[Question] How do I infer from numpy array/tensor?

[vmiller987]

Hello,

I apologize for making this an issue, but the new discussion button seems to be missing. I even checked on another machine.

Is there a simple way to infer from a numpy array/tensor and not from file? My datasets are given to me as an x-ray image with multiple objects in this image. We use another model to bounding box and crop out individual items and this stays in memory.

If someone could point me in the right direction, I'd greatly appreciate it.

[sherpan]

I believe one is supposed to use the

anomalib.deploy.TorchInferencer class for that

https://anomalib.readthedocs.io/en/v2.0.0/markdown/guides/reference/deploy/index.html#anomalib.deploy.TorchInferencer:~:text=3%2C%20224%2C%20224%5D)-,predict(image),-%23

[vmiller987]

Excellent. Thank you for pointing this out.

[samet-akcay]

@vmiller987, we have moved to a new org and creating a discussion may be disabled. I'll double check.

Regarding the inference, I agree with @sherpan. You could pass your array/tensor to TorchInferencer. To use TorchInferencer, you need to export your model to torch though. Either via model.to_torch(..) or Engine.export(..)

[vmiller987]

I still don't have the ability to make discussions it seems. I had one more question I was hoping you could point me in the right direction.

This is a bit of a novice question I think. From what I understand, PaDiM does one epoch for feature extraction to generate it's model on. I have a dataset with ~13000 images. I must split this dataset down to ~9500 images in order to avoid OOM issue on the GPU. In standard CNN models, usually you can adjust the batch size or have to tile/patch/chunk the images. I haven't tried Anomalib's tiler yet, as part of me assumes this would run into the same issue, however I could be wrong. I've tried halving the batch_size to see no change. I do know that other models that do the 1 epoch kind of run into similar issues.

What would be the appropriate way to handle this situation?