Unsupported Layers Update for Segment-Anything Models

[junaidsiddiqui]

We are aiming to deploy one of the more well-known CV models on an MCU, the segment-anything model by Facebook, but there are layers that are unsupported, and we can't exactly modify layers as it would greatly affect performance. Fortunately, time is not a concern for us in inference so the speed of the hardware is not a major challenge (as opposed to a GPU).

Is there some timeline or plan, or some way we can implement these layers ourselves internally?

LayerNormalization, MatMulInteger, DynamicQuantizeLinear, OneHot, ConvInteger, and Range

[Julian E.]

Hello @junaidsiddiqui,

 

Sadly, it seems very complicated to do something here... It is too early to use your model I think with the ST Edge AI Core.

 

First, more and more layers are added as the stedgeai core is updated but I can't tell you the order or priority of which they are treated.

 

You can find the list of ONNX supported layers here:https://stedgeai-dc.st.com/assets/embedded-docs/supported_ops_onnx.html

 

It is maybe possible to replace the unsupported layers by equivalent operation using other supported layers. I am not sure that this is really possible. For example, for MatMulInteger, we support MatMul but only for float32 or QLinearMatMul but I don't know if it is the same.

 

It is probably a lot of work, and it will for sure have impact on the accuracy of the model..

 

Have a good day,

Julian

[junaidsiddiqui]

Hi Julian,

Thanks a bunch!

I've used alternative layer operations for other AI architectures I customized on my own which I can tolerate some model accuracy.

I'll try to investigate other avenues if the case is as you put it. Appreciate all your help, Thanks!