[Explainer] PyTorch starter 0.857 F1-Score on public LB

Seismic Facies Identification Challenge
#1

Baseline to take part in the challenge based on pytorch+argus+albumentations+BloodAxe’s toolbelt
Key component to achieve score more than 0.75 is a crop-predict (look at ImageSlicer).

Notebook on COLAB (train + inference)

12 Likes
📝🏆 Seismic Facies Id Community Prize Winners (Round 1)
📝 Explained by the Community | Win 4 x DJI Mavic Drones
#2

Great code! Very simple. Thanks.

#3

Your code looks really clean and nice. Can you please elaborate more on the process of inference? I am asking because I’ve come to the opposite conclusion: the fewer overlaps I do, the better the final score is :thinking:

#4

I tried different crop/overlap and only params like in notebook get score more than 0.7.
I have an intuition, thats because preprocessing lack.
My “postprocessing” must be achieved in learnable way, mb there is some possibilities to apply window function through Z axis (in learnable way it could be like in hypercolumn) or just use RMS/MPA preprocessing from @leocd post (thereby narrowing feature extraction process for network, extraction vs regularization)

1 Like