bloodaxe / kaggle-2020-alaska2

🏔️ Alaska 2 Solution

This repository contains a prize winning solution code for ALASKA 2. Team ABBA McCandless.

Key elements

• Trust your CV
• Don't resize, think twice before using hard image augmentations
• Don't use standard image I/O libraries (avoid rounding and clipping pixel values to [0..255])
• Use CNNs without pooling layers in the stem
• Higher resolution for deeper layers is better
• Build a diverse ensemble
  • EfficientNet
  • MixNet
  • SRNet
  • Hand crafted features (DCTR/JRM)

Documentation

See ./documentation/ABBA_McCandless_documentation.pdf for the solution documentation. A short description is also available on the kaggle forum.

Installation

Eventually run:

bash system_requirements.sh

And install pip requirements:

pip install -r requirements.txt

Inferencing

For sake of convinience, we attach pre-trained models in models/ and abba/weights/, so you may use them right away:

export KAGGLE_2020_ALASKA2=/path/to/alaska2/dataset

sh abba_predict.sh
sh eugene_predict.sh

After running inferencing scripts, final submissions can be found in submits/ folder.

Training

export KAGGLE_2020_ALASKA2=/path/to/alaska2/dataset

# This will take couple of hours to extract DCT matrices from JPEG and save to disk
sh eugene_preprocess.sh

# This will train models from our ensemble. Requires 4-GPU machine and plenty of time
sh abba_train.sh
sh eugene_train.sh

Hardware requirements

Mostly trained on 4xTitan V100 and 3xTitan RTX.

kaggle-2020-alaska2's Issues

Ask about the torch and torchvision version

May I ask which version of torch and torchvision you are using? Thank you!!!

Support In understanding certain aspect of this work

Hi all , I am doing a research on image steganalysis I stumbled on this work while analysing ALASKA compeititon , really nice work
I have being having certain trouble in understanding bit of the work you have done here. Is it possible to contact you to understand these bits.

Thanks in advance