2343597_en-US

[#1]

EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Dear NXP Support,

I would like to report an issue with EfficientDet-Lite0 on i.MX95 using the Neutron conversion flow.

My setup is:

• Board: i.MX95 EVK

• BSP: Linux imx95evk 6.12.34-lts

• Original model: efficientdet_lite0_int8.tflite

• Application: NNStreamer pipeline with TensorFlow Lite and external Neutron delegate

• Input: live camera stream from /dev/video13

I compared the original model and the Neutron-converted model using the same application and added debug prints for both the model input tensor and output tensors.

Findings:

1. The original model efficientdet_lite0_int8.tflite runs correctly on CPU.

2. For the original model, the input tensor changes frame by frame, and the output tensor statistics also change frame by frame.

3. For the Neutron-converted model, the input tensor still changes frame by frame, but the output tensors remain almost constant across frames.

4. As a result, detections are static or empty even though the live camera image is changing.

I tested two Neutron paths:

A) Standalone Neutron SDK 3.0.1

• Converted with eiq-neutron-sdk-windows-3.0.1\bin\neutron-converter.exe

• Runtime on target updated to 3.0.1

• Result: converted model loads, but outputs remain nearly constant and produce static boxes

B) Rolled back to the 6.12.34-era toolchain

• Converted with:
  C:\nxp\eIQ_Toolkit_v1.17.0\bin\neutron-converter\MCU_SDK_25.09.00+Linux_6.12.34_2.1.0\neutron-converter.exe

• Runtime rolled back accordingly

• Result: converted model loads, input changes, but outputs are still nearly constant and detections remain empty

Example behavior with the rolled-back conversion/runtime:

• Input changes:

  • INPUT range: min=0 max=255 mean=141.903 sum=43592724

  • INPUT range: min=0 max=255 mean=137.091 sum=42114370

  • INPUT range: min=1 max=255 mean=105.373 sum=32370478

• Output remains nearly constant:

  • classes_raw range: min=0.000000 max=0.019531 mean=0.000009

  • boxes_raw range: min=-2.926135 max=0.535502 mean=-0.30798

The conversion statistics for the rolled-back path show:

• operators after extract: 33

• Neutron graphs: 4

• operators total: 371

• operators converted: 342

• operators NOT converted: 29

This suggests that the live input reaches the model correctly, but the converted Neutron model path is not producing frame-dependent outputs for EfficientDet-Lite0.

Could you please advise:

1. Whether EfficientDet-Lite0 is expected to be supported correctly on i.MX95 with BSP LF_6.12.34_2.1.0.

2. Whether this model is known to have issues with Neutron conversion on i.MX95.

3. Whether there is a recommended reference object detection model for Neutron on this BSP that is known to behave correctly with live video.

4. Whether this behavior indicates a converter issue, a delegate/runtime issue, or an unsupported operator pattern for EfficientDet-Lite0.

Thank you in advance!

Best regards,
EOF

Re: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0Hi @Rita_Wang,
Thank you, I look forward for your reply.
Have a nice day you too!
Best regards,
EOFRe: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Hi @EOF ，

We will check in our side and then give you reply, please wait.

Wish you have a nice day

Best Regards

Rita

Re: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Hi @Rita_Wang,

I am using the i.MX 95 19mm x 19mm evaluation kit (EVK) (i.MX 95 19mm x 19mm Evaluation Kit | NXP Semiconductors)

With 

• BSP: Linux imx95evk 6.12.34-lts

I've attached the converted model (trained in COCO dataset) in this message. 

This model was converted using (no additional flag only --input,--ouput,--target):

• C:\nxp\eIQ_Toolkit_v1.17.0\bin\neutron-converter\MCU_SDK_25.09.00+Linux_6.12.34_2.1.0\neutron-converter.exe

Best regards, 

EOF

Re: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Hi @EOF ，

Here we need more information, Could you please provide the customer's model for further analysis?

Wish you have a nice day

Best Regards

Rita

Re: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Hi @Rita_Wang ,

Thank you for your reply.

Yes, the non-coverted file is attached to this message. Thank you again, I look forward for your reply. 

Have a nice day!

Best regards,

EOF

Re: EfficientDet-Lite0 Neutron conversion on i.MX95 produces static outputs on LF_6.12.34_2.1.0

Hi @EOF ,

Could you provide us the source model ,not the neutron converted model to us ?  We can  try to convert the model on our side and validate if there are any accuracy issues. Thanks.

Wish you have a nice day

Best Regards

Rita