Edge AI

What I have checked through the user manual UM2526 and FP-AI-VISION1 examples is that I am using X-cube-ai version 6.0.0 for the 32bit floating point model (the model data type I am currently using), so I want to read data from external SDRAM in order to reduce inference time. I know that it is optimized so that the inference time is faster than reading from the internal flash memory. Currently, the model I trained is Keras 2.3.0, which is not compatible with X-cube-ai 5.0.0.Purpose : To reduce the inference time, I want to save the initial weight&bias table to the external Q-SPI flash memory and move it to SDRAM as in the example. Question 1 : The inference time is faster when it is placed in the external memory when the data type is float, but in the table12, internal flash memory shows faster results , I don’t understand why. (reference : UM2611 manual)Q2 : In general, what clock delay is the read time latency of the internal flash memory?Q3 : If I use external SDRAM for re

Hello STM Community,I am currently trying to upload a large model to my STM32H745I-DISCO. Given the size of the model, I chose the [Use external flash] option, which, as per the user manual, led to the generation of a separate network_data.bin file. As per my understanding, this network_data.bin file should contain the weight files that were split, although I invite corrections if my assumption is incorrect.The user manual further emphasizes that the network_data.bin file must be manually programmed on the board's external Flash memory. Tools such as the STM32CubeProgrammer (STM32CubeProg) are recommended for this task. An interesting note I came across was that using automatic validation on the target leads to the network_data.bin file being automatically programmed on the external Flash memory on the board.Given this information, I'm left with some uncertainty: if I don't manually code to explicitly utilize this network_data.bin file, will the model still be able to run inferenc

 Hi STM Community,I'm currently leveraging an STM32H745I-DISCO to run our custom DNN model, which is designed for a 4-class classification task and accepts multivariate time-series data of the shape (3, 2048). While I've successfully analyzed and validated the model with random input data on the target platform, I'm finding it challenging to upload and run inference using our own custom data (over 100 instances).Considering my limited experience in embedded engineering, I'm in need of detailed guidance on how to accomplish the following:Data Uploading: What are the steps and the preferred format to upload our custom data onto the STM32 platform?Data Implementation: How can I integrate this data into the main script for the model to run inference?Data Reshaping: The first reference below seems to illustrate the use of 1-dimensional data, while our model uses 2-dimensional data. Could you provide advice on reshaping the data appropriately for our model?Here are a few resources I hav

Hello! Im using STM32Cube IDE version 1.12.1.Im having isssue with including the header file "ai_platform.h " and "ai_datatypes_defines.h".I have installed X-cube-AI extensions  and all other required extension of Artificial Intelligence.I am getting the error message of "no such file or directory" .I request you to look into the issue.Thank you !

Good Morning , I am working on the following board:  https://www.st.com/en/evaluation-tools/stm32wb5mm-dk.htmlI need to activate Unicleo-GUI and MLC core on the board.  Do you know how to get it ?Thank you very much. Best regardsLuigi Bibbo'PS: my email is luigi.bibbo@unifi.it

How to deal with fatal error: ai_platform.h: No such file or directory? I can't find this file anywhere in the project.I tried using different versions of CubeIDE and X-CUBE-AI, added include paths in project properties. I checked on F7 and H7. I have Core and ApplicationTemplate selected in Software Packs Component Selector.The problem started with the inability to use USART along with the X-CUBE-AI package enabled. Then no error appeared, only communication over USART did not work. I am currently using CubeIDE 1.12 and X-CUBE-AI 8.0, and have updated what I can. The last time AI worked for me I was still using Eclipse (SW4STM32), but it no longer supports the latest chips I use.Please point out what could be the problem.

If i want to send the result of (Food-Recognition)of FP-AI-VISION1_V3.1.0,to achieve to object of  to communicate some event that the network recognizes (for example the detection of a certain object) and send it out of the panel.The complete project is shown: development board + external board (raspberry or arduino or whatever).Hay some software o method o project to achieve this?

Hi I have a really simple Neural Network (Keras) with 3 float32 inputs and 3 float32 outputs. Size: 2.2KB RAM 16kBFlash and 1600MACC, really small network, in my opinion it should not be a problem at all to run that.For testing i wanted to try the simplest way possible to run that Network on a STML4 MCU.I used CubeMX, integraded the Model and created and built an example Project by using a "ApplicationTemplate".All went prefect. so know i have an example Project where everything is generated. especially a file called "app_x_cube-ai.c".There is a init function called: "MX_X_CUBE_AI_Init()" and a RunAI function called " MX_X_CUBE_AI_Process", looking like this void MX_X_CUBE_AI_Process(void){    /* USER CODE BEGIN 6 */  int res = -1;   printf("TEMPLATE - run - main loop\r\n");   if (network) {   /* 1 - acquire and pre-process input data */  res = acquire_and_process_data(data_ins);  /* 2 - process the data -

Hello everyone!At the moment I am developing an AI system for the STM32H7A3 using STM32 Model Zoo und Cube AI. For testing purposes I trained the SSD Mobilenet V1 model for 192x192 picture resolution and implemented it in a firmware app for testing interference. For training I followed the tutorial in github.The output of an interference is of the shape 3830x29. Where can I find out what that means and how can I calculate detected class, probabilty and boxes of detection?If you need further informations just ask.Anybody can help me? The info in the github just says "Coming soon".

I'm trying to do a program like a baby crying and i'm looking for the code of babycrying program. I need to do a program that detects unwanted noises in cars, so i think that maybe is similar than the babycrying app.I thought it was possible using the expert mode, but now I think it is not possible with it. If someone could help me pls, maybe with some information or the babycrying code I would appreciate it.

Hello everybody,I started to use the AI Cube MX extension and developed an application on the NUCLEO-F746ZG Board. It all worked very well.Now I am trying to integrate this small application into a bigger, existing project.There seems to be a problem with the Init function of the AI, it causes a reset after a few seconds.There are a  few possible problems I identified:- the project was configured with a CubeMX Version the AI extension is not compatible with- the clock configurations of the two project parts seem to clash (HSE_ON vs. HSE_BYPASS)Does anyone have experience with that?Does anyone know any clock configuration etc. that the AI extension is compatible with? I cannot find much in the documentation...  

Good morning, I am working on this Firmware https://github.com/stm32-hotspot/STM32WB-BLE-AI-MotionSense.I created the Neural Network but I don't see the Classification on The Monitor.Can you help me to solve the problem ?Thank you.Best regardsLuigi Bibbo' PS: my e-mail is   luigi.bibbo@unifi.it 

Hi,I have managed to program the STM32F746G-DISCO with a simple neural network but when I try to do it using the external Flash by using the option in advanced settings (External NOR Flash) it doesn't work. Any idea why? Is there anything else I need to do?I'm using STM32CubeMX and Keil.Thanks!

I want to know when will stm cube ai support keras 2.12.0 because for now it only support up till 2.11.0. I can't seem to downgrade my keras version as it require to downgrade more dependencies.Furthermore I converted my lstm keras (h5) model to tflite but it gives me "transpose of batch size not supported" error on analyze any solution?

Hi, I have two onnx models. One of them is U-Net like architecture and another is just Conv1d + Linear. I have onnx file and try to analyze the models.When I analyze U-Net like model, the analyzer show "INTERNAL ERROR: tuple index out of range" error.Also, when I analyze Conv1d+Linear, the analyzer show "NOT IMPLEMENTED: Order of dimensions of input cannot be interpreted".  I am using 8.0.1 version.

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: When I run the model batch is returned as 1, and the error type is 17 and code 16, which means: AI_ERROR_INVALID_VALUE        = 0x17, AI_ERROR_CODE_ARRAY        = 0x0016I can't find any documentation on the error codes. I also can't seem to fix the problem as everything else works OK.Is

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？