Trainers: add Instance Segmentation Task

[ariannasole23]

No description provided.

[adamjstewart]

Looks good, just need to make the code and tests match our existing trainers and run ruff: https://torchgeo.readthedocs.io/en/latest/user/contributing.html#linters

[adamjstewart]

To solve the import issue, you also need to add 2 lines to torchgeo/trainers/__init__.py, just copy the style used by other classes.

[adamjstewart]

Can you run ruff on the code to make it more uniform? This will make it easier to review.

[adamjstewart]

Yet another test that downloads data on the fly (#1088), but let's fix that another day when we figure out how to support custom weights.

[adamjstewart]

Hmm, looks like FTW isn't supported because it only has masks, not bboxes.

[adamjstewart]

Tests are finally passing! @ariannasole23 maybe you can check VHR-10 again to make sure I didn't break anything and it still works. Would like a couple other people to review this too, but otherwise this is ready to merge once you accept the CLA.

[ariannasole23]

@microsoft-github-policy-service agree

[adamjstewart]

Will give others a few days to test before merging.

[ashnair1]

The current InstanceSegmentationTask implementation isn't really flexible. I tried training a model with it recently and found the following:

• Training a model with weights=True i.e. with a pretrained backbone crashes as the FPN Weights trained on COCO require number of classes to be 91 whereas the VHR10 has 11 classes (incl background). Currently, the only way to train this model would be to have weights=None. See below for sample training config
• Using the Resnet50_FPN_Weights means that users can only use the resnet50 backbone. The ObjectDetectionTask allows users to choose their preferred backbone i.e. resnet18, resnet34, resnet50 etc.

Train config used:

trainer:
  min_epochs: 15
  max_epochs: 100

model:
  class_path: InstanceSegmentationTask
  init_args:
    model: 'mask-rcnn'
    backbone: 'resnet50'  # Only resnet50 can be used
    weights: None  # Using pretrained weights (weights:True) will not work in this implementation
    num_classes: 11
    lr: 3.0e-5 

I'd suggest following the implementation of how Faster-RCNN is defined in ObjectDetectionTask while defining Mask-RCNN in InstanceSegmentationTask.

[adamjstewart]

• weights=True, num_channels!=91 has been "fixed" and tested in the latest commit. I say "fixed" because I just skip the decoder weights when num_channels!=91. A smarter solution would be to drop the last layer of the weights, but I was lazy.
• Yep, this trainer currently lacks a lot of options, and only supports a single model and backbone. I agree that the implementation in ObjectDetectionTask is more flexible, but @ariannasole23 is busy with final exams and I'm busy preparing the release. I would suggest merging this as is and you can open a follow-up PR to add more backbones if you want.