README.md

Bag-of-Tricks-and-a-Strong-Baseline-for-Fungi-Fine-Grained-Classification

paper link: http://star.informatik.rwth-aachen.de/Publications/CEUR-WS/Vol-3180/paper-182.pdf

This work is sponsored by Natural Science Foundation of China(62276242), CAAI-Huawei MindSpore Open Fund(CAAIXSJLJJ-2021-016B), Anhui Province Key Research and Development Program(202104a05020007), and USTC Research Funds of the Double First-Class Initiative(YD2350002001)”。

1. Environment setting

1.0. Package

• Several important packages

  • mindspore == 1.8.1
  • mindcv

• Replace folder mindcv/ to our mindcv/ folder (We made some changes to the original Mindcv framework, such as adding TA, etc)

  MindCV is an open source toolbox for computer vision research and development based on MindSpore. mindcv

1.1. Dataset

In this project, we use a large fungi's datasets from this challenge to evaluate performance:

• Fungi2022

1.2. OS

• Windows10
• Ubuntu20.04
• macOS (CPU only)

2. Train

• Single GPU Training
• DataParallel (single machine multi-gpus)
• DistributedDataParallel

2.1. data

train data and test data structure:

├── DF20/
│   ├── img20001.jpg
│   ├── img20002.jpg
│   └── ....
├── DF21/
│   ├── img21001.jpg
│   ├── img21002.jpg
│   └── ....
└──

Training sets and test sets are distributed with CSV labels corresponding to them.

2.2. configuration

you can directly modify yaml file (in ./configs/)

2.3. run.

take the SwinTransformer training process as an example.

1. we train a basic model by dividing the training set and verification set 9:1

python train.py ./Fungidata/DF20 -c configs/swin_large_384.yaml \
        --freeze-layer 2 \
        --batch-size 32 \
        --lr 0.01 \
        --decay-rate 0.9 \
        --output ./output/Swin-TF/DF20/freeze_layer_2

2. Add data augment and continue fine-tuning

python train.py ./Fungidata/DF20 -c configs/swin_large_384.yaml \
        --output ./output/Swin-TF/DF20/All_aug/freeze_layer_2 \
        --initial-checkpoint ./output/Swin-TF/DF20/freeze_layer_2/20220430-123449-swin_large_patch4_window12_384-384/Best_Top1-ACC.pth.tar \
        --freeze-layer 2 \
        --lr 0.001 \
        --batch-size 32 \
        --warmup-epochs 0 \
        --cutmix 1 \
        --color-jitter 0.4 \
        --reprob 0.25 \
        --aa trivial \
        --decay-rate 0.9

3. Modify the loss function and continue fine-tuning

python train.py ./Fungidata/DF20 -c configs/swin_large_384.yaml \
        --output ./output/Swin-TF/DF20/new_loss/freeze_layer_2 \
        --initial-checkpoint ./output/Swin-TF/DF20/All_aug/freeze_layer_2/20220430-123449-swin_large_patch4_window12_384-384/Best_Top1-ACC.pth.tar \
        --freeze-layer 2 \
        --lr 0.001 \
        --batch-size 32 \
        --warmup-epochs 0 \
        --cutmix 1 \
        --color-jitter 0.4 \
        --reprob 0.25 \
        --aa trivial \
        --decay-rate 0.9 \
        --Focalloss

4. Fine-tuning with full data sets

python train_all.py ./Fungidata/DF20 -c configs/swin_large_384.yaml \
         --batch-size 32 \
         --img-size 384 \
         --output ./output/Swin-TF/DF20/All_data/swin_large_384 \
         --freeze-layer 2 \
         --initial-checkpoint ./output/Swin-TF/DF20/new_loss/freeze_layer_2/20220502-114033-swin_large_patch4_window12_384-384/Best_Top1-ACC.pth.tar \
         --lr 0.001 \
         --cutmix 1 \
         --color-jitter 0.4 \
         --reprob 0.25 \
         --aa trivial \
         --decay-rate 0.1 \
         --warmup-epochs 0 \
         --epochs 24 \
         --sched multistep \
         --checkpoint-hist 24 \
         --Focalloss

5. Two-stage training

python train_all.py ./Fungidata/DF20 -c configs/swin_large_384.yaml \
         --batch-size 32 \
         --img-size 384 \
         --output ./output/Swin-TF/DF20/two_stage/swin_large_384 \
         --freeze-layer 2 \
         --initial-checkpoint ./output/Swin-TF/DF20/All_data/freeze_layer_2/20220504-115867-swin_large_patch4_window12_384-384/Best_Top1-ACC.pth.tar \
         --lr 0.001 \
         --cutmix 1 \
         --color-jitter 0.4 \
         --reprob 0.25 \
         --aa trivial \
         --decay-rate 0.1 \
         --warmup-epochs 0 \
         --epochs 5 \
         --sched multistep \
         --Focalloss

2.4. multi-gpus

change train.sh

python -m mindspre.distributed.launch --nproc_per_node=4 train_all.py

3. Evaluation

for details, see test.sh

sh test.sh

Use the sh test.sh command to execute the reasoning file (the checkpoint in the google cloud disk will be used, please download it in 5. Challenge's final checkpoints and logits)

4. Model Ensemble

run model_ensemble.ipynb

Execute 3. Evaluation to get the logits file required by model_ensemble.ipynb (you can also download it directly from google cloud disk, see 5. Challenge's final checkpoints and logits for details)

5. Challenge's final checkpoints and logits

It can be downloaded from Google Cloud Disk: https://drive.google.com/file/d/1v9SlsjXXKI5kizg9BTMWfA9faaawAgpv/view?usp=sharing

It can be directly used for model ensemble reasoning.

Acknowledgment

• Thanks to mindcv for Mindspore implementation.