We'd like to introduce Yauhen Babakhin, Senior Data Scientist at H2O.ai, and winner of the CGIAR Wheat Growth Stage Challenge. Read on to see how he came out on top and learn from his solutions.
Please introduce yourself to the Zindi community.
Hi! My name is Yauhen Babakhin (bes). I orignally hail from Belarus, and now live in the Czech Republic.
Tell us about your background as a data scientist.
I have a Master’s Degree in Applied Data Analysis, and have over 5 years of working experience in Data Science. I worked in Banking, Gaming and eCommerce domains, and am currently a Senior Data Scientist at H2O.ai. I'm also the first Kaggle competitions Grandmaster in Belarus, having gold medals in both classic Machine Learning and Deep Learning competitions.
Tell us about your solution for the CGIAR Wheat Growth Stage Challenge.
My solution is based on an ensemble of multiple CNN models (ResNet50, ResNet101, ResNeXt50), together with some tricky augmentation techniques. It also uses a multi-stage training process utilising the bad quality labels that were available in the competition dataset.
I've published the code and a detailed solution here: https://github.com/ybabakhin/zindi_wheat_growth
I think what set my solution apart from the rest is that I realised it was important to utilise bad quality labels in some way (I used them as a pre-training step). And also a list of augmentations that reflected the peculiarities of the data.