search

ecker-lab / sensorium_2023

repo for the sensorium 2023 competition
24 stars 13 forks source link

data scaling was nans for some np versions, this is more robust #13

Closed pollytur closed 4 months ago

pollytur commented 4 months ago

Before scaling was applied before cutting videos to chunks of n frames. As it includes interpolation and original videos have nans, some version of the library resulted in nans being interpolated everywhere. So now scaling is the last data augmentation before putting data to tensors.

pollytur commented 4 months ago

this does not influence benchmark results as we used scaling=1 anyway

yongrong-qiu commented 4 months ago

Fine from my side.