[BUG] Dictionary attribute error when using the new sf.mil.get_mil_tile_predictions function

[sz3029]

Description

Thank you for the recent update of v2.3. I was trying to use the new function sf.mil.get_mil_tile_predictions() for CLAM model, but had encountered an error

To Reproduce

The code I used was:

...
clam_config = mil_config(
        model = 'clam_sb',
        trainer = 'fastai', # default, preferred
        lr=1e-4,
        epochs = 100,
        bag_size = 512,
        batch_size = 64,
        model_size = "small" # default to small
    )

sf.mil.get_mil_tile_predictions(
    weights=os.path.join(
            "/projects/slideflow_ML/CLAM", 
            "CLAM_D1_training",
            backbone_model,
            magnification,
            'mil',
            "00000-"+
            "_".join(
                ["CLAM_D1_training",
                magnification,
                backbone_model,
                normalization_type,
                ]
            ),
        ),
    outcomes="status",
    config=clam_config,
    dataset=dataset_test,
    bags=os.path.join(
            "/home/CLAM/TEST_HOLDOUT",
            magnification,
            "clamfeatures",
            backbone_model,
            normalization_type,
        ),
)

But it produced the following error:

File ~/.conda/envs/slideflow2-2/lib/python3.9/site-packages/slideflow/mil/eval.py:531, in get_mil_tile_predictions(weights, dataset, bags, config, outcomes, dest, uq)
    527 for i, (bag, slide) in track(enumerate(zip(bags, slides)),
    528                         description="Generating tile predictions",
    529                         total=len(bags)):
    530     if is_clam:
--> 531         pred_out = _predict_mil_tiles(model, bag, use_first_out=True, uq=uq)
    532     else:
    533         pred_out = _predict_mil_tiles(
    534             model,
    535             bag,
   (...)
    538             uq=uq
    539         )

File ~/.conda/envs/slideflow2-2/lib/python3.9/site-packages/slideflow/mil/eval.py:1130, in _predict_mil_tiles(model, bag, use_lens, device, apply_softmax, use_first_out, attention, attention_pooling, uq)
   1128     y_pred = y_pred.cpu().numpy()
   1129 if y_att is not None:
-> 1130     y_att = y_att.cpu().numpy()
   1131 if uncertainty is not None:
   1132     uncertainty = uncertainty.cpu().numpy()

AttributeError: 'dict' object has no attribute 'cpu'

It looks like the returned y_att is not a tensor object.

Expected behavior

Output: pandas dataframe of tile-level predictions

Environment (from the output of sf.about()):

• Slideflow Version: 2.3.1
• OS (e.g., Ubuntu): linux
• How you installed Slideflow (pip, source): pip
• Python version: 3.9.13
• Backend: cucim, torch
• CUDA version: 11.3
• GPU models and configuration: Tesla V100-SXM2-32GB