Cfa predictions are NaN when using lightning inference

[lweggartner]

Hey all,

I am trying to implement Cfa. I am already having difficulties when using the "getting_started" notebook. ( I have changed it slightly to use the lightning inference. You can see the modified version in the attachement)

I have noticed that the pred_scores are NaN.

# load model for inference
model_path = "results/Cfa/axi_images/latest/weights/lightning/model.ckpt"
model_path = '/home/azureuser/cloudfiles/code/Users/weu4rt/anomalib/results/Cfa/MVTec/bottle/v5/weights/lightning/model.ckpt'
model_pred = Cfa.load_from_checkpoint(model_path)

# inference
prediction = engine.predict(model = model_pred, datamodule= datamodule, ckpt_path= model_path)

# anomaly map of first image
(prediction[0]['anomaly_maps'][0]) # results in following tensor
tensor([[[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]]])

Has anyone faced similar issues with Cfa or has an idea how to fix this behaviour?

Thanks in advance!

Attachement:
getting_started.txt