Please help! I have 'ValueError: No samples to concatenate' error while using the trainer.test functionality

[1amrutesh]

I want to use the anomalib library on a custom dataset. However, even after I specify the path, I get the runtimeerror. The structure of the directory is

• Bad
• • -Images
• Good_train
• --Images
• Good_test
• --Images
  The images are in the Images folder. There are 200 abnormal images and 1800 normal images with 80:20 split in good_train and good_test respectively. I used the Padim model.
  I have tried to implement this on Kaggle with P100 GPU. I carefully followed the [kaggle guide]

I have also attached the screenshot of the error. I shall be obliged if anyone helps me.

[djdameln]

Hi, it's hard to say for sure without knowing the contents of your config.yaml, but the most likely explanation is that the task type for your dataset is configured as segmentation, while your dataset does not have ground truth pixel-level annotations available. As explained in this answer, you will have to set the dataset.task parameter in the config.yaml to classification in this case.

[1amrutesh]

Yes, thanks a lot for your help. After setting the task 'classification' instead of 'segmentation', I can run the code trainer.test(model=model, datamodule=datamodule) successfully, and the above error does not occur. However, the classification task can only classify whether the image is anomalous; it cannot localize it. Is it possible to localize the anomalies in the image without the 'pixel-level annotations'?

[djdameln]

Yes, after training the model with the classification task type, you can use the model to generate pixel-level predictions during inference. Just create a new config.yaml in which you set the task parameter to segmentation and run the inference entrypoint script.

Please be aware that Anomalib uses the validation set to adaptively compute an optimal value for the anomaly score threshold on both the image- and pixel-level. In the absence of pixel-level annotations, the image-level threshold will be re-used as pixel-level threshold, which may negatively impact the quality of the segmentation predictions.