Data

Tools

Indexes

Applications

Experiments

Open Source Science

https://github.com/cedrickchee/tgs-saltnet

Kaggle | 22nd place solution for TGS Salt Identification Challenge

https://github.com/cedrickchee/tgs-saltnet

Science Score: 10.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • ○
    CITATION.cff file
  • ○
    codemeta.json file
  • ○
    .zenodo.json file
  • ○
    DOI references
  • ✓
    Academic publication links
    Links to: arxiv.org
  • ○
    Academic email domains
  • ○
    Institutional organization owner
  • ○
    JOSS paper metadata
  • ○
    Scientific vocabulary similarity
    Low similarity (7.2%) to scientific vocabulary

Keywords

computer-vision deep-learning fastai image-segmentation kaggle-competition pytorch
Last synced: 5 months ago · JSON representation

Repository

Kaggle | 22nd place solution for TGS Salt Identification Challenge

Basic Info
  • Host: GitHub
  • Owner: cedrickchee
  • Language: Jupyter Notebook
  • Default Branch: master
  • Size: 34.2 KB
Statistics
  • Stars: 1
  • Watchers: 4
  • Forks: 1
  • Open Issues: 0
  • Releases: 0
Fork of svishnu88/TGS-SaltNet
Topics
computer-vision deep-learning fastai image-segmentation kaggle-competition pytorch
Created over 7 years ago · Last pushed over 7 years ago

https://github.com/cedrickchee/TGS-SaltNet/blob/master/ 

# TGS-SaltNet

Kaggle competition22nd place solution for [TGS Salt Identification Challenge](https://www.kaggle.com/c/tgs-salt-identification-challenge).

## General

I recently participated in a Kaggle competition, TGS Salt Identification Challenge and reached the 22nd place. This repository contains the final code which resulted in the best model. The code demonstrates usage of different important techniques using [fast.ai](http://www.fast.ai/) and [PyTorch](https://pytorch.org/).

1. Use ResNet model as an encoder for U-Net.
2. Add intermediate layers like [Bottleneck Attention Module(BAM)](http://bmvc2018.org/contents/papers/0092.pdf), [Squeeze & Excitation](https://arxiv.org/abs/1803.02579) blocks in a ResNet34 model which can be easily replicated for other network architectures.
3. Show how to add [Deep supervision](https://www.kaggle.com/c/tgs-salt-identification-challenge/discussion/65933) to the network, and calculate loss and combine loss at different scale. 

## Main software used

1. fastai version 0.7
2. PyTorch version 0.4
3. Python version 3.6

## Hardware required

The code was tested with TitanX GPU / 1080ti.

## Thanks

A special thanks to Heng CherKeng for his generous contributions to different ideas in the competition, for a long list of amazing Kaglle community members, Jeremy and fast.ai community for the amazing and flexible fastai framework.

Owner

  • Name: Cedric Chee
  • Login: cedrickchee
  • Kind: user
  • Location: PID 1
  • Company: InvictusByte

Lead Software Engineer | LLMs | full stack Go/JS dev, backend | product dev @ startups | 🧑‍🎓 CompSci | alumni: fast.ai, Antler.co

GitHub Events

Total
Last Year