WinCLIP few_shot_mode issue

[a-koe-753]

Hey guys :) I want to use WinCLIP for my custom dataset. In order to "calibrate the thresholds" I use engine.validate before prediction as described e.g. in (#2446 #2701) For zero-shot mode this works as expected. However I just cannot bring it to work for few-shot mode when adding custom reference images.

This is my code. When adding the few_shot_source I always run in a problem with the pre_processor. It seems that in this case the test_transform attribute is missing. Any ideas how I can solve this?
Any help would be greatly appreciated!

from anomalib.models import WinClip
from anomalib.engine import Engine

engine = Engine()
model = WinClip(class_name="my class", k_shot=5, few_shot_source="./my class/normal_resized")

torch.save(model.state_dict(), "winclip_probe-mark.pt")

from anomalib.data import Folder
from torchvision.transforms.v2 import Resize, InterpolationMode

resize_tf = Resize(size=(240,240), interpolation=InterpolationMode.BICUBIC)

datamodule = Folder(
    name="my class",
    root="./my class",
    normal_dir="normal",
    abnormal_dir="anomalous",    
    mask_dir="masks/anomalous", 
    augmentations=resize_tf,
    seed=42
)

validations = engine.validate(model=model, datamodule=datamodule)
print(validations)

---

AttributeError Traceback (most recent call last)
Cell In[15], line 1
----> 1 validations = engine.validate(model=model, datamodule=datamodule)
2 print(validations)

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/anomalib/engine/engine.py:460, in Engine.validate(self, model, dataloaders, ckpt_path, verbose, datamodule)
458 if model:
459 self._setup_trainer(model)
--> 460 return self.trainer.validate(model, dataloaders, ckpt_path, verbose, datamodule)

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/trainer.py:663, in Trainer.validate(self, model, dataloaders, ckpt_path, verbose, datamodule)
661 self.state.status = TrainerStatus.RUNNING
662 self.validating = True
--> 663 return call._call_and_handle_interrupt(
664 self, self._validate_impl, model, dataloaders, ckpt_path, verbose, datamodule
665 )

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/call.py:49, in _call_and_handle_interrupt(trainer, trainer_fn, *args, **kwargs)
47 if trainer.strategy.launcher is not None:
48 return trainer.strategy.launcher.launch(trainer_fn, *args, trainer=trainer, **kwargs)
---> 49 return trainer_fn(*args, **kwargs)
51 except _TunerExitException:
52 _call_teardown_hook(trainer)

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/trainer.py:705, in Trainer._validate_impl(self, model, dataloaders, ckpt_path, verbose, datamodule)
701 download_model_from_registry(ckpt_path, self)
702 ckpt_path = self._checkpoint_connector._select_ckpt_path(
703 self.state.fn, ckpt_path, model_provided=model_provided, model_connected=self.lightning_module is not None
704 )
--> 705 results = self._run(model, ckpt_path=ckpt_path)
706 # remove the tensors from the validation results
707 results = convert_tensors_to_scalars(results)

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/trainer.py:973, in Trainer._run(self, model, ckpt_path)
970 log.debug(f"{self.class.name}: preparing data")
971 self._data_connector.prepare_data()
--> 973 call._call_setup_hook(self) # allow user to set up LightningModule in accelerator environment
974 log.debug(f"{self.class.name}: configuring model")
975 call._call_configure_model(self)

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/call.py:110, in _call_setup_hook(trainer)
108 _call_lightning_datamodule_hook(trainer, "setup", stage=fn)
109 _call_callback_hooks(trainer, "setup", stage=fn)
--> 110 _call_lightning_module_hook(trainer, "setup", stage=fn)
112 trainer.strategy.barrier("post_setup")

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/lightning/pytorch/trainer/call.py:177, in _call_lightning_module_hook(trainer, hook_name, pl_module, *args, **kwargs)
174 pl_module._current_fx_name = hook_name
176 with trainer.profiler.profile(f"[LightningModule]{pl_module.class.name}.{hook_name}"):
--> 177 output = fn(*args, **kwargs)
179 # restore current_fx when nested context
180 pl_module._current_fx_name = prev_fx_name

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/anomalib/models/image/winclip/lightning_model.py:150, in WinClip.setup(failed resolving arguments)
146 if self.few_shot_source:
147 logger.info("Loading reference images from %s", self.few_shot_source)
148 reference_dataset = PredictDataset(
149 self.few_shot_source,
--> 150 transform=self.pre_processor.test_transform if self.pre_processor else None,
151 )
152 dataloader = DataLoader(reference_dataset, batch_size=1, shuffle=False, pin_memory=True)
153 else:

File ~/projects/WINCLIP_try2/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py:1964, in Module.getattr(self, name)
1962 if name in modules:
1963 return modules[name]
-> 1964 raise AttributeError(
1965 f"'{type(self).name}' object has no attribute '{name}'"
1966 )

AttributeError: 'PreProcessor' object has no attribute 'test_transform'