Problem during openVINO inference

[NotAleric]

Dear Anomalib Team,

I am using Anomalib to detect anomalies in my dataset, which contains maps of displacements. I have followed the getting_started file from the notebooks section and encountered an issue during the inference phase using OpenVINO.

Here is a brief overview of my process:

Dataset and DataModule: I used the Folder class to create the datamodule with the following code:

datamodule = Folder(
    name="abejas",
    root=Path.cwd() / "mapas/datos_noviembre",
    normal_dir="normales",
    abnormal_dir="anomalas_rural",
    task=TaskType.CLASSIFICATION,
)
datamodule.prepare_data()  # Create train/val/test/predict dataloaders

datamodule.setup()  # Split the data into train/val/test/prediction sets.

Model Creation: I created a basic Padim model:

model_padim_final = Padim()

Training the Model: I used the following code to train the model:

engine = Engine(task=TaskType.CLASSIFICATION)
engine.fit(model=model_padim_final, datamodule=datamodule)

The training process completed successfully.

Testing the Model: The testing phase also yielded good results:

test_padim = engine.test(
    model=model_padim_final,
    datamodule=datamodule,
    ckpt_path=engine.trainer.checkpoint_callback.best_model_path,
)

Results:

Copiar código ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Test metric ┃ DataLoader 0 ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │ image_AUROC │ 0.9957906007766724 │ │ image_F1Score │ 0.9931600689888 │ └───────────────────────────┴──────────────────────────┘

Exporting the Model to OpenVINO: I used the following code to export the model to OpenVINO:

engine.export(model=model_padim_final, export_type=ExportType.OPENVINO)

output_path = Path(engine.trainer.default_root_dir)

openvino_model_path = output_path / "weights" / "openvino" / "model.bin"

metadata = output_path / "weights" / "openvino" / "metadata.json"

inferencer = OpenVINOInferencer(
    path=openvino_model_path,  # Path to the OpenVINO IR model.
    metadata=metadata,  # Path to the metadata file.
    device="CPU",  # Run on an Intel CPU.
)

Inference Issue: When I use the inferencer to predict a new image from the test data, the prediction always indicates a 100% probability of the image being abnormal. I used the following code for inference:

i, data = next(enumerate(datamodule.test_dataloader()))
images_test = data['image']
image_numpy = images_test[0].numpy()

predictions = inferencer.predict(image=image_numpy)

print(predictions.pred_score, predictions.pred_label)`

Results of Prediction:

1.0 LabelName.ABNORMAL

Dataset Verification: I have verified the number of images in each class to ensure the Folder class accesses the folders correctly:

Number of normal images: 2327 Number of abnormal images: 727

Image Shape Verification: I have checked the shape of the images at various stages:

Before using the Folder class: (768, 1366, 4) In the Folder class: torch.Size([3, 768, 1366]) In the test datamodule: torch.Size([3, 768, 1366]) Transposed for OpenVINO inference: (1366, 768, 3)

Installation: I installed Anomalib from source using the following commands in a new conda environment:

!git clone https://github.com/openvinotoolkit/anomalib.git

%cd anomalib

%pip install .

%anomalib install -v

Despite achieving good results during testing with various models (Dfm, PathCore, Padim), the inference using OpenVINO consistently fails, indicating the problem is likely not with the model itself. I have tested with many images and used various methods for loading the images (from the Folder datamodule and directly from my original folder), but the issue persists.

I dont know if the problem is related with the dataset or even if the model is overfiiting.

I have attached a file with an example from my dataset in case the problem is related to the images.

I would appreciate any guidance or suggestions you could provide to resolve the inference issue with OpenVINO.

Thank you for your assistance.

Best regards, Alex

[him192021]

I got this problem too. I checked the read_image() function and I found that they divide the image by 255. After I divided the image(in np.array format) by 255, this problem was solved,

[haimat]

I got this problem too. I checked the read_image() function and I found that they divide the image by 255. After I divided the image(in np.array format) by 255, this problem was solved,

@him192021 Hi, could you please elaborate on exactly what steps you take to correctly infer an image?

[him192021]

@haimat something like: image = Image.open(path).convert("RGB") image=np.array(image) / 255.0

In my case, the result was fine when using anomalib.data.utils.read_image. However, the result was always 100% defective when I use cv2/PIL to read the image. Then I printed the heat map and found that the image becomes strange. It seems that the OpenVINOInferencer.predict expect the input between 0 and 1.

[haimat]

@him192021 Thanks for your response!

[NotAleric]

Thanks for the replies @him192021 and @haimat.

In my case it still doesn't work despite rescaling the images between [0,1] so that the inference with OPENVINO works correctly.

In this image you can see how I am reading the images (normal and anomalous) with the function read_image of anomalib.data.utils.

When doing inference of the normal image, it still comes out abnormal 100% in spite of having images scaled between [0,1].

It is probably a problem with my data as they are quite complex to analyse by neural networks.