TFLite running on ESP32 error: output_dim != expected_dim

Help
#1

Hello all,

I’m new to EI but have a question about a model that I’ve converted to TFLite to run on an ESP32. The model is not mine so I can’t share unfortunately however the details I can share is that it’s converted to TFLite using:

converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func_basic])
converter.target_spec.supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS, # enable TensorFlow Lite ops.
tf.lite.OpsSet.SELECT_TF_OPS # enable **TensorFlow** ops.
]

Note: I had to enable TensorFlow Ops for it to convert, its quite a complex model. This TFLite model works OK on a desktop machine but when running on the ESP32 generates the error:

tensorflow/lite/micro/kernels/pad.cpp:93 output_dim != expected_dim (69 != 67)
Node PAD (number 11f) failed to prepare with status 1
AllocateTensors() failed

This error is generated at this line of code & occurs at the pad operation seen in the image below. Its a little surprising as its a simple Pad operation. The only possible explation I can see is the Pad operation version:

DenseNet/tf.compat.v1.pad/Pad/paddings;StatefulPartitionedCall/DenseNet/tf.compat.v1.pad/Pad/paddings

From here it seems to be depreciated, I’m not sure if this is just the website I’m using to view the model or something else.

If anyone had some pointers it would be gratefully appreciated!

Regards,

Owen

image

#2

Hi @leo738,

How did you convert the TFLite model to a TFLite Micro model? Did you use the TensorFlow conversion process or something with Edge Impulse?

As we are not associated with Arduino_TensorFlowLite_ESP32, I don’t know how much we will be able to help. You might have better luck posting an issue in that repo.

#3

@shawn_edgeimpulse Thanks for the response & apologies for the late reply…

I converted the model using the TensorFlow conversion process. I suspect the model is too complex to convert to TFLite & that may be causing the errors but not certain.

Regards,