Vesuvius Challenge - Ink Detection (9th place solution)

https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection

For solution discussion, refer to https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/discussion/417361

1. Hardware

• GPU: 2x Nvidia Quadro RTX 8000, each with VRAM 48 GB
• CPU: Intel® Xeon(R) Gold 6240 CPU @ 2.60GHz, 72 cores
• Memory: 376 GB RAM

2. OS

• ubuntu 18.04.5 LTS

3. Set Up Environment

• Install Python >=3.8.10
• Install requirements.txt in the python environment
• Set up the directory structure as shown below.

└── solution
    ├── src 
    ├── results
    ├── data
    |   ├── vesuvius-challenge-ink-detection
    |   |   ├── test
    │   |   ├── train
    │   |   ├── sample_submission.csv
    │   |   
    |   ├── pretrained   
    |       ├── pvt_v2_b3.pth
    | 
    ├── r050_resnet34-unet-mean32-pool-05.sh   
    ├── r090-pvtv2-daformer-pool-02a.sh  
    ├── LICENSE 
    ├── README.md 

• The dataset "vesuvius-challenge-ink-detection" can be downloaded from Kaggle:
  https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/data

• Pretrained model can be download from PVT (Pyramid Vision Transformer) repository:
  https://github.com/whai362/PVT/releases/download/v2/pvt_v2_b3.pth

4. Training the model

Warning !!! training output will be overwritten to the "solution/results" folder

• Use the 2 commands to train the weights of the deep neural nets. The bash script will call "run_train.py" with appropriate configure file.

usage: python run_train.py <configure>

#pwd = /solution
>> bash ./r050_resnet34-unet-mean32-pool-05.sh
>> bash ./r090-pvtv2-daformer-pool-02a.sh

• This will produce the 4 model files used in submission:

└── solution
    ├── ...
    ├── results
        ├── r050_resnet34-unet-mean32-pool-05
        |   ├── fold-1 
        |   |   ├── stage2_0/checkpoint/00018924.model.pth
        |   |      
        |   ├── fold-2aa
        |       ├── stage2_0/checkpoint/00014850.model.pth
        |   
        ├── r090-pvtv2-daformer-pool-02a
            ├── fold-1 
            |   ├── stage2_0/checkpoint/00029376.model.pth
            |      
            ├── fold-2aa
                ├── stage2_0/checkpoint/00009159.model.pth

• optional: if you are interested in the parameters used in training the different stages of the model , you can refer to "hyper-parameters.pdf" in the repo.

5. Submission notebook

• The public submission notebook (clean version) is at:
  https://www.kaggle.com/code/hengck23/9th-place-final-ensemble-v1?scriptVersionId=136499959

• The same learned weights from this repository are used in the submission notebook via the public dataset https://www.kaggle.com/datasets/hengck23/ink-weight-05a

  [private / public score] 0.654354 / 0.738625

6. Local validation

• Run the "run_infer_ms.py" python sript for local validation.

usage: python run_infer_ms.py <configure> <model_file>

#pwd = /solution
>> python src/r050_resnet34-unet-mean32-pool-05/run_infer_ms.py config_fold2aa_stage2_0 results/r050_resnet34-unet-mean32-pool-05/fold-2aa/stage2_0/checkpoint/fold-2aa-Resnet34MeanPool-00014850.model.pth

7. Reference train and validation results

• Reference results can be found at the google share drive. It includes the weight files, train/validation logs and visualisation images. You can use this to check your results.

  [google drive link (1.8 GB)]

8. Presentation slides:

• Please refer to "kaggle-winner-call-version-july-05.pdf" in the repo.

Authors

• https://www.kaggle.com/hengck23

License

• This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgement

"We extend our thanks to HP for providing the Z8-G4 Data Science Workstation, which empowered our deep learning experiments. The high computational power and large GPU memory enabled us to design our models swiftly."