Edge AI

..

Hello,I'm a newcomer in the STM32 machine learning world and I tried to make a AI model work on my STM32F429I-DISCO. Unfortunately the "Validate on target" option does not work... Errors are E200 and E801. Here is my thought process :Iinstall a ftdi to my board to send data trough serial port : Then I launch stm32cube mx, select the board STM32F429I-DISC1 and start project without initialize all peripheral with their default mode. Then :And finally : I also tried validate on target on my other COM (COM4) and with communication port on the target : USART 1.Each time I get the same result : I don't know what to do, can someone know the solution of my problem ?Thank in advance !Best regards,Martin

Hi, I created my own model and saved it as .h5.However, when I analysed I got the error below. E010(InvalidModelError): Couldn't load Keras model D:\motion.h5,E010(InvalidModelError): Model saved with Keras 2.9.0 but <= 2.6.0 is supportedDo you have any ideas on how I could handle this situation?Thanks,

Hi STM community,Is anyone here familiar with adding the CMSIS-NN functions in a source code in STM32CubeIDE ?If you are, could you please point me to or provide me with the proper steps to do so ?As I am facing compile error "undefined reference to ..." when I attempt to use these functions e.g. This code is configured to run in the NUCLEO-F446RE board in the STM32CubeIDEa prompt reply is much appreciatedThanks

The "validate on target" option doesn't work and I was wondering if it was possible to work around the problem by sending some data manually. I thanks you in advance.

where can I find more detalis for this project, which sensor? I have VL53L0X. it is OK? or VL53L5CX is better? what is configuration of tof sensor for data logging?

Hello, the question may be stupid but everywhere I go I see people using the STM32H747 for the different examples in the FP-AI-VISION1 that can be found here. But I only have a STM32F429I-DISCO that is compatible for the X-CUBE-AI pack.The part that interests me is the MobileNetv2_Model.I thank you in advance for helping me !

Hello,I have created an Artificial intelligence project to be implemented on Nucleo-144 (stm32h745) board. The stm32cubeide created a set of related source and header files all grouped in the X-CUBE-AI folder.I defined variables in my main.c following an online tutorial. But when I run build I received the error bolded below indicating something wrong in one of the X-CUBE-AI folder created by the software.The error is as follows: "../X-CUBE-AI/App/gru_network.h:42:3: error: invalid initializer". This error is due to lines 41 and 59 in gru_network.h, as you can see in the attached figures. Note that gru_network.h is a header file auto-generated by the stm32cubeideWhat would be the issue?Thank you.

Hi, it seems to be a common error in STM32 devices. It happens when I try to validate on target after analyze the model. I tried to fix it according to the answers in community. But those solutions given in community or other websites doesn't help. At least for my board they doesn't help. Infomation:  The device: STM32H735G-DK MCU: STM32H735IGK6 CubeMX 6.5.0 with X Cube AI v7.1.0, v7.0.0 & v 6.0.0 Model: The model given in official tutorial of X-cube-ai.   CN15 STLK √ JP7: CHGR/USBFS SW1: Set to 1 USART3 Enabled: PD8 & PD9 Baud Rate: 115200 Word Length:8 bit No Parity Stop bits: 1   ST-Link connectivity works normally. (Able to connect with CubeProgrammer). It can be built in Keil and could be load to the board, with LD6 blinking in red and green. I am looking forward to solutions to this problem. Thanks 

I have a model file(.h5) and Analyze is failing with keras(.h5) model at STMCubeMX and throwing INTERNAL ERROR: The layer has never been called and thus has no defined input shape.Need help in understanding the issue and resolution for the same.My model is attached here for referenceThanks in Advance

I have an AI model that detects the human activity and saved the model in .h5 format and converted it into .tflite image. When I upload the image into STMCubeMX and analyze the model, analyze is failing and showing as NOT IMPLEMENTED: Transpose of batch not supported.Need help with the resolution for above issue. Thanks in Advance.

Hi, am using STM32H747- Discovery board. FP-AI-VISION1 function pack which is already running in that board. Now I want use that board in PicoVoice. Am not able to connect the device in Picovoice. Am not able to upload the firmware. I tried installed the St-link upgrade firmware but still am not able to connect the device. Like for sensortile box, there is firmware for board recovery. Is there any firmware to recovery the STM32H747-DISCO board?Please help.Regards,Shivashankari

E200(ValidationError): stm32: Unable to bind the STM AI runtime with "network" c-model: []  connection to "serial:115200"/"115200" run-time fails  E801(HwIOError): Invalid firmware - COM4:115200I have already tried the following:Using CubeMX and IAR IDEUninstalling CubeMXUsing older versions of STM IDEUsing older versions of X-CUBE AIConfirmed that I am using M7Checked clocksChecked ports for any errors thereUsing other networksDirectly hooking up to D0 and D1 for USART6ResetsMany combinations of the aboveIn general I work with: STM32F746G-DiscoSTM32CubeIDE version 1.9.0X-cube AI version 7.1.0LeNet-5 for my networkKeras for my training and suchIf anyone has any insights into the issue, that would be much appreciated. Attached is the .ioc file.

We have created an LSTM model using Keras, and decreased its size to use it on our STM32F746G-Disco board. Our model can actually process dynamic size records. This means that, we do not need fix the size of the record. We could realize it on Google Colab platform. We however encountered the error "not implemented: dynamic input dimensions not supported: shape is ('batch: none', 'ch: 16', 'h: none') (channel last)", the last channel varies because record lengths are different. How can we deal with this? Is it possible to pass dynamic input to LSTM model?

..

or any hardware at all ? like stm32l4 dev board ?

Hi all,I would like to ask if there is a suggested procedure to move the network weights, the one generated inside network_data.h, to external memory-mapped QSPI FLASH. I've already moved both input/output data and activation to external SRAM, but I'm facing with a compile time error when I do the same with weights (initializer element is not constant).Before start changing things to solve it, maybe someone of you can help!I've analysed the example FP-AI-VISION1, but it seems to move only input/output data and activations, not weights.Thank you and kind regards.

Hi,I have selected the right target board in NanoedgeAI studio and compiled the library and header files to be included in the STM32CubeIDE project. I am also using the right board: STWINKT1B. Project builds and debugs successfully, but every call from the Nanoedge API returns "NEAI_BOARD_ERROR". For example, this error was observed when calling "neai_anomalydetection_init", "neai_anomalydetection_learn", and "neai_anomalydetection_detect". Any idea how I can troubleshoot this?Thanks.

..

I am looking for a development board that can run a deep learning model with a CSI interfaced CMOS camera.I think if chips can support DCMI(digital camera interface), it means it is available for CSI interfaced CMOS cameras. I hope my idea is right. Therefore, I found that many chips support DCMI such as F7 series, F2 series, part of F4 series, L4 series, and H7 series.Like this board "NUCLEO-L432KC" which can seem like can run TunyML and supports DCMI.https://www.digikey.com/en/products/detail/stmicroelectronics/NUCLEO-L432KC/6132763?WT.mc_id=frommaker.ioWithin those chips supporting DCMI, is there any other recommended development board for users to apply deep learning applications and CMOS cameras?Thanks

/* USER CODE BEGIN PV */ static ai_buffer *ai_input; static ai_buffer *ai_output; /* USER CODE END PV */   int main(void){ ... // Chunk of memory used to hold the intermediate values for neural network AI_ALIGNED(32) ai_u32 activations[AI_NETWORK_DATA_ACTIVATIONS_SIZE];   // Buffers used to store input and output tensors AI_ALIGNED(32) ai_i32 in_data[AI_NETWORK_IN_1_SIZE_BYTES]; AI_ALIGNED(32) ai_i32 out_data[AI_NETWORK_OUT_1_SIZE_BYTES];   // Pointer to our model ai_handle my_model = AI_HANDLE_NULL;   // structs that point to our input and output data and some meta data ai_input = ai_network_inputs_get(my_model , NULL); ai_output = ai_network_outputs_get(my_model , NULL);   .... }As seen the function ai_network_inputs_get(ai_handle network, ai_u16 *n_buffer) resides in the network.c file and it makes a call to the ai_platform_inputs_get(network, n_buffer); in file ai_platform_interface.hIt doesnt give me any errors while running but the code breaks w

 Is it needed to do the learning on every device? Or can the library be used on all devices of the same class?

I experienced NEAI Studio license activation issues, can you please help ?

I am using your NanoEdge AI Studio Software for classification and I am looking at the FFTs displayed in the signals segment after collecting data.How, specifically are these FFTs generated from the raw data ? Also, what the units on the axis are as I am unsure what unit 'fs' is ?

I am trying out a code example from Nano Edge AI studio and want to download and run that on the STWINKIT1B. I figured one way would be to use the HSDatalog example and add a serial port to it on the STM32CubeIDE but something simpler would be fine as well. I just want to be able to read the accelerometer output put that in the data to the buffer for the Nano Edge library, get inferences and display on serial port. Searched around for some simple startup examples but could not find one, so hoping some one could point in the right direction. Thanks cn