Huan Ren, Han Jiang, Wang Luo, Meng Meng, Tianzhu Zhang (USTC)

This repository contains the code for the FungiCLEF2023 competition from UstcAIGroup. The majority of the code in this repository is sourced from fgvc9_fungiclef. The main differences lie in the custom_loss.py and post_avg_entropy.py files.

• In the custom_loss.py file, we have implemented the poisonous/edible classification loss for enhanced identification of poisonous species. Additionally, we have included a uniform distribution constraint specifically for the novel category in the validation set.

• In the post_avg_entropy.py file, we have implemented the Entropy-guided Unknown Identifier to leverage entropy for distinguishing novel categories.

You can get started by following these steps:

1. Create a new conda environment and activate the new environment:

conda create -n MetaFormer python=3.8
conda activate MetaFormer

2. Install PyTorch:

conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

3. Install additional required packages using pip:

pip install -r requirements.txt

4. Install Apex:

git clone https://github.com/NVIDIA/apex
cd apex
# if pip >= 23.1 (ref: https://pip.pypa.io/en/stable/news/#v23-1) which supports multiple `--config-settings` with the same key... 
pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./
# otherwise
pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --global-option="--cpp_ext" --global-option="--cuda_ext" ./

If you encounter the error No module named 'packaging', you can refer to this issue. One way to fix it is by running conda install packaging beforehand.

1. Download the challenge image data (we use the full size version) and metadata from competition website.

2. Download the CSV file indicating whether each category is poisonous.

3. Download the pretrained model from the model zoo of MetaFormer.

4. Place the datasets inside datasets/fungi/challenge_data/ and pretrained model into pretrained_model/. Make sure the data structure is as below.

├── datasets
│   └── fungi
│       └── challenge_data
│           ├── DF20
│           │   ├── 2237851949-74654.JPG
│           │   └── 2237851951-222637.JPG
│           ├── DF21
│           ├── ├── 0-3008822340.JPG
│           ├── └── 0-3008822343.JPG
│           ├── poison_status_list.csv
│           ├── test.csv
│           ├── train.csv
│           └── val.csv
├── pretrained_model
│   ├── metafg_0_inat21_384.pth
│   └── metafg_2_inat21_384.pth

bash run_train.sh

bash run_inference.sh

After running inference, you will get result{0-rank}.pkl which indicate the output of a single model. Here we give an example:

├── fungi_pkl_ensemble
│   ├── MetaFG_meta_0_384_bs36_epoch80_poison_trainval
│   │   ├── result0.pkl
│   │   ├── result1.pkl
│   │   ├── result2.pkl
│   │   └── result3.pkl
│   ├── MetaFG_meta_2_384_bs18_epoch64_poison_trainval
│   │   ├── result0.pkl
│   │   ├── result1.pkl
│   │   ├── result2.pkl
│   └── └── result3.pkl

You can average ensemble the results and post process with our proposed Entropy-guided Unknown Identifier:

python post_avg_entropy.py

We referenced the repos below for the code.

• fgvc9_fungiclef

• MetaFormer