Normalization code request
dear all ,
i have done the same experiment on different dataset but surprisingly the experiment with normalization is much worse than without it since the normalization is important to be used in foundation models i was wondering why there is a problem in normalization case. i go through your files but i couldn't find the normalization code to see where i made a mistake so if possible would you please share your code or at least tell me how you have computed.
what i have done is to consider each patch or satellite chip band by band and compute the mean and standard deviation in your case you do have 18 values in 3 time frames but i have used just one time frame so i do have 6 values for mean and 6 values for standard deviation.
i am looking forward to your response.
bests,
Hi
@Erfan2e
,
not sure if the person who build the model is still part of the team. But you can find the config for training Prithvi 2 on this dataset here: https://github.com/NASA-IMPACT/Prithvi-EO-2.0/blob/main/configs/multicrop.yaml
It is using the following data module in TerraTorch: https://github.com/terrastackai/terratorch/blob/main/terratorch/datamodules/multi_temporal_crop_classification.py
Regarding normalisation, it's applied in the self.aug which comes from torchgeo. Basically we have one mean/std value band (6, not 18. The data is resized to 3x6 before normalization).
Hope that helps!