Edge AI

I am using the XCUBE-AI software to deploy an SNN (Spiking Neural Network) on my STM32H747I-DISCO board. The SNN is in the form of a .h5 file, and was converted from a CNN using the snn toolbox. The SNN ofcourse, has a lot of custom layers. I want to analyze this network using XCUBE-AI. I used two methods:1) Choose the 'keras model and topology' option, and create the config file which contains the python files containing the class definitions of the custom layer.2)Convert the .h5 file into tflite and upload the tflite model.However, both these approaches give me 'error' while analyzing. It doesn't tell me what went wrong. Is there a log of some kind where I can see what is actually going wrong? Otherwise it's hard to debug.

When I try to Validate model on target using CubeIDE tool I am facing this error.

I have a saved tflite model , which I want to deploy on my STM32H747I-DISCO board. For now, I just want to run validation on target using one custom input. I have the input as a numpy array (saved as a .npy file, it's a 4D array). Can someone please highlight the steps to do so?

..

Now I am working on x-cube-ai on STM32F407,but I find little information about the APIs,please help me. Thanks!

I'm working on a TinyML project with an Arduino Nicla Vision (STM32H747AII6) and I need to store a model with size 608kB in RAM. In the Nicla Vision datasheet, it is stated that 864kB of SRAM are available, but it is segmented in three non contiguous parts (AXI SRAM, SRAM1-3, SRAM4).When I try to load the model in RAM, I get the error 'linker_script.ld:110 cannot move location counter backwards (from 240aa610 to 2407fc00)' because only the AXI SRAM (524kB) is used.Is there a way to distribute the model over all the available non contiguous RAM?

Hi, I am trying to implement an audio classifier on stm32f746g-discovery board. I converted my keras model into a tflite file and successfully loaded it in the board.I used integer only quantization on my model in order to decrease ram usage. I think it needs uint8_t type input. Currently I am having an issue converting float values to uint8_t type.The model seems to be giving some outputs when I give some uint8_t array as an input but the outputs are nowhere near expected. Here is how I try to convert float values to uint8_t:   float array[4290];  uint8_t sound[4290];  for(int i=0; i<4290; i++){  array[i] = 255*(array[i]-min)/(max-min);  sound[i] = (int)array[i];  }Thanks a lot.

I am currently trying to run an ML model on the STM32H7A3. The model is uploaded correctly and successfully analysed.The first issue is that when generating the code, the middleware folder which contains ST>ai (which contains ai_platform) is deleted by CubeIDE. I've had to overcome this by copying the middleware folder from an older project, as there is no way to generate it. Hopefully, ST knows about this.Anyway, when I run the model, the correct input and outputs are present. This means I can fill in the input buffer. I do not enable the use activation buffer for input buffer as shown in the screen show:

Hi.I'm trying to run a neural network model using X-Cube-AI, but at the point of performing the inference, my program gets stucked in the Hard Fault handler routine. I attach the details given by the Hard Fault Analyzer in STM32CubeIDE. I have read these issues are mostly related to the heap and stack sizes, but I have considerably increased their size and the problem persists. I would appreciate any help on how to solve this issue?Thanks in advance. Regards.

screen MB1166-a09,FP-AI-VISION1_V3.1.0,camera mb1683-CAMS-A02, the screen is a mess.No camera picture.I'm using cubeide for run( flash the borad),should i change the version of FP-AI-VISION1? where i goes wrong？

Hello:After a basic understanding of X-CUBE-AI, I am not sure if I have a sufficient understanding of this project. I seem to have found that the generated network. c and related model files do not particularly rely on the HAL library and MCU, so I tried to port them to other companies' MCUs. I copied all the header files and. c files related to the model, but encountered some issues as follows. I would like to know if this action is feasible.Thanks�?

Hi, I have followed the tutorial https://www.youtube.com/watch?v=grgNXdkmzzQ&t=4s to test X-CUBE-AI with CubeMX5.0.1, but the step 'Analysis' just end with an error message without any details. I have received the same error with CubeIDE which comes with integrated CubeMX. If I try to run 'Analysis' of the same model after selecting the board I receive the error as shown below. Could anyone please help me in figuring out where could be the problem? Do I require to install Python to run X-Cube-AI? That is not mentioned in the users manual or tutorial.Thank you in advance!AchyutThe error:--------------Analyzing model Neural Network Tools for STM32 v1.1.0 (AI tools v4.1.0) TOOL ERROR: Traceback (most recent call&nb

Hi, I have a problem validating my classifier on xcube ai (both validation on target and validation on desktop). It shows error scores but accuracy is n.a. persistently. I have tried setting "--classifier" option in settings, restarting cubeIDE, nothing works. It shows that option --classifier is used during compilation but still, n.a. is showed under acc upon finishing validation. My model is a binary classifer model with sigmoid function as output, producing either 0 or 1. Does anyone know how to fix this issue? Accuracy score is crutial for me to see, as I want to compare the accuracy score on test set on quantized model in my thesis.Also, is there any way to show binary crossentropy (or any crossentropy error) among the default scores in validation? It would be useful to see.

I'm trying to run simple validation on target in XCubeIAI on my model. Upon building I get the following error: On my other pc with the same project and IDE version the validation is succesfull. Any ideas?

Hello, I am new to deployment of neural networks on mcu's. I am doing a speech enhancment project and want to deploy my noise classificator on my nucleo L496ZGP board. I have 0 exprience with this, how can I do tests on my custom data and (prefferebly) classify recorded sounds in real time in xcube-ai?

When the cubeai tool is used to analyze the model, the unquantized model can be analyzed normally, but the static quantization analysis can make mistakes. I use the onnx model. The following is my quantified model. I hope someone can help me.

Hi,I am trying to simulate the behavior of different STM32 MCUs to see which unit meets my expectations for an object detection task.I was wondering if there is a way to simulate their behavior on object detection problems with tflite model + validation dataset given? If so, can you help me about it? Thanks a lot in advance!

..

I'm trying to use the benchmarking tool in the developer cloud (https://stm32ai-cs.st.com).I have a model exported in several formats (fp32, fp16, full 8 bit, etc.)The input shape is (1, 128, 384, 3).The model works fine and other tools like Edge Impulse are able to benchmark my model without issues on a variety of hardware.However, when I upload the model and try to run in developer cloud, I get the following output:>>> stm32ai validate --model EfficientNet_v1_full_integer_quant_with_int16_act.tflite --allocate-inputs --allocate-outputs --relocatable --compression none --optimization balanced --name network --workspace workspace --output output Neural Network Tools for STM32AI v1.7.0 (STM.ai v8.0.1-19409) INTERNAL ERROR: index 1 is out of bounds for axis 0 with size 1I'm not sure why the tool is over-indexing in the first dimension.Can someone point me in the right direction?

STM32H747I-EVAL  Could conectar a some kind of camera to use FP-AI-VISION1_V3.1.0? I don't see the conector to B-CAMS-OMV

When using the STM32Cube.AI online tool to benchmark chips for running a neural net the max file size is limited to 32MiB. I have a models that is 93MiB, is there a way to run a model this size? What is the reason for the size limit? Is it related to the available hardware or is there another reason?https://stm32ai-cs.st.com/home

Sorry if this question is basic, but I was not able to find an answer in the documentation. I have a very simple Keras model with one input and one output, for your reference the architecture is:model = tf.keras.Sequential()model.add(tf.keras.layers.Dense(16, activation='relu', input_shape=(1,)))model.add(tf.keras.layers.Dense(16, activation='relu'))model.add(tf.keras.layers.Dense(1))I have two vectors x_test and y_test with shape (200,), the same shape of my validation arrays used during the training phase, and I saved a .npz dataset to quantize the network as follows:np.savez('testdata.npz',x_test=x_test,y_test=y_test)I am using the STM32.AI developer cloud to quantize the network. Quantization fails, and the log displays the message:Executing with: {'model': '... Converting original model to TFLite... cannot reshape array of size 200 into shape (1,1)So apparently

The approach described on the official XCUBE-AI documentation is not working for my case.

Dear Sir/Madam, I am getting an error "Analyze fail on the AI model" when I analyze a network using "STM32Cube.AI runtime" (see the snapshot below). I just train a simple AI model (https://colab.research.google.com/drive/1HBsuTW2gGAoPgWnkpwXiXI7JHXR05lrG?usp=sharing#scrollTo=eFymXEWBwarM) and want to execute a model (sine_model.tflite) on the board. However, when I want to analyze the network, I get this error "Analyze fail on the AI model".I am using the latest version (1.12.1) of STM32CubeIDE and X-Cube-AI(8.0.1). I work on Ap