Skip to content

[FEATURE] timm.models.adapt_input_conv: beyond RGB weights #2445

Open
@adamjstewart

Description

@adamjstewart

Is your feature request related to a problem? Please describe.

TorchGeo provides a number of model weights pre-trained on non-RGB imagery (e.g., Sentinel-2, 13 channels). Oftentimes, when dealing with time-series data, we would like to stack images along the channel dimension so that we end up with $$B \times TC \times H \times W$$ inputs. However, we don't yet have an easy way to adapt our pre-trained weights to match.

Describe the solution you'd like

timm.models.adapt_input_conv provides a powerful tool for repeating and scaling weights to adapt to changing in_chans, but only seems to support 3-channel weights if in_chans > 1. I would like to extend this to support any number of channels. Would this be as simple as replacing 3 with I throughout the function?

Describe alternatives you've considered

We could write our own functionality in TorchGeo, but figured this would be useful to the broader timm community.

Additional context

@isaaccorley @keves1 may also be interested in this.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions