Edge AI

Hi,I'm planning on implementing a motor fault detection through vibration signals on an embedded system and I found that ST has some good support for this, but what would be the ideal microcontroller/dev-board in order to implement this?On the AI page, there's only one board listed (B-L475E-IOT01A) but I see here in this forum that people are using very different ones? Any reasons why or why not?Thanks.

I'm using STM32L476G-DISCO , running the human activity keras model on Cube MX 5.2 and Using the AI package 4.0.0 , the validation on desktop is successful but I got a "read time error" everytime i run the validation on target also the validation on target runs successfully using the AI Package 3.4.0 ...Any help? or there still bugs using the package?this is the report of the validation on the attached file.

Hello,I have a question concerning the memory footprint of a neural network.Given a pre-trained neural network topology, how can I estimate the needed memory footprint.Are there any application notes that describes this in more detail.All feedback is welcome.Best regards,Joeri

I have a problem with STM32; when i connect my FC the STM32 it isn't detected from my pc, it doesn't appear in device manager. What is the problem? I have windows 10 a OS. I tried out to install cube programmer, zadig, but STM32 isn't detected. Thanks for the support!!

I'm trying to create the KWS model found onhttps://github.com/Shahnawax/KWS-ANN-KERASand I'm trying to implement it on an STM32F412G-DISCO MCU.I've followed the lines of https://www.st.com/content/ccc/resource/technical/document/user_manual/group1/69/bb/ec/5d/78/16/43/ce/DM00570145/files/DM00570145.pdf/jcr:content/translations/

Hello,I'm using STM32Cube.AI to implement a Keras model. It comprises three parallel branches and interaction takes only at the end place where I concat the branches.This is a simplified version to show the structure of the model:Unfortunately Concatenate is not supported :(. [AI:Generator] NOT IMPLEMENTED: Unsupported cOde generation for layer 'Concat' [AI:Generator] Python generation ended [AI:Generator] Invalid networkI think there is a demand to support the Concatenate function :) and it sholdn't be a big deal to integrate it in the code generator.The only solution I have at this moment i

I have two different input data (voltage, current) and output data in excel file. I need help to training the data using neural network method and how to program this in micro-controller and tested with real-time (machinery) input datas.RegardsHemavathi

..

..

..

Hello,I used the Pruning API of Tensorflow and reduced the weights of my model. If i import the saved TensorflowLite model to the MCU, the deleted weights get initialized with zero. Is there a possibility only to use the other weights? Because I can save a lot memory if this would work.Tim

Following the steps in the video “Getting Started Video with STM32Cube.AI�?, I can see ‘har_github.h5’ which have model picture as follows:https://github.com/Shahnawax/HAR-CNN-KerasBut when I analyze my project which

Hello everyone,It's good to see a number of tutorials/implementations posted by community members which encourage beginners like me to implement AI at the microcontroller level.Is there any complete image classification example or tutorial where ST microcontroller capture images via DCMI interface and immediately classify images using the pre-trained network?If available, I sincerely request to share the link.Thanks in advance,M

Hi all.I'm trying to experiment with the stm32f7 discovery board and the MNIST dataset. I've trained the model and everything seems to be working fine and I'm running the validation app from the cubemx on the stm32.My problem is that I haven't found any documentation on the serial protocol that cubemx uses to communicate with the stm32. The console refers to a stm32com-base interface (i.e. Python stm32comm module, protocol version 1.0). This is the output of the console:# Note: At this point, default ASCII-base terminal should be closed # and a stm32com-base interface should be used # (i.e. Python stm32com module). Protocol version = 1.0Where can I find that tool? I've search with no luck.Thanks in advance.

We have a trained keras model, and we we are trying to deploy it on the stm32. We successfully loaded the model on cubeMX and can see the generated code in Keil.(attached .h and .c files).WE are having difficulty understanding what to pass as arguments to the create function below - Any suggestions would be of great help.ai_error ai_rnn2_create( ai_handle* network, const ai_buffer* network_config);thank yousiva

I am trying to validate Keras model using STM32 cubeMX on STM32f407 discovery kit. I am able to generate code and flash into board. But, after that when I tried to validate on target it gives me error as STM32com exception: Invalid payload: ack instead node. I search on internet about these but nothing is given. Also, in User manual, Nothing is given about these error.I have attached screenshot of error. Can anyone please help me to get out of this?

I am trying to use X-cube-ai in STM32CUBEIDE but when I try to configure I am getting this error " Warning: This IP has no parameters to be configured" (image attached)The "x -cube-ai" pack has been installed I have checked in "manage embedded software packages" as well in this folder \Users\***\STM32Cube\Repository\Packs\STMicroelectronics\X-CUBE-AI\4.0.0\CubeMX\plugins\ip .....jar file is availableAlso when I do the same process with STM32CUBEMX I am not getting any error.(Image attached)Pl let me know what to do?

Hello, I have a STM32F746G-DISCO which has a microcontroller with 1 Mbytes of Flash memory and 340 Kbytes of RAM. However, I would need more space for my bigger Keras/TensorFlow models. I have noticed that this Discovery board comes with 128 Mbit Quad-SPI Flash memory, as well as 64 Mbit SDRAM. Would it be possible to use the QSPI Flash memory to store the weights and the SDRAM as an extension of the 1 Mbytes RAM of the microcontroller? I have seen examples of code in the STM32Cube F7 Firmware package that makes use of the QSPI Flash memory and SDRAM, however I am not sure if those examples could be integrated with the X-CUBE-AI generated code in the way that I described, and what the implications of that would be. I know that there is already the possibility to quantize Keras weights, but I would still be interested in using those extended memory if possible.Thank you and kind regards

Good morning,I have created a project with CubeMX.AI based on an .h5 file constructed in keras following the steps shown on ST tutorialhttps://www.youtube.com/watch?v=grgNXdkmzzQ (Getting Started with STM32Cube.AI).Is it possibile to have an example showing how to exploit the generated code in order to solve a generic problem of classification for a new sample of data?Best regards

I have an LSM6DSOX on a STEVAL-MKI197V1 on a breadboard. I am able to communicate with the device (reading device ID returns the correct value, all driver calls return 0), but I am not getting an interrupt generated on either INT1 or INT2. I'm using the following to initialize the sensor. Any help that anyone could provide would be greatly appreciated. The following code is based on the ST example, read_data_interrupt.c.bool lsm6dsox_Init(lsm6dsox_ctx_t* ctx) { ctx->read_reg = platform_read; ctx->write_reg = platform_write;   uint8_t devId = 0; uint8_t rst = 0; lsm6dsox_pin_int1_route_t int1_route; int ret = 0;   // Reset device ret |= lsm6dsox_reset_set(ctx, PROPERTY_ENABLE); do { ret |= lsm6dsox_reset_get(ctx, &rst); } while (rst);   // Disable I3C interface ret |= lsm6dsox_i3c_disable_set(ctx, LSM6DSOX_I3C_DISABLE);   // Enable block update ret |= lsm6dsox_block_data_update_set(ctx, PROPERTY_ENABLE);   // Se

I converted a keras model to a quantized TFLite model to load this on my MCU. But i think Cube.ai doesnt yet support fixed Point TFLite models.Will this feature be enabled in the future?

Hello,i used the stm32ai software to quantize my MNIST model. I can load the quantized model on my STM32H743-Nucleo Board and with the Validation Software I can validate my quantized model on target.But I would like to create my own Application, so I used the Application Template.I can already run my "normal" model with floating point input-vecors.But if I want to feed my quantized net with fixed point input vector, I only get the same probabilityon every number (see picture). I think the problem is my input data type. The Application Template expects ai_i8. But this data type is to small for grey-scaled Mnist Digits

I'm running a Cifar10 Keras CNN Pre-trained model basically an image recognition model my labels are : 0 - airplane1 - automobile2 - bird3 - cat4 - deer5 - dog6 - frog7 - horse8 - ship9 - truckThe image array is stored as HWC format or Height-Width-Channel, a 32x32 RGB color image with 32 x 32 x 3 = 3072 values.how can I enter the input image in the custom data test for validation in MXCUBE (x-cube-ai package) do I need to normalize the data ? for example cat = {226,221,195,247,241,227,246,239,227,243,237,225,243,237,222,243,236,220,241,234,218,239,229,213,237...183}

Unico 9.0.0.0 / STEVAL-MKI197V1 (LSM6DS0X) / Machine Learning Core (MLC)When entering Features, how do you end? There should be a "End Feature Configuration" selection but I do not see it in version 9.0.0.0. If I just "Next" another feature selection opens.

Hello,Is there a way to give X-CUBE-AI a model with quantized-weights? Right now the tool is able to do a custom post-training quantization for a given Keras model. However, what if we have a model that already has quantized weights (which is the case when we do a quantization-aware training with Tensorflow Lite for example)? Is there any solution/workaround to make use of "quantization-aware weights"?Thank you and kind regards