Use in production

[terrarier2111]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

[xrli-U]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

[terrarier2111]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

Alright I tried modifying the MvtecAD config to use my custom data set and it somewhat worked, however unfortunately the model didn't save it's predictions for each image at the very end but only calculated AUROC (and similar) scores, but i want to get the predictions the model makes as an output, how would i do this?

[xrli-U]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

Alright I tried modifying the MvtecAD config to use my custom data set and it somewhat worked, however unfortunately the model didn't save it's predictions for each image at the very end but only calculated AUROC (and similar) scores, but i want to get the predictions the model makes as an output, how would i do this?

If all goes well, the visualization function on line 96 of the models/musc.py executes and saves the prediction for each image. You can focus on debugging this function, such as the problem about the save path.

[terrarier2111]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

Alright I tried modifying the MvtecAD config to use my custom data set and it somewhat worked, however unfortunately the model didn't save it's predictions for each image at the very end but only calculated AUROC (and similar) scores, but i want to get the predictions the model makes as an output, how would i do this?

If all goes well, the visualization function on line 96 of the models/musc.py executes and saves the prediction for each image. You can focus on debugging this function, such as the problem about the save path.

This does produce some heatmaps, however unfortunately it also doesn't save and of the internal anomaly scores (values from 0.0 to 1.0) of musc, is there a way to do that (note that i am not interested in the delta between the ground truth and the predictions, but only the predictions itself, as my ground truth is a dummy)
is it sufficient to just store scores_cls from inside the visualize function and how would i know which score is associated with which image?

[xrli-U]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

Alright I tried modifying the MvtecAD config to use my custom data set and it somewhat worked, however unfortunately the model didn't save it's predictions for each image at the very end but only calculated AUROC (and similar) scores, but i want to get the predictions the model makes as an output, how would i do this?

If all goes well, the visualization function on line 96 of the models/musc.py executes and saves the prediction for each image. You can focus on debugging this function, such as the problem about the save path.

This does produce some heatmaps, however unfortunately it also doesn't save and of the internal anomaly scores (values from 0.0 to 1.0) of musc, is there a way to do that (note that i am not interested in the delta between the ground truth and the predictions, but only the predictions itself, as my ground truth is a dummy) is it sufficient to just store scores_cls from inside the visualize function and how would i know which score is associated with which image?

The pixel-level prediction results are stored in pr_px and the image-level prediction results are stored in pr_sp.
image_path_list stores the paths for each image in the same order as in the prediction results pr_px and pr_sp.

[terrarier2111]

Do you think this implementation is usable in real production circumstances?
Also how can i give this model a path to an image and get an anomaly score as an output?

Yes, this implementation is able in real production circumstances, and some companies are already using our algorithm. For the issue of single image inference, you can refer to issue 30 (#30). In addition, we have improved the code for single image inference, but it is testing on the production line. So I am sorry that this new code is unavailable for the time being.

Alright I tried modifying the MvtecAD config to use my custom data set and it somewhat worked, however unfortunately the model didn't save it's predictions for each image at the very end but only calculated AUROC (and similar) scores, but i want to get the predictions the model makes as an output, how would i do this?

If all goes well, the visualization function on line 96 of the models/musc.py executes and saves the prediction for each image. You can focus on debugging this function, such as the problem about the save path.

This does produce some heatmaps, however unfortunately it also doesn't save and of the internal anomaly scores (values from 0.0 to 1.0) of musc, is there a way to do that (note that i am not interested in the delta between the ground truth and the predictions, but only the predictions itself, as my ground truth is a dummy) is it sufficient to just store scores_cls from inside the visualize function and how would i know which score is associated with which image?

The pixel-level prediction results are stored in pr_px and the image-level prediction results are stored in pr_sp. image_path_list stores the paths for each image in the same order as in the prediction results pr_px and pr_sp.

Thank you very much, this is such a great project, thank you for sharing your research :)