[Bug]: Lightning vs exported models have huge pred_score gap

[FedericoDeBona]

Describe the bug

Train a model and inference with engine.predict() gives the correct pred_score, then exporting models in torch, openvino and ONNX, all three give the same wrong pred_score (always higher).
I expect results of the exported models to be equal to the lightning one.

Train code

augmentations = transforms.Compose([transforms.Resize((256, 256))])
datamodule = Folder(
	name="test",
	root=f"/home/rdvisiofy/biiond-worker/local_dataset/64",
	normal_dir="good",
	abnormal_dir="defected",
	augmentations=augmentations
)
model = Patchcore()
engine = Engine()
engine.fit(datamodule=datamodule, model=model)
#Run inference
da = PredictDataset(path = "/home/rdvisiofy/biiond-worker/local_dataset/64/good/01JT3F31K0BWQ1DH2T2RBZBVV7.jpeg")
da = DataLoader(da, collate_fn=da.collate_fn)
pred = engine.predict(dataloaders=da)
#OUTPUT: pred_score 0.2846

Inference code

image = cv2.imread("/home/rdvisiofy/biiond-worker/local_dataset/64/good/01JT3F31K0BWQ1DH2T2RBZBVV7.jpeg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (256,256))

# OPENVINO
pred = OpenVINOInferencer(path="/home/rdvisiofy/biiond-worker/deleteme/results/Patchcore/test/latest/weights/openvino/model.bin", device="CPU") .predict(image)
#OUTPUT: pred_score 0.4146552

pred = TorchInferencer("/home/rdvisiofy/biiond-worker/deleteme/results/Patchcore/test/latest/weights/torch/model.pt").predict(image)
#OUTPUT:pred_score 0.4146552

# ONNX
image = image.astype(np.float32)/255.0
image = np.transpose(image, (2, 0, 1))
image = np.expand_dims(image, axis=0)
ort_session = ort.InferenceSession("/home/rdvisiofy/biiond-worker/deleteme/results/Patchcore/test/latest/weights/onnx/model.onnx")
outputs = ort_session.run(None, {"input": image})
print("ONNX", outputs)
#OUTPUT: pred_score  0.41465414

Dataset

Other (please specify in the text field below)

Model

PatchCore

Steps to reproduce the behavior

see above

OS information

OS information:

• Python version: 3.10.0
• Anomalib version: 2.0.0, via pip
• PyTorch version: 2.6.0
• GPU models and configuration: NVIDIA GeForce RTX 4090
• Any other relevant information: I'm using a custom dataset

Expected behavior

I expect results of the exported models to be equal to the lightning one.

Screenshots

No response

Pip/GitHub

pip

What version/branch did you use?

2.0.0

Configuration YAML

.

Logs

.

Code of Conduct

• I agree to follow this project's Code of Conduct

[samet-akcay]

thanks for reporting this @FedericoDeBona. We'll investigate this