Edge AI

I’m using X-Cube-AI Toolbox 9.1 to realize Vision Transformer deployment but got stuck. I exported from torch using ONNX opset=15. I then used onnx2tf to convert ViT to .tflite format, which is obviously a valid model as shown in Netron. 1.When I’m trying to analyze it in .tflite, there popped up an error says “BATCH_MATMUL not supported.”2. Without conversion, analyzing .onnx model popped up something like “tuple has no method get_shape” and failed.3. I further converted it to .keras and .h5, but .keras says: Unable to load the file, and then the whole CUBE IDE got stuck and expects a restart. Meanwhile .h5 only says something like the Length is four.All of the trials above deploying ViT model failed with only a single line of error note during model analysis. So I have the following questions:A) I tried to unroll the matmul in FOR loops, but unfortunately  LOOP operator is also not supported. B) I think one important thing is that the scaled dot product attention

Hi, I am currently working with the STM32 H747I-DISCO board, which uses the MB1166-A09 LCD daughter board. I attempted to update the driver by following the instructions from this resource: https://wiki.st.com/stm32mcu/wiki/AI:How_to_perform_people_counting_using_FP-AI-VISION1_and_STM32H747XI#Driver_update_-28required_only_when_using_MB1166_version_A09-29  Despite following the steps, I am still experiencing a noisy display when running the people counting project in STM32CubeIDE. Could someone please provide guidance on how to resolve this issue?

I am running a model which uses batch_matmul though the batch_matmul is already defined why am i getting such error.

Hello, this is a Stm32h747i-disco user.https://www.digikey.kr/en/maker/projects/tinyml-getting-started-with-stm32-x-cube-ai/f94e1c8bfc1e4b6291d0f672d780d2c0I'm trying to study using the sine model example in the link above, but x-cube-ai 5.1.2 is not confirmed.What is the reason? 

I've successfully implemented a small machine learning model on the STM32F303ZE, and it's working well so far. However, I’m curious about the maximum size of the model that this microcontroller can handle. Has anyone experimented with larger models on the STM32F303ZE? What are the limitations in terms of model size, and at what point should I consider switching to another board?

im trying to train a model based on pascal 2012 dataset (training mode). after the system laod the datasets. when it starts training 1/50 epoque   i encounter several errors about cudnn(please check image) im using cudnn v9.4 . i found that tensorflow 2.8.3 (modelzoo work with) is compatible only with cudnn v8.1 is it true. and how to install v8.1 because when i download teh v8.1 from nvidia website . i dont find the setup file to execute. 

hello,I want to train using Model Zoo with a custom dataset. I'm trying to convert the data from COCO format to Darknet format, but I encountered an error with the YAML file (see attached).

We reported the following error during the "optimize" step of generating an ONNX model: (ST Edge AI Core v1.0.0-19894  NOT IMPLEMENTED: Unsupported layer types: ScatterND).How to solve this issue? Is there any function or tool that can replace ScatterND? 

hello,I want to deploy an object detection application on the STM32H747 board. I used YOLOv5n; can I use FP-AI-Vision like in the example from the tutorial "How to use transfer learning to perform image classification on STM32".Best Regard

Hello,I have STM32H7S78-DK board and its have ARM cortex M7 core. As per research I saw that AI model is work on the same M7 cortex but I didn't know how it will work? would you please let us know STM32H7S78 supports AI model implements or not and if yes then can anyone provide steps to how we can implements on the board. Thanks.

I was trying to reconfigure and customize the HMI given by HSDATALOG2 interface , i have been looking into utilities/HSDPython_SDK/examples ; utilities/HSDPython_SDK/st_dtdl_gui but i can't see how i can manage the controll over the widgets colours and texts . i' ve consulted the scripts written in python , JSON files but nothing changes while executing the main . is there a file that contains all the widgets , how may i make my own changements.       @SimonePradolini @Federica Bossi 

Hello everyone,I am learning how to use the ST Edge AI Developer Cloud tool and, after downloading my model that I developed with TensorFlow and selecting the STM32 MCUs platform, I wanted to try the post-training quantization but I encountered the following error:Executing with: {'model': '/tmp/quantization-service/970d7e77-a0e3-4d23-a8ec-5e7e52d9a260/1dcnnrnn_01.h5', 'data': '/tmp/quantization-service/970d7e77-a0e3-4d23-a8ec-5e7e52d9a260/dataset4quantization.npz', 'input_type': tf.float32, 'output_type': tf.float32, 'optimization': <Optimize.DEFAULT: 'DEFAULT'>, 'output': '/tmp/quantization-service/970d7e77-a0e3-4d23-a8ec-5e7e52d9a260', 'disable_per_channel': False} Error when deserializing class 'InputLayer' using config={'batch_shape': [None, 9999, 1], 'dtype': 'float32', 'sparse': False, 'name': 'input_layer_4'}. Exception encountered: Unrecognized keyword arguments: ['batch_shape']The error occurs regardless of weather I provide a data set for quantization or not.Can anyone

I want to implement air gesture recognition using accelerometer and gyroscope data. With different movements you should be able to make settings on the handheld device. Movements to be recognized are, for example, tipping twice or moving up and down twice. My findings so far:The first attempt was made with LSM6DSO and it's Finite State Machine (FSM) by using Unico tool. Basically it works, but the reliability is too low. The second attempt was made with LSM6DSOX and it's Machine Learning Core (MLC) by using Unico tool. Reliability is too low in this case as well.The third attempt was made with Qeexo AutoML using the SensorTile Box MCU for processing the sensor data. Similar to this description. https://community.st.com/t5/mems-and-sensors/how-to-detect-hand-air-gestures-with-the-sensortile-box-and/ta-p/49547The result was relatively good. However, just one tipping triggers the double tipping event. My biggest problem is, that I don't want to

Hello,I am using the model from 'STM32 AI Model Zoo' located at 'stm32ai-modelzoo/image_classification/pretrained_models/mobilenetv2/ST_pretrainedmodel_public_dataset/flowers/mobilenet_v2_0.35_128_tl' for image classification. I have installed and tested the Device Application using the Template Application in X-CUBE-AI 9.0.0.I verified that the input data is correctly received, and I understand that the output should be quantized integer values according to the quantization results. However, I have found that all the output values in the first tensor of the model's data_outs are 0.- 1. Can this happen if I don't use the stm32h747i board?- 2. Did I misunderstand and write the post process function incorrectly?Thanks 

When I use multiple networks in one program, I get the following errorAI platform (API 1.1.0 - RUNTIME 8.1.0)Discovering the network(s)...Found network "network_fe"Creating the network "network_fe"..Initializing the networkE: AI error (ai_mnetwork_init) - type=0x30 code=0x13 

Hello,Please, I would like to know if Edge Impulse is compatible with the STM32H747-DISCO, and more specifically, if the FOMO model can be deployed in this board.Best regards

Hello,I would like to run a ML model which I had created using tflite on the STM32H7 series. I have installed x-cube-ai 8.1. Are there any tutorials or manuals where I can understand the functions to be used.There are a few videos but those are the old ones and I am getting errors while using those functions / modules.Can anyone please help me with this?-RegardsHrishikesh

Hello,I want to implement an application for detecting items in the kitchen on an STM32-H747i-DISCO board. I installed Python 3.10.2, TensorFlow 2.8, CUDA 11.7, and cuDNN v8.8.1. After installing the model zoo , I tried running the command in  Command Prompt python stm32ai_main.py , from this path " ..\stm32ai-modelzoo\object_detection\src".  But I encountered the attached errorI

I want to run Cube AI Model zoo examples on STM32H745. The current example in Model zoo is for STM32H747 with a display, I can replace the display and print the output on UART.  Is there an example or tutorial that explains the changes required to run AI on STM32H745? I would appreciate a detailed answer with a guide, as I am new to STM MCUs. Development kits: NUCLEO-144 STM32H745ZI EVAL BRD

im trying to run a model that distuingish pasta types using stm32h747i and camera . i followed this guide  https://wiki.stmicroelectronics.cn/stm32mcu/index.php?title=AI:How_to_use_Teachable_Machine_to_create_an_image_classification_application_on_STM32&sfr=stm32mcu to be short , i followed every step and the model worked .  but when u tried to run the project the next day the camera didnt work properly .  i tried other projects and the camera work fine  its just in this project any help? i should mention that i didnt follow the part where i need   to change the flipmirror function because its mot clear.  but the camera worked properly the first time without doing that part .       

It seems like a used ram size fit to External SDRAM in STM32H757i- EVAL Board.However, I can see the error when doing analyze.Also, I cannot see any TFLite Micro runtime generated code.I can see empty function.I was totally fine with STM32CUBE AI runtime.  

How to figure out which model can be deployed on the board, according to the available RAM on the board.The ST-yolo-LC model from the ST model zoo seems to fit the board's RAM availability,(i) is it possible to train this model for my own usecase i.e fall detection and implement the same?(ii) And if I retrain the model for my own use, can I still use the application as it is from the AI function pack examples.(iii) What model is st-yolo-LC, is it a particular of yolo models or a custom model?

When I used validation app I could see like  notes: - the output of the reference model is used as ground truth/reference value         - 10 samples (5 items per sample)   acc=100.00%, rmse=0.000000147, mae=0.000000066, l2r=0.000000430, nse=1.000, cos=1.000   5 classes (10 samples)   ---------------------------------   C0        0    .    .    .    .   C1        .    0    .    .    .   C2        .    .  &nb

Hi,I'm encountering an issue while porting a trained TensorFlow-Lite model, obtained from a research paper, onto a NUCLEO-F401RE board. The process goes smoothly when analyzing and validating the code on my desktop, generating the code, and downloading it to the board. However, when I attempt to perform a "validate of target," I receive an "Invalid firmware" error (screenshot attached). The model is neural network based on U-Net architecture with shallow structure, single-channel input, and output.Board used: NUCLEO-F401RESTM32MX : Version 6.12.0STM32CubeIDE : Version: 1.16.0Could you please provide guidance on how to resolve this issue? Any suggestions for modifications or troubleshooting steps would be greatly appreciated.Thank you!

I wanted to share some feedback regarding the use of ModelZoo, as you previously recommended. Unfortunately, after dedicating significant time—about a month—to working with it, I have encountered numerous challenges that have made it difficult to achieve the intended outcomes.The errors and bugs I've faced have been persistent, and despite consulting with various ST employees and even the ST community, I have not been able to resolve them. It appears that the structure of the ModelZoo folder differs significantly from what is presented in the tutorials, rendering the tutorials largely unhelpful. Additionally, the README files provided do not offer the necessary guidance to overcome these issues.One particular area of concern is the YAML files, which do not align with those shown in the tutorial. This discrepancy has led to repeated errors, including proxy errors, that are not addressed or explained in the existing documentation.Given these difficulties, I believe that ModelZoo requires