Dimitris Mantas

### Summary Currently, the `SemanticSegmentationTask` recognises the `ignore_index` parameter only when cross entropy and focal loss is used. However, `smp.losses.JaccardLoss` implicitly supports this option via its `classes` parameter, which is...

Add Inference Example

1

### Issue I think the documentation would greatly benefit a complete example on how to properly load a trained checkpoint, test, and finally use it for inference on a single...

We should throw specific exceptions when possible as per the [PEP8](https://peps.python.org/pep-0008/#programming-recommendations) guidelines which state that ```python try: ... except Exception: ... ``` blocks are a code smell.

Replace ``encoder_weights: str | None = "imagenet"`` with ``pretrained: bool=True``

2

The ``encoder_weights`` parameter in the model initializers is a bit ambiguous/strange, especially given its lax type hint. At least to me, it looked like you could pass strings to represent...

I noticed that the DeepLab ASPP has a hardcoded 50% dropout layer at the exit. I'm not sure whether this is mentioned in the original papers, but I think it...

In my opinion, the most attractive use case for certain timm encoders such as ResNet-18, which is also available in torchvision, is that timm generally allows for various additional configuration...