EIQ-Neutron conversion problem from int8 to uint8 for Softmax Layer

fylips98 — Tue, 11 Mar 2025 11:59:54 GMT

Hello, I wanted to ask a question about the conversion tool in eIQ portal, and in particular to the conversion "eiq converter-neutron".
I am bringing my own model, training a time series model that uses a Transformer structure, and converting it in TFlite with my own dataset outside the eiq portal.

This quantized model has all the layers inputs and outputs converted to int8 as verified by the graph in the eiq portal. The problem is that when I use eiq converter-neutron the softmax output gets converted into uint8, and this produces an error in developing phase, since TF lite for microcontrollers wants the softmax layer to have input and output of the same format.

Are there some workarounds to this?
Thank you in advance for your help!

Re: EIQ-Neutron conversion problem from int8 to uint8 for Softmax Layer

Celeste_Liu — Fri, 14 Mar 2025 06:11:36 GMT

Hello @fylips98 ,

Thanks for your post.

Please try to quantize original Tensorflow model with eIQ and set the data type to int8 in quantization option.

Best Regards,
Celeste

-----------------------------------------------------------------------------------------------------------------
Note: If this post answers your question, please click the "ACCEPT AS SOLUTION" button. Thank you!
-----------------------------------------------------------------------------------------------------------------

topic EIQ-Neutron conversion problem from int8 to uint8 for Softmax Layer in MCX Microcontrollers

EIQ-Neutron conversion problem from int8 to uint8 for Softmax Layer

Re: EIQ-Neutron conversion problem from int8 to uint8 for Softmax Layer