Unable to use the tflite files for my Tensorflow Lite object detection

[Osakun]

**Question/Issue: Is there any way to train a model to get a Tensorflow Lite model that can be used for non-Edge Impulse TensorflowLite? TensorFlow lite converter? Or do i have no choice but to retrain using Tensorflow Lite google colab?

Context/Use case:
Hi,i am very new to Tensorflow Lite. I’m currently working on an object detection project on my Raspberry Pi 4, following the guide here (https://github.com/EdjeElectronics/TensorFlow-Lite-Object-Detection-on-Android-and-Raspberry-Pi/blob/master/deploy_guides/Raspberry_Pi_Guide.md) and all is working fine.

Now, I have trained a few models that I have previously used on my ESP32-Cam and would like to use the same models for my Raspberry PI. I have used the Lite file on the dashboard and renamed them tflite but there were some issues getting them to work. I have tried the ones u can get from the deployment tab “OpenMV Library” but its not working as well. For the dashboard files:

Using the quantized int8 model, the error shows “ValueError: Cannot set tensor: Got value of type UINT8 but expected INT8 for input 0, name: serving_default_x:0”

While if I were to use the float32 model, the error shows to be Bounding box coordinates of detected objects IndexError: List index out of range.

[Eoin]

Hello @Osakun, welcome to the community!

Nice work! @shawn_edgeimpulse would be best to chime in here, as he has experience with the Coral and created a lot of content detailing the setup required for it.

You should be able to export and import the model via our Python SDK in your colab.

Edge Impulse Python SDK

Let us know how you get on with integrating our python sdk into your colab project!

Best

Eoin

[shawn_edgeimpulse]

Hi @Osakun,

The error you are seeing is likely due to a mismatch in runtime versions. You must use the correct TensorFlow Lite engine version that matches with the model version. To check your model version, I recommend loading the .lite file into netron.app and clicking on the input layer at the top. You should see the required runtime version listed. For example, here is a model I trained recently loaded into netron that requires runtime v1.5.0.