ehudbaumatz/ML_ShipDetection

Detectron2 and FastAI based ship segmentation for Kaggle dataset.

Instance segmentation framework for kaggle ship detection¹ using Pytorch based toolboxes Detectron2 and Fastai.

Final presentation ²

Setup

• Requirements:
  • Unix based OS (Linux/Mac OS).
  • Nvidia GPU with cuda 10.1 installed, and at least 10GB of vram.
  • Docker (version == 19.03) and docker-compose (version == 3.6)
  • 40GB of free space: ~10GB for docker image, ~30GB for data

*All other dependencies are taken care of by docker-compose.

Data

kaggle competitions download -c airbus-ship-detection

Use the Kaggle API to download the dataset ³

Put it in folder named input

Model preprocessing / training and Kaggle submittion

docker-compose up 

• Training: In the container, open the notebook module_notebook and run all.
• Metrics: tensorboard for metrics, use another jupiter notebook and !tensorboard --logdir=runs --host=0.0.0.0 and if run on a local machine use a browser to go to http://0.0.0.0:6006.

Quick overview:

• Data Loader (module_preprocessing) ⁴
• Classifier (FastAI) ⁵
  • Output probability of ship on image.
  • Resnet34
• Instance Segmentation/Object Detection (Detectron2) ⁶
  • Output pixel mask of ships.
  • Augmentations ⁷⁸
    • Flips (Vertical 50%, Horisontal 50%)
    • Rotation (-20/+20 random rotation)
    • Random lighting (0.1 standard deviations)
  • Mask RCNN 50 layer pretrained on Coco .
    • 3 stage training (256x256,512x512,756x756) ⁹
  • Validation every 5k iterations.
• Predictions on validation (module_submit)
• Kaggle submission (module_submit) ⁴

Project

• Kaggle proptotype
• Initial presentation
• Data Loader module
• Training module
• Add data Augmentations
• Submit module
• Classification module
• Dockerize project
• Argumentize code
• Train
• Inference
• Submit

Footnotes

1. https://www.kaggle.com/c/airbus-ship-detection/overview ↩

2. https://docs.google.com/presentation/d/1O3DLAQ9SKukEivKACH7JCZqiKiia6ratmZuc1KuovQk/edit?usp=sharing ↩

3. https://github.com/Kaggle/kaggle-api ↩

4. https://github.com/pascal1129/kaggle_airbus_ship_detection ↩ ↩²

5. https://www.fast.ai/ ↩

6. https://github.com/facebookresearch/detectron2 ↩

7. https://jss367.github.io/Data-Augmentation-with-Detectron2.html ↩

8. https://link.springer.com/chapter/10.1007/978-981-15-5558-9_9 ↩

9. https://www.kaggle.com/iafoss/unet34-dice-0-87/notebook ↩