ICLR Workshop Challenge #1: CGIAR Computer Vision for Crop Disease
$5,000 USD
Identify wheat rust in images from Ethiopia and Tanzania, and win a trip to present your work at ICLR 2020 in Addis Ababa.
820 data scientists enrolled, 306 on the leaderboard
29 January—29 March
Increasing the base line model to around 0.24 and 0.31
published 23 Mar 2020, 11:19

In the likes of competitiveness, it will be good to give more insight into how to improve the model for those below 0.4 scores on the leaderboard. The major challenges is you need to iterate fast and try different methods, and if you are you are using colab, then you and I know that colab is slow.

https://github.com/steveoni/ICLR_models contain some models I used to experiment on some basic stuff.

And to see more tricks that you can try before using ensembles check out this post : https://medium.com/@steveoni/tricks-for-improving-your-image-classification-model-cd1f588602ba

Following this trick, I was able to have a single model that gave me my highest score of 0.24 on the leaderboard, using Mixup data augmentation.

And to see how to use zindi dataset directly on colab, check out the discussion forum for the topic on that.

You are a champion. Thanks

thanks for this, but it's think this is coming late 5 days to goto share such score might not be really that fair. please always make it quicker on next competition

anyone who found it's notebook useful should give credit by star the notebook

See https://zindi.africa/competitions/iclr-workshop-challenge-1-cgiar-computer-vision-for-crop-disease/discussions/728

Likely your loss figures are impacted by duplicates in the train set that you have in your valid set.

yeah, i wasn't sure of the best way, i wanted to use scikit learn version before.

Thanks. tho am not submitting again

also set the random seed so you get the same examples in validation set across experiments

Please if some of these tricks works, Please do well to let me know, especially that of using Mixup for data augmentation, and is Test Time Augmentation, actually increasing the score

Thank you for sharing.However i think sharing such a huge score towards the end of a competition is evil to the community.Make them quicker in future competitions.cheers!