https://github.com/khan-yin/kaggle-cassava-leaf-disease-classification

NoteBook for the competition "Cassava Leaf Disease Classification" held in Kaggle. Private Leaderboard 369/3900(Top10%) Bronze Medal
https://github.com/khan-yin/kaggle-cassava-leaf-disease-classification

computer-vision jupyter-notebook kaggle pytorch

Last synced: 6 days ago
JSON representation

NoteBook for the competition "Cassava Leaf Disease Classification" held in Kaggle. Private Leaderboard 369/3900(Top10%) Bronze Medal

• Host: GitHub
• URL: https://github.com/khan-yin/kaggle-cassava-leaf-disease-classification
• Owner: khan-yin
• Created: 2021-04-27T07:09:03.000Z (almost 4 years ago)
• Default Branch: main
• Last Pushed: 2021-05-24T07:38:55.000Z (over 3 years ago)
• Last Synced: 2024-12-03T16:46:18.936Z (2 months ago)
• Topics: computer-vision, jupyter-notebook, kaggle, pytorch
• Language: Jupyter Notebook
• Homepage: https://www.kaggle.com/c/cassava-leaf-disease-classification
• Size: 186 KB
• Stars: 4
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# kaggle-Cassava-Leaf-Disease-Classification

NoteBook for the competition "Cassava Leaf Disease Classification" held in Kaggle. Private Leaderboard 369/3900(Top10%) Bronze Medal.    

The Competiton was held from November 20th in 2020 to February 18th in 2021.It is the first time for me to join the competition in Kaggle. Fortunately, I won the bronze medal in this competition.🤣

# solution

My final submission for the competition got 0.8980 on private datasets,which got 0.9039 on public datasets. but my best solution got 0.8987 on private datasets.`new_predict.ipynb` which would be up to Top7% on leaderboard.the code released on GitHub contains training part and testing part. 

Here are some trciks I used on my notebook:

- Data Augmentation,I tried some new methods such as mixup, cutmix(but it didn't work well) while using the traditional methods in `torchvision.transforms` and `albumentations`.

- K-fold cross-validation `k=5`.

- Model Ensemble: EfficientB4 + Vision Transformer.

- use Focal Loss(Kaiming He's work) to ease class imbalance problem.

- Test Time Augmentation: I used `TTA=10` for every batch and calculate the mean probability for the result.

- AMP in pytorch for faster training.

# Hope

In the competition, I learned a lot of methods for better performance and standard code style from other kaggler's wonderful ideas all over the world.I will keep on reading and learning new technique. Hope I can get better score next time!😃