Refactor CropModel to infer num_classes from checkpoint during loading

[naxatra2]

Fixes #960

This PR makes num_classes optional in the CropModel class and infers it from the checkpoint when loading via load_from_checkpoint. This improves usability, especially for users who do not know or remember the original number of classes at the time of inference.

Changes Made

1. Made num_classes an optional argument in the CropModel constructor.
2. In the constructor, if num_classes is not provided (None), we defer model/metrics initialization until on_load_checkpoint. Otherwise, we initialize them immediately.
3. Added an on_load_checkpoint method that reads num_classes from the checkpoint if it is still None, then sets up the model and metrics.
4. Added tests to verify that:
   • We can load from a checkpoint without passing num_classes.
   • Users can still pass num_classes explicitly if they want.
   • Calling forward without having a valid model raises a clear error.

[bw4sz]

This is important and should be merged, two additions. We need to use

self.save_hyperparameters()

at the end of init documented here https://lightning.ai/docs/pytorch/1.6.2/common/hyperparameters.html#lightningmodule-hyperparameters

The second is the label dict, the crop model uses torchvision dataloader that loads folders from file, that could easily not be the same as the label_dict that as user could optionally pass. We should remove label_dict as an argument from init. Then we should add a self.label_dict = <> here

DeepForest/src/deepforest/model.py

Line 157 in f1c724d

I believe its under self.train_ds. class_to_idx.

https://pytorch.org/vision/main/generated/torchvision.datasets.DatasetFolder.html#torchvision.datasets.DatasetFolder

Now, the question I have is if a use creates a model, saves the hyperparameters, loads a dataset, saves a checkpoint, and then reloads the model, is the label_dict still there?

[bw4sz]

Can we also fix a small bug related to the crop label.

This is poor logic, not sure what i was thinking, we should never have hidden conditional formatting. The cropmodel should return the same dtype and it should insist on a label dict, this explains the issue #705.

DeepForest/src/deepforest/predict.py

Line 237 in f1c724d

results["cropmodel_label"] = crop_model.numeric_to_label_dict[label]

There should not be a if statement here. CropModels should have label dicts and they should return labels, not numeric.

The function seems to expect batch_size 1, that is wrong, it should be

results["cropmodel_label"] = [crop_model.numeric_to_label_dict[x] for x in label]

to work for larger batch sizes. Let's confirm this with tests. @naxatra2 if this is too much, I can take over, this PR is important.

[naxatra2]

If it is not urgent then can I try doing this @bw4sz, it will be a great learning opportunity. I will update my progress here.