openvino error about ngraph HOT 16 CLOSED

The problem is from the input to PriorBoxClustered; 2x2760708352 doesn't look like the shape you would expect as an input.

from ngraph.

Thank you, so you say that is not from your library, is from the wrong format that the input layer receives no, so the problem is in the openvino path, correct?

from ngraph.

Not exactly; something went wrong with the size of the input to PriorBoxClustered, probably a bad signed/unsigned conversion. Wherever that big number came from is the op with the problem. If there's a PriorBox in that area, @4745 might be the fix and it might not be in the OpenVINO sources you are using yet.

from ngraph.

Ah ok, i understand and thanks for your reply!!

You know if there is a patch i can apply ?

from ngraph.

I should have said #4745 . It's a pretty simple change.

from ngraph.

Thank you so much!! i will try to apply that patch and compile again

I see today openvino repo was updated and i see they removed ngraph module.... i have to take a look

from ngraph.

Sorry to disturb again, the branch that was linked in openvino was the master branch, in that branch the files mentioned in that pull request are in a different location, they are under a folder named "experimental". Will work if i apply those patches in that location or i will need to use your dynamic branch instead master?

Thanks again

from ngraph.

It should work; it's just making the computation signed so that negative numbers don't turn into large positive numbers

from ngraph.

sadly it doesn't work, keep trowing same error

from ngraph.

Do you know the shape of the PriorBox input and the attributes it of the PriorBox op?

from ngraph.

No, sorry

But if you tell me how to get that info i will do it

from ngraph.

If helps error seems related to only raspberry enviroment and "caused by the new IR v10 format.."

from ngraph.

Are you starting with an OpenVINO graph? I think the PriorBox attributes (min_size, aspect_ratio, flip, clip, step, offset, variance, scale_all_sizes, fixed_ratio, fixed_size, density) would be almost the same in that graph, as would the inputs to the PriorBox (output_size and image_size tensors).

from ngraph.

You mean to use the workbench to visualize a graph of the model? Sadly you can not do that on raspberry, as the raspbian package do not have those tools....

I can do that in a ubuntu machine if helps....

https://docs.openvinotoolkit.org/latest/_docs_Workbench_DG_Visualize_Model.html

from ngraph.

I checked with some OpenVINO people, and they ask that you move the conversation over there since there are multiple places things could be going wrong (and any fixes will be in the ngraph in the OpenVINO repo).

from ngraph.

Thank you, i appreciate your help!!

from ngraph.