If we use only machine learning such as Light GBM model on the image pixel values read as array, we can reach <1 RMSE values. While this implementation is inferior to deep learning models, it could have some usage in resource constrained environment where availability of GPU is an issue.
Sample code here:
https://github.com/anindabitm/CGIAR-Zindi/blob/master/Tree_model.ipynb
Thank you, I am wondering about the training time, since you're not using any gpu and the dimensiality is so high?
Training time is not too high about 20-30 mins for entire pipeline
thanks that good
how you came up with this (7839, 3136) (7839, 3136) (7839, 3136)
7839 is just the size of low quality images. And 3136 is 56*56
This is pretty interesting.
Good work! Nice to see different approaches used :) Thanks for sharing
Thanks
I tried it, with multiples variation, and I noticed that it fits the training data very quikly, but overfit so hard, I tried diffrent methods, and they don't seem to work good, but the technique defentley can get less than 1 rmse
Try attempting to use extracted image embeddings (new to Zindi and this piqued my interest, an example being the following https://www.kaggle.com/s/2543927)