Reshape error in ssd_mobilent_v3 object detection model in TensorflowLite Micro

[Ramson]

Hi team,

We have trained the object detection model on VOC dataset using eIQ toolkit and exported the model as .tflite file which I have attached below. When running the model on RT1170EVK. We're facing an error in the reshape operator as shown below:

.reshape.cpp:70 num_input_elements != num_output_elements (126 != 21)
.reshape.cpp:70 num_input_elements != num_output_elements (126 != 21)
.reshape.cpp:77 ReshapeOutput(context, node) != kTfLiteOk (1 != 0)
Node RESHAPE (number 147f) failed to prepare with status 1
AllocateTensors() failed
Failed initializing model

When we looked into the model using model tool, we think this is the node.

How to solve this error? Please guide us. Thanks in Advance.

Regards,

Ramson Jehu K

[anthony_huereca]

Hi Ramson,
  We were able to recreate the error with your model and our development team is working on a fix. I'll provide an update when I have more information. 

-Anthony 

[Ramson]

Thanks Anthony, waiting for your response only.

[anthony_huereca · Accepted solution]

Hi Ramson,

  The root cause is an issue in the TFLite conversion step when exporting the model as a .tflite file. It makes use of the TensorFlow tools and this bug has now been logged in the Tensorflow github: https://github.com/tensorflow/tensorflow/issues/52028 

  In the meantime, if not tried already, try using the TOCO converter which may be able to avoid this issue.  You can also try using the DVRT inference engine so you don't have to go through the conversion step with the model. 

-Anthony 

[Ramson]

Hi Anthony, 

Thank you so much for the insights that you have provided. Sure I will try using TOCO convertor and check. 

I'm also kinda stuck with the output interpretation which i have mentioned here : https://community.nxp.com/t5/eIQ-Machine-Learning-Software/What-is-the-output-signature-of-the-ssd-m... . It would be a great help if you could get us through it.

Thanks in Advance

Ramson