Best single model (average of 5 folds on the test set) achieves 0.40327 RMSE on Private LB. It was ResNet101

Thank you

It depends on the use-case you have. If you're interested in fast real-time predictions, then, of course, large ensembles is not a way to go. However, if you don't have strict time limits, but interested in the highest quality possible, then ensembles could help.

yeah you're right in deployment I don't think we'll want to have 3 models having to make predictions first before inference can me made as that will increase latency.

Thank you

Yes, that's right, but Data tab told that: `All the test images have reliable labels`. It means that it's better to train on good quality labels to obtain a better performance on the test dataset.

thank you so much

Hi @eaedk! It really depends on the problem type.

But usually I begin with the exploratory data analysis part followed by hypothesis generation, establishing local validation, and trying different ideas and approaches. I create a document where I store all these thoughts, assumptions, papers, and resources that, I think, could work for this particular problem. Sometimes such a document can even contain up to 20 pages of text. Then I have a list of priorities and try one idea after another with the higher priority (and save all the results).

I guess, it's just a natural idea. We have 5 classes in the good labels: 2,3,4,5,7. Classification gives us the softmaxed probabilites for each class. So, we could take a weighted average of the labels to obtain the final prediction.