Export the torch model of EfficientAD

[wwt123-star]

I have no idea why the export torch.pt of efficientad cannot work on inference?
The commond I used is as followed:
inferencer = TorchInferencer(path='model.pt')

[blaz-r]

Can you provide more details please.

[samet-akcay]

@wwt123-star, can you use the issue template to provide more details?

[wwt123-star]

When I run the demo, this error occurs:

[blaz-r]

Hello. This happens because EfficientAD returns a dictionary. You can modify the code of torch inferencer in the following way to avoid this:

predictions = predictions["anomaly_map_combined"] 

before if statements, or change the model to return only anomaly_map_combined.

I think this was discussed somewhere before, but I can't find that thread. @alexriedel1 @samet-akcay do you think that this should be changed to just return combined map, since other two are not used anywhere as far as I see?

[alexriedel1]

the student and autoencoder maps are used for calculating the quantiles after training
https://github.com/openvinotoolkit/anomalib/blob/9a33639d4daaf521e15ae6e38d25bc11428a5462/src/anomalib/models/efficient_ad/lightning_model.py#L172-L173

the torch model could be modified to output only the combined map at inference regardless of this, but for a quick fix the inferencer can be modified @wwt123-star

[blaz-r]

Thanks for clarifying, I didn't see that it's used there, but it makes sense. I think something will need to change, because right now I think that Efficient AD can't really be used with torch or openvino inferencers as-is.

[alexriedel1]

Yes, i think the way will be to add an argument to the torch model forward method that specifies whether the st and ae maps are returned of just the combined map.