Didn't find op for builtin opcode 'SUM' version '1'.

[javierFerreroM]

@tensorflow/micro

System information

• Host OS Platform and Distribution: Linux Ubuntu 18.04:
• TensorFlow installed from binary
• Tensorflow version : 2.3
• Target platform: ARM 64

Describe the problem
In the following lines I proceed to describe the steps followed:

1. Adapt MobileNetv2 model. Loaded for training with Keras API. Take out the last layer by setting include_top = false, and adding a customized last Conv2D layer with the Functional API. This is done in order to apply the convolutional sliding window approach. (Bigger input image, than the images used for training).
2. Training done with success. Translation of the model from TF to TF Lite and it runs and provides a sensible matrix of results.
3. Translation of TFLite model to TF Tiny model (model.cc) by using the xxd -i ... command provided in the documentation.
4. When allocating the model, we encounter the following failure:

libraries ready
STM32 Tensorflow Lite test
Model working
Interpreter working
Didn't find op for builtin opcode 'SUM' version '1'. An older version of this builtin might be supported. Are you using an old TFLite binary with a newer model?

Failed to get registration from op code SUM

Failed starting model allocation.

Allocate working
Type input: 1
Bytes input: 1310720
Size input: 4
Dim input 0: 1
Dim input 1: 512
Dim input 2: 640
Dim input 3: 1

As seen in the message, dims are correct, type input is correct (Float) and everything else seems to be working fine.

Prior to this issue, a similar fault appeared as it requested to have the op code EXP. This one is already available in Tensorflow Git and it has been already implemented. (exp.cc, exp_test.cc, exp.h, AddEpx() included in AllOps file and make file modified).

The same procedure has been handled for the MobileNetV2 with no modifications and it worked fine.

I have two main questions:

• Why do we need extra operators, if we only added one extra Conv2D layer?
• Is SUM implemented, and will we have further operators needed?

Please provide the exact sequence of commands/steps when you ran into the problem

Model configuration in python (Tensorflow + Keras API):
model = MobileNetV2(include_top=False,weights=None,input_tensor=Input(shape = (512,640,channels) ,dtype = 'float32'),pooling = None,classes = len(class_labels))
last = Conv2D(filters = 5,kernel_size = 3, padding= 'valid',strides=(1,1),activation='softmax', input_shape = (model.layers[-1].output.shape))(model.layers[-1].output)
model = Model(model.input, last)

Model Conversion from Tensorflow to Tensorflow Lite:
converter = tf.lite.TFLiteConverter.from_keras_model(loaded_model)
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS]
model_no_quant_tflite = converter.convert()
Model Conversion from Tensorflow Lite to Tensorflow Tiny:
xxd -i model_mobilenet_sliding2.tflite > model.cc

Model loading in Tensorflow Tiny (C++):

namespace{ tflite::ErrorReporter* error_reporter = nullptr;
const tflite::Model* model = nullptr;
// This pulls in all the operation implementations we need
tflite::AllOpsResolver resolver;
constexpr int kTensorArenaSize = 4 * 1024 * 1024;
uint8_t tensor_arena[kTensorArenaSize];
uint8_t* img = nullptr;
uint64_t timeImage;
uint64_t timeInvoke;
TfLiteTensor* model_input = nullptr;
TfLiteTensor* model_output = nullptr; }

// Set up logging (modify tensorflow/lite/micro/debug_log.cc) static tflite::MicroErrorReporter micro_error_reporter; error_reporter = &micro_error_reporter;

// Say something to test error reporter error_reporter->Report("STM32 Tensorflow Lite test");

model = ::tflite::GetModel(model_mobilenet_sliding2_tflite); cout << "Model working"<<endl;

tflite::MicroInterpreter interpreter(model, resolver, tensor_arena, kTensorArenaSize, &micro_error_reporter);

cout << "Interpreter working" << endl;

interpreter.AllocateTensors();

cout << "Allocate working" << endl;

model_input = interpreter.input(0);

// Get image from provider.
cout << "Type input: " << model_input->type << endl;
cout << "Bytes input: " << model_input->bytes << endl;
cout << "Size input: " << model_input->dims->size << endl;
cout << "Dim input 0: " << model_input->dims->data[0] << endl;
cout << "Dim input 1: " <<model_input->dims->data[1] << endl;
cout << "Dim input 2: " <<model_input->dims->data[2] << endl;
cout << "Dim input 3: " <<model_input->dims->data[3] << endl;

[Saduf2019]

@javierFerreroM
Could you please try on tf 2.4.1 or 2.5rc1 and let us know if you still face the issue.

[javierFerreroM]

Good morning @Saduf2019 ,

I just tested both versions and in both we encounter the same issue. The project is generated out of the hello_world example using the following command:

make -f tensorflow/lite/micro/tools/make/Makefile TARGET_ARCH=x86_64 generate_hello_world_make_project

I used the same model.cc file for the three TF Tiny versions ( I am assuming that the model is not the issue, but the missing libraries).

An interesting point I'd like to highlight is that I have checked that in version 2.5rc1, exp.cc is already present in the kernels folder, and it is already integrated in the make file. The only missing part is the AddExp(); that I added by hand to the All_Ops_resolver.cc

Thus, I integrated it and so, the error of missing operand EXP disappears, but we arrive to the same point with the missing SUM operand.

libraries ready
STM32 Tensorflow Lite test
Model working
Interpreter working
Didn't find op for builtin opcode 'SUM' version '1'. An older version of this builtin might be supported. Are you using an old TFLite binary with a newer model?

Failed to get registration from op code SUM

Failed starting model allocation.

Therefore, I guess that the same procedure should be followed, to include the SUM operand, and I am not sure whether new faults of the same characteristics may rise.

Thank you very much, I stand by awainting your response.

[mohantym]

Hi @javierFerreroM !
We are checking to see whether you still need help in this issue .
Earlier, The work around was to add the op from OpResolver in the C++ code.

Attached similar issues for reference.

1, 2

Feel free to test documentation from tflite-micro repo and post on the same repo for further assistance.

Thank you!

[google-ml-butler]

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

[google-ml-butler]

Closing as stale. Please reopen if you'd like to work on this further.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No