the following structure:

CUB_dataset_root_folder/
    └─ images
    └─ image_class_labels.txt
    └─ train_test_split.txt
    └─ ....

FoodX Dataset

Download FoodX dataset from here. After extracting the files, the root folder should have following structure:

FoodX_dataset_root_folder/
    └─ annot
        ├─ class_list.txt
        ├─ train_info.csv
        ├─ val_info.csv
    └─ train_set
        ├─ train_039992.jpg
        ├─ ....
    └─ val_set
        ├─ val_005206.jpg
        ├─ ....

Stanford Dogs Dataset

Download the dataset from here. The root folder should have following structure:

dog_dataset_root_folder/
    └─ Images
        ├─ n02092339-Weimaraner
            ├─ n02092339_107.jpg
            ├─ ....
        ├─ n02101388-Brittany_spaniel
            ├─ ....
        ├─ ....
    └─ splits
        ├─ file_list.mat
        ├─ test_list.mat
        ├─ train_list.mat

Training

To train a ConvNeXt-base (384x384) model on CUB dataset, run the following command

python main.py --model convnext_base --data-set CUB --drop_path 0.8 --input_size 384 --batch_size 16 --lr 5e-5 --update_freq 2 --warmup_epochs 0 --epochs 60 --weight_decay 1e-8  --layer_decay 0.7 --head_init_scale 0.001 --cutmix 0 --mixup 0 --output_dir /path/to/save/checkpoints --finetune /path/to/convnext/pretrained/imagenet/weights.pth --data-path /path/to/dataset

To train a ConvNeXt-base (384x384) model on CUB+Dog dataset, run the following command

python main.py --model convnext_base --data-set CUB_DOG --drop_path 0.8 --input_size 384 --batch_size 16 --lr 5e-5 --update_freq 2 --warmup_epochs 0 --epochs 60 --weight_decay 1e-8  --layer_decay 0.7 --head_init_scale 0.001 --cutmix 0 --mixup 0 --output_dir /path/to/save/checkpoints --finetune /path/to/convnext/pretrained/imagenet/weights.pth --data-path /path/to/CUB/and/Dog/dataset/seperated/by/space

Note: For CUB + DOG dataset, please provide both paths in the --data-set parameter, seperated by a space.

For example: /l/users/u21010225/AssignmentNo1/CUB/CUB_200_2011/ /l/users/u21010225/AssignmentNo1/dog/

Similarly to train ConvNeXt on FoodX dataset, replace CUB by FOOD in the --data-set argument and corresponding dataset path in above sample commands.

Evaluation

To evaluate a ConvNeXt trained model on CUB + DOG dataset, run the following command

python main.py --model convnext_base --data-set CUB_DOG --eval true --resume /path/to/trained/model --input_size 384 --drop_path 0.2 --data_path /path/to/CUB/and/Dog/dataset/seperated/by/space

---

Acknowledgement: This code repo is forked and modified from official ConvNeXt repository