MSE Loss function with SemanticSegmentationTask

[ghost]

Is there a way to run SemanticSegmentationTask with RMSE or MSE loss with torchgeo? Currently, supports 'ce', 'jaccard' or 'focal' loss.

[isaaccorley]

We have torchgeo.trainers.PixelwiseRegressionTask specifically for this. Let me know if you encounter any issues.

[adamjstewart]

We should think about ways to more easily integrate all metrics in case someone wants a different classification metric. Currently the only way to do this is something like:

trainer = SemanticSegmentationTask(...)
trainer.train_metrics = MetricCollection(...)
trainer.val_metrics = ...
trainer.test_metrics = ...

[ghost]

Thanks very much for your reply. torchgeo.trainers.PixelwiseRegressionTask seems what I need for my project. But is not available for me. I am using torchgeo version 0.3.1. Also, is there a notebook that shows end2end implementation ?

[adamjstewart]

But is not available for me. I am using torchgeo version 0.3.1.

PixelwiseRegressionTask will be added in the 0.5.0 release (due sometime next month if all goes according to plan). Until then, you can install a development version. See https://torchgeo.readthedocs.io/en/stable/user/installation.html for instructions to do that.

is there a notebook that shows end2end implementation ?

See #575