[Task]: Inference on a test set

[dangmanhtruong1995]

What is the motivation for this task?

The example on the github page shows only inference for an image.

Describe the solution you'd like

If anyone know how to do inference on a test set that would be great.

Additional context

No response

[samet-akcay]

Hi @dangmanhtruong1995, --input and --output flags checks whether it is an input to an image or folder containing images. Assuming that you train a padim model on MVTec bottle category, you could run torch_inference.py to perform an inference on a number of images from a directory.

python tools/inference/torch_inference.py \
    --config anomalib/models/padim/config.yaml \
    --weights results/padim/mvtec/bottle/run/weights/model.ckpt \
    --input datasets/MVTec/bottle/test/broken_large \    # Path to broken_large images in the test set.
    --output output/broken_large    # An output directory to save the predicted images.

Let me know if this is what you are looking for.

[samet-akcay]

I'm converting this to a Q&A in discussions. Let's continue there.

[dangmanhtruong1995]

Hi @samet-akcay , so I have a test folder containing images in "abnormal" and "normal" subfolders, I want to do inference on all of those images and then output F1 score, AUC, .... as when we trained before.

[samet-akcay]

Hm, sounds like tools/test.py might suit better in this case.

[dangmanhtruong1995]

Hm, sounds like tools/test.py might suit better in this case.

So how do I run using tools/test.py ? Because on the readme it does not talk about this.

[samet-akcay]

https://github.com/openvinotoolkit/anomalib/blob/main/tools/test.py#L23-L25

You will need to provide --model or --config and --weight_file as arguments, such as the following:

python tools/test.py \
    --config </path/to/your/config/file>
    --weight_file </path/to/your/weight/file>

if you use this script, make sure that your the data that you want to test follows the same structure as specified in your config file (mvtec, btech, folder etc).

[dangmanhtruong1995]

Hi, I have managed to run it as you said, thanks a lot.