Building model for 4-band data?

[NoaMillsUSDA-ARS]

I'm trying to train rastervision to do semantic segmentation on 4 band imagery. I obviously can't use resnet50 for this task, since resnet50 is built for 3 band imagery. How can I create a custom neural network model to use in resnet50's place?

[AdeelH]

Raster Vision is quite capable of handling arbitrary number of bands. See the multiband argument is isprs_potsdam.py for an example.

To use a custom model architecture, see the external_model argument in the same file.

[NoaMillsUSDA-ARS]

Thanks for the info! I was able to get my model running through the train and bundle stages but I'm getting a runtime error at the prediction stage. I'll start a new discussion thread to follow up.

[NoaMillsUSDA-ARS]

Also, for documentation, I am running rastervision in a singularity container on a compute cluster, and in order to get the model to train, I had to chmod +w the files in pytorch-fpn/.git/objects/pack because I was getting permission denied errors. Would love to know what these files are for.

Further, I was getting another permission denied error at the bundle stage, and found that the model was trying to write to the /local/scratch folder, which is in the singularity container and by default read-only. I was able to fix this by bind mounting to the /local folder in the container.