Edge AI

 

I have a TFLite that I can successfully analyze (targeting NUCLEO-L4R5ZI-P board) with X-CUBE-AI. I then quantize the TFLite model using the "full integer quantization" method from the docs. However, when I analyze this quantized model, I'm met with "TOOL ERROR: cannot reshape array". I validate using random numbers. I don't understand the problem as the quantized version runs absolutely fine on my laptop (with the exact same input + dataloader as the un-quantized version).I'd be really grateful for any help with this - thanks!  

Hii! I'm trying to implement a tensorflow ML model in NUCLEOF401RE Board using XCube.AI and stm32CubeMX. when i download the XCube.AI package and try to add the zipper file in CubeMX im facing an error that "file cannot be deflated". how to solve this error. The zipper file is installed in proper location ""C:\Users\Lakshana\STM32Cube\Repository\Packs\STMicroelectronics\X-CUBE-AI\8.1.0\stm32ai-windows-8.1.0.zip" please help me resolve this error 

I have a LSTM model which was quantized using TensorFlow dynamic, full and float16 quantization. After that, I wanted to benchmark these three quantized models on the STM32AI model zoo. But after importing these models into the web UI and pressing "Start", I get the following error message: >>> stm32ai analyze --model dummy_model_0_stateful_int.tflite --allocate-inputs --allocate-outputs --compression none --optimization balanced --series stm32f4 --target stm32f4 --name network --workspace workspace --output output Neural Network Tools for STM32 family v1.7.0 (stm.ai v8.1.0-19520) NOT IMPLEMENTED: Unsupported layer types: READ_VARIABLE, VAR_HANDLE, ASSIGN_VARIABLE, CALL_ONCE, WHILE, stopping. I would be really grateful for ideas or insides on the problem :) 

Hello everyone,I'm currently facing an issue while working with the LSM6DSOX accelerometer.In my setup, the LSM6DSOX accelerometer is not consistently triggering an interrupt on INT1 when the specified threshold is exceeded. Code, connections, and power supply: everything appears to be in order. I am trying to implement a wake-up feature similar to the example in AN5273 but with a different threshold (1.0g +/- 0.1g). The FSM operates at an ODR of 104 Hz, and the accelerometer shares the same ODR with a full-scale range of 4g. I used the Unico tool to generate the .ufc file for configuration, imported it into my code, and also experimented with the UNICLEO tool and an eval-board for DIP24 on a different setup, encountering the same issue.Config:FSM ODR: 104 HzAccelerometer ODR: 104 HzGyro OFFFull-Scale: 4gThreshold: 1.0g +/- 0.1gFSM as per wakeup example in application note

In the STM32Cube.AI workshop - 6 Neural Network (NN) Model creation - YouTube, tutacousticscenes2016 does not work. librosa returns an error message: ModuleNotFoundError: No module named '_soundfile_data'.  Please suggest a workaround.

Hello, I have a question regarding the evaluation parameters for decision tree results. As you can see in the picture below, the Mean Absolute Error and Root Mean Squared Error are high, but the accuracy is also high. I am confused because I thought that low error and high accuracy are supposed to go hand in hand. Can you please clarify this for me?  Kindly,Fehmi

I am curious about these parameters,I don`t konw where they are used.These are auto generated by following the official guide, but if I want to build a project with x-cube-ai and CubeMX my own and still use the model from model zoo, how shoul i deal with these two par

Hi STM Community-Context: X-Nucleo-53L8A1 / Nucleo-L476RG16 zones per line of distance measurements are being captured by NanoEdgeAI in an attempt to generate an anomaly detection AI model. 16 axes correlate to 16 measurement zones. Measurements are in mm.  The data seems to be properly formatted; however, when the "Signal" is added, NanoEdgeAI indicates there are not sufficient values for FFT to be performed. MIN, MAX, & MEANS are calculated.Files attached. Pls advise.Thanks in advance,Kelly 

Which software is best to do tiny ML and other algorithm for ISPU ?How can we implement our model in to it ?How can we create our model ?

The following is the code generated by X-CUBE-AI of STM32CubeIDE ,but I can't understand it. /** Array declarations section **********************************************/ /* Array#0 */ AI_ARRAY_OBJ_DECLARE( input_1_output_array, AI_ARRAY_FORMAT_FLOAT|AI_FMT_FLAG_IS_IO, NULL, NULL, 4096, AI_STATIC) /** Tensor declarations section *********************************************/ /* Tensor #0 */ AI_TENSOR_OBJ_DECLARE( input_1_output, AI_STATIC, 0, 0x0, AI_SHAPE_INIT(4, 1, 1, 64, 64), AI_STRIDE_INIT(4, 4, 4, 4, 256), 1, &input_1_output_array, NULL)    DER 

Hi!I have generated an n-Class Classification static library with NanoEdge AI Studio using data from ISM330ISN. I want to use this library in the ISPU template project that comes with X-CUBE-ISPU (X-CUBE-ISPU\1.1.0\Ispu\ism330is_lsm6dso16is\template\ispu). I have pasted the .h files (NanoEdgeAI.h and knowledge.h) and the static library (libneai.a) in the ispu_utils folder and have linked the static library by going to Properties > C/C++ Build > Settings > Tool Settings > GNU STRED Cross C Linker > Libraries and inserting the following information:But after hitting "Apply", "Rebuild Ind

Hi, im using the FP-SNS-DATALOG1 firmware for measuring sensor data and another custom firmware for other ML tasks. Furthermore, i am implementing a bootloader to switch between both softwares.  I want to measure automaticaly and in order to do so i need to trigger the sampling via software (not with pushing the USR button). How can i implement that? I compiled the Project file from your github repository and i would now try to create a new binary (.elf) to use as firmware on my evaluation board. Is there an easier/quicker way?

Hi everyone, I am working on an extrapolation task as part of a research project at the university and I compare the different models, in this case random forest models. I assume that the model complexity of the model is related to the size of the knowledge array. How can it be that the execution speed of the models on the microcontroller does not increase with increasing model size? Are calculations performed in parallel? For example, models with a size of 200 kB can be executed much faster than models with a size of 80 kB. I work with STM32 Nucleo-144 development board with STM32H723ZG MCU.Thanks in advance.

 Hi,I am currently working on my Final Year Project, which involves implementing a plant disease detection model on the STM32H747i-Disco board. My project is inspired by an image classification example from the model zoo.Back in October 2023, I successfully deployed the sample model (specifically, the provided flower model) on the STM32H747i-Disco board, using the B-CAMS-OMV camera. However, I have recently encountered an issue where I am unable to replicate this deployment.Although the CMD window indicates that the model is deployed successfully (figure 1), the LCD screen on the board remains blank (figure 2). This issue persists despite the fact that there are no error warnings in the CMD window.I am aware that there have been updates to the LCD screen of the board. My LCD screen is the A03 model, which, to my understanding, should not be affected by these updates. In an effort to troubleshoot, I followed the steps provided in

I am using the STEVAL-STWINKT1B evaluation board to gather microphone data with a 192kHz sampling rate. The data is saved in a .dat file which i want to read and process. However, when reading the IMP23ABSU_MIC.dat file i get completely different results. Using a freeware hex editor i opened the .dat file and this is how it looks: As the samples are saved in a 16-bit integer (according to the mic's datasheet), which corresponds to 2 bytes, im expecting to read 0x31FF. This is valid since the HSDatalog software also presents the same value in int16 (e.g. -207). This is what i have tried so far in STM32C

Hello everyone, I have a pytorch model (mymodel.pth) obtained by running the mixed-precision quantization of this algorithm: * https://github.com/eml-eda/q-ppg. I followed the instructions in the readme section.Python models are not natively supported by X-CUBE-AI so I convert mymodel.pth to ONNX.But if I open mymodel.onnx with Netron, I can see that each convolutional layer has 2 inputs (see image) and STM32CubeIDE doesnt' support this.  * Reading the paper, I know that the quantized model obtained by Q-PPG has been deployed on STM32WB55. So how can I deploy the mo

Hi, I have a self-trained image classification model with two classes with dynamic range quantization(unquantized model is about 1.5mb and it is 500kb after quantization). And this model takes too much flash and ram after benchmarking. Is there any method for reducing the flash and ram consumption? The model I am using is mobilenet, should I use other model for image classification? Or is there any other model for recommendation?Also note that STM32 developer cloud does not support quantization for my self-trained model so that I quantize the model by myself(using dynamic range quantization). Thank you!

Hi, I am trying to create an STM32Cube.AI application using X-Cube-AI. I am following the documentation under Content 6. I trained my own model and configure the .ioc file as shown below. I wanted to get input from the ADC and feed directly into the model to perform classification. I am using the latest X-Cube-AI version 8.1.0. Under my X-Cube-AI folder, this is what I got:

Hi folks,I gently ask you some help related to a computer vision application, based in artificial inteligence, that I have been developed at STM32H747i-DISCO.  I am developing my code following the steps presented in a stm tutorial that is available in this link : https://www.youtube.com/watch?v=NDshmSH7WnA . The code that I am using is the same presented in the tutorial and available to download from STM website, in the following link: 

Hello, I have implemented https://wiki.st.com/stm32mcu/wiki/AI:How_to_perform_motion_sensing_on_STM32L4_IoTnode this project. When I run the program I had a hard fault error when function enters  ai_platform_network_process().  I have incremented the heap and stack sizes. I don't know what else I can do to run properly. As below, after ai_network_run() it enters ai_platform_network_process(). This function is source coded and cannot see the content of this function. If you can help me, I will be very grateful...Regards. static void AI_Run(float *pIn, float *pOut){ai_i32

Hi, I am currently using SensorTile.box Pro with FP-SNS-DATALOG2 FW. After acquisition, I would need to convert the data from the microphone from .dat to .wav and to do this I am using the hsdatalog_to_wav.py example provided. Unfortunately, the resulting .wav seems to be empty. Duration is correct but no sound is present. Attached a picture of the .dat file converted to .csv (phases with no sound and with sound are visible) and the .wav file. Thank you for the support

Hi, I am using NanoEdgeAIStudio to detect anomalies on microphone signal. I am particularly interested in logging ultrasounds (from 20kHz to 100kHz). I used the Datalogger tool (in NanoEdgeAIStudio) to set up the datalogger and generate the neai_datalog.bin file. In this tool, I have noticed the only microphone available is "MP23ABS1" (see the screenshot below). I have read MP23ABS1's doc, though it is not clear to me if it is able to log ultrasounds ? Can you clarify this pls ? Also, I have read STWINKT1B's doc and I know there is also microphone I

The deploy.py script in stm32ai-modelzoo/image_classification/getting_started generates an extra file, ai_model_config.h which is convenient as it contains the class names and other usefull defines.Is there a way to generate this file from cubeMX?Moreover, at the end of this config.h file, there are some defines which are related to the specific board (a nucleo H743ZI2 in my case). Where do these come from and how can someone modify them?For example I would like to use an OV7670 camera in place of the one supported in the example.