[Bug]: OpenVINO vs Torch inferencer on CPU gives significantly different results

[FedericoDeBona]

Describe the bug

From #2447, training a model, then exporting it in both OpenVINO and Torch the visual result and pred_score are a lot different. Looks like a normalization problem but in the metadata of both models, the values are the same.

OpenVINO inference - pred_score: 0.7632662161599639

Torch inference - pred_score: 0.6897885799407959

Full code

from anomalib.data import MVTec
from anomalib.models import Patchcore
from anomalib.engine import Engine
from anomalib.deploy import ExportType
from anomalib.deploy import OpenVINOInferencer, TorchInferencer
from PIL import Image
from glob import glob
from anomalib.utils.visualization import ImageVisualizer
from anomalib.utils.visualization.image import  VisualizationMode
from anomalib import TaskType
import os
import cv2

CATEGORY="screw"

datamodule = MVTec(category=CATEGORY)
model = Patchcore()
engine = Engine()

engine.fit(datamodule=datamodule, model=model)
engine.export(model=model,export_type=ExportType.TORCH) 
engine.export(model=model,export_type=ExportType.OPENVINO) 

vino_inferencer = OpenVINOInferencer(
    path=f"/home/trainer/trainer_engine/results/Patchcore/MVTec/{CATEGORY}/latest/weights/openvino/model.bin",
    metadata=f"/home/trainer/trainer_engine/results/Patchcore/MVTec/{CATEGORY}/latest/weights/openvino/metadata.json",
    device="CPU")
torch_inferencer = TorchInferencer(path=f"/home/trainer/trainer_engine/results/Patchcore/MVTec/{CATEGORY}/latest/weights/torch/model.pt", device="cpu")
print(vino_inferencer.metadata)
print(torch_inferencer.metadata)
#Output
"""
{'task': 'segmentation', 'image_threshold': 30.5954647064209, 'pixel_threshold': 34.377254486083984, 'anomaly_maps.min': 9.125215530395508, 'anomaly_maps.max': 49.421302795410156, 'pred_scores.min': 25.861536026000977, 'pred_scores.max': 52.29983139038086}
{'task': <TaskType.SEGMENTATION: 'segmentation'>, 'image_threshold': 30.5954647064209, 'pixel_threshold': 34.377254486083984, 'anomaly_maps.min': 9.125215530395508, 'anomaly_maps.max': 49.421302795410156, 'pred_scores.min': 25.861536026000977, 'pred_scores.max': 52.29983139038086}

"""


visualizer = ImageVisualizer(mode=VisualizationMode.FULL, task=TaskType.SEGMENTATION)

for defect_path in sorted(glob(f"/home/trainer/trainer_engine/datasets/{CATEGORY}/test/*")):
	defect = os.path.basename(defect_path)
	print(f"===== {defect} =====")
	for img_path in sorted(glob(f"/home/trainer/trainer_engine/datasets/{CATEGORY}/test/{defect}/*")):
		img_name = os.path.basename(img_path)
		vino_res = vino_inferencer(cv2.imread(img_path))
		torch_res = torch_inferencer(img_path)
		print("VINO")
		print(vino_res.pred_score)
		display(Image.fromarray(cv2.resize(visualizer.visualize_image(vino_res), (500*3,125*3))))
		print("TORCH")
		print(torch_res.pred_score)
		display(Image.fromarray(cv2.resize(visualizer.visualize_image(torch_res), (500*3,125*3))))

Dataset

MVTec

Model

PatchCore

Steps to reproduce the behavior

See above

OS information

OS information:

• OS: Ubuntu 24.04
• Python version: 3.10.14
• Anomalib version: 1.2.0.dev0
• PyTorch version: 2.4.0+cu118
• CUDA/cuDNN version: 11.8
• GPU models and configuration: GeForce RTX 3090 Ti

Expected behavior

Not sure, I suppose the exported models should have almost equal result in both OpenVINO and Torch

Screenshots

No response

Pip/GitHub

GitHub

What version/branch did you use?

1.2.0.dev0

Configuration YAML

-

Logs

-

Code of Conduct

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