TOOL ERROR: operands could not be broadcast together with shapes (32,7200) (32,)

[rickforescue]

I have a convolutional model with convolutional layers bach normalization layers and Dense layers at the end. The model is converted to a tflite model. The inferencing works perfect on computer using tflite but When I try to deploy it on the nucleo h743zi2 I get this error.

The network layers ans its shape look like it is shown in the pic. Has anyone come across this problem?

As far my understanding goes, I did not do wrong model creation. It is some bad interpretation from STM Cube library.

Additional Info: I am using STM Cube AI version 7.1.0

Thanks in advance

Rick

[fauvarque.daniel]

Can you share the model so I can reproduce and have the development team fix the problem ?

Thanks in advance

Regards

Daniel

[rickforescue]

Hello @fauvarque.daniel​,

Thanks for quicl reply. Should I share you the keras .h5 file ?

[fauvarque.daniel]

yes please

[rickforescue]

Here is the .h5 keras file in zipped format. The model is relatively big to fit in internal flash. I quantized it using tflite Let me know if you have some other questions

Thanks

Rick

[fauvarque.daniel]

If I may, if you could provide also the quantized tflite so I have exactly the file you are using.

Daniel

[fauvarque.daniel]

I've reproduced the problem with the h5, I let you know if there is a workaround

[rickforescue]

Ok Thankyou @fauvarque.daniel​ 

[fauvarque.daniel]

The problem comes from the optimization to fold the batch normalization.

With the undocumented option "--optimize.fold_batchnorm False" the model is analyzed correctly.

You can pass the option directly to the stm32ai command line or if you are using X-Cube-AI inside STM32CubeMX you can add this option in the first screen of the advanced parameter window

Regards

Daniel

[DanF]

Is this still the correct syntax? I am using Cube AI 8.1.0 and when I add: --optimize.fold_batchnorm False" I get an unrecognized argument error. 

[DanF]

A little more information. The error itself only occurs when I attempt to quantize the model into 8 bit data types by adding these lines when creating the TF Lite model:

 

converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]

converter.inference_input_type = tf.int8 # or tf.uint8

converter.inference_output_type = tf.int8 # or tf.uint8

 

Without those lines there is no TOOL ERROR at all.

 

Update: It's only this line causing the problem:

converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]

I suspect that STM code is using some other ops and thus this fails. 

[rickforescue]

Thanks a lot @fauvarque.daniel​ . The solution works. :)

Well, I am curious. Can you give little bit more insight about it. What do you mean by folding the bachnorm ?

Thanks

Rick

[fauvarque.daniel]

During code generation there is an optimization phase that can merge some layers, a typical case is a Conv2D followed by a batchNormalization followed by a ReLU, The optimized graph will just have a Conv2D

For example for this part of the model

After the optimizer the model will look like

 

Regards

Daniel

[rickforescue]

Thanks. I see. Its interesting inisight.