Unable to train the model

[a14en9]

Hi Kris,

I downloaded the dataset from HKH Glacier Mapping to try to train the model. If I understand correctly, I can call train.py under the scripts folder to train the downloaded data. However, I encountered the following problems:
Given groups=1, weight of size [16, 11, 3, 3], expected input[16, 15, 512, 512] to have 11 channels, but got 15 channels instead
I tried to go back and look directly at the image in the npy file and found their dimensions are in [512,512,15]. From my understanding, I might forget the normalisation process. But I am not very clear about this part, can you please elaborate, especially about the stats_train.json file in postprocess.yaml.

Best wishes,
Oliver

[Aryal007]

Hello Oliver,

The configuration for U-Net based model is specified in the file glacier_mapping/conf/train.yaml
Given that the downloaded images have 15 channels, I would suggest you change inchannels: 11 on line 4 to inchannels: 15. The given configuration file is expecting input image to have just 11 channels instead of the downloaded 15.

The stats_train.json file contains the mean and standard deviation across all channels and is calculated using generate_stats function in glacier_mapping/data/process_slices_funs.py. These pre-calculated stats are later used to normalize the training samples during training.

Please let me know if you are still experiencing the issue.

Regards,
Bibek

[a14en9]

Hi Bibek,
Thank you for your explanations.
Oliver