⏯ Kaggle - Structural Defects Network (SDNET) 2018

" Train a neural network to automatically detect structural defects with 96% Val_accu! "

_{Created by @Shuhui}

---

Introduction

• An exploration towards Structural Defects Network (SDNET) 2018
• Utilize data augmentation with geometric transformations using Torchvision in Python
• Train the data with Resnet34 and achieved a classification of cracked/non-cracked structures with an accuracy of 96% after parameter optimization

Requirements

• Python >= 3.7.0
• CUDA >= 11
• Pytorch >= 1.7.0

Environment

• video card: NVIDA Tesla K80 × 2
• Video Memory: 24G
• Memory: 16G

Usage

To use time-frequency module, place the contents of this folder in your PYTHONPATH environment variable.
To detect single image, use predict.py and change the img_path：

  if __name__ == '__main__':
    img_path = "../test_pic.jpg"
    main(img_path)

To detect a batch image, use batch_predict.py and change the img_path_list：

if __name__ == '__main__':
    img_path_list = ["../tulip.jpg", "../rose.jpg"]
    main(img_path_list)

Then get classified image with certainty:

Transfer Learning

Thanks to the model provided by Pytorch.