Issue with AI Model Validation on STM32H747I-DISCO

[Dresult]

Hi everyone,

I'm trying to perform a validation test on an STM32H747I-DISCO, but I haven't been able to get it working.

Here's what I'm doing:

1. On the ST Edge AI Developer Cloud, I select the model "CNN2D_ST_HandPosture_8classes_hand_posture_ST_VL53L5CX_handposture_dataset.h5", then proceed to quantize and optimize it.
2. I run a benchmark and then download the project for STM32CubeMX 6.14.
3. I open the project and, in the Software Packs, I select "Device Application" for the M7 core and enable the "Validation" option under STMicroelectronics.X-CUBE-AI. I then save the project.
4. Next, I open the generated project in STM32CubeIDE 1.18, compile the code for the M7 and M4 cores in Release mode, and program the STM board. So far, everything works fine. When I press the reset button I can see 

#
# AI Validation 7.1
#
Compiled with GCC 13.3.1
STM32 device configuration...
 Device       : DevID:0x0450 (STM32H743/53/50xx and STM32H745/55/47/57xx) RevID:0x2003
 Core Arch.   : M7 - FPU  used
 HAL version  : 0x010b0500
 SYSCLK clock : 400 MHz
 HCLK clock   : 200 MHz
 FLASH conf.  : ACR=0x00000032 - latency=2
 CACHE conf.  : $I/$D=(True,True)
 Timestamp    : SysTick + DWT (delay(1)=1.002 ms)

ST.AI RT
--------------------------------------------------
 tools version   : v2.0.0
 network rt lib  : v10.0.0-fd22b7f9
   compiled with : GCC 11.3.1

Discovering the network(s)...

Found network "network"
Creating the network "network"..
Initializing the network
Network informations...
 model name         : network
 model signature    : 0x8f1b6d2aab7b9895e6460ef525012bb0
 model datetime     : 2025-03-23T17:34:47+0100
 compile datetime   : Mar 23 2025 17:35:55
 tools version      : 2.0.0
 complexity         : 8200 MACC
 c-nodes            : 4
 map_activations    : 1
  [0]  @0x20000000/1152
 map_weights        : 1
  [0]  @0x8010860/2896
 n_inputs/n_outputs : 1/1
  I[0] (1,8,8,2)128/i8 1:(0.003921,-128), @0x20000400/128
  O[0] (1,1,1,8)8/i8 1:(0.003906,-128), @0x200003E8/8

-------------------------------------------
| READY to receive a CMD from the HOST... |
-------------------------------------------

# Note: At this point, default ASCII-base terminal should be closed
# and a serial COM interface should be used
# (i.e. Python ai_runner module). Protocol version = 3.1
​

• Now, I open from the IDE the .ico file, navigate to Pinout & Configuration -> Middleware and Software Packs -> X-CUBE-AI, and try to perform a validation. However, I get the error: LOAD ERROR: STM32 - read timeout 50007.9ms/50000ms

I tried to send random data (in the shape expected from the network) using python Serial and then reading them through picocom. In that case I can see unknow characters as the bytes are converted in ASCII. So, I suppose that the network gives some output...

Any suggestions on how to fix it?

Thanks in advance!

[Julian E.]

Hello @Dresult ,

Thank you for your message.

I have seen it one or two times, but I never got to have the user model to test it out.

I don't have a H7 with me today, I will test it out tomorrow.

I'll update you when I know more

Have a good day,

Julian

[Dresult]

Hi Julian,

thanks a lot! Let me know if you need more information or if I can help in someway.

Looking forward to your update, have a good day too!

[Julian E. · Accepted solution]

Hello @Dresult ,

There is indeed a timeout error when using the .ioc from the dev cloud. It may be because of the memory  pool settings. I am still investigating this.

In the meantime, if you really need to do a validation on target, you just need to create a new CubeMX project, set everything by default, activate X Cube AI with the validation template. Generally, CubeMX ask if to set the optimal parameter after that, I generally click no. Then you import the model, click validate on target, check the option to flash the code and select STM32CubeIDE.

Have a good day,

Julian

[Dresult]

Hi Julian,

Thanks a lot for your time! I tried your solution, and now the validation works as expected. Please let me know if you find a solution for the .ioc file generated from the dev cloud :)

@Julian E. I have a question related to validation but not directly correlated to this issue...

I'm trying to deploy a model on the boards of the dev cloud, the initial layers are these:

however, when it comes to the quantization steps if I add an npz file for validation I receive this error:

[ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Invalid rank for input: input Got: 2 Expected: 3 Please fix either the inputs or the model.

the file in this case is generated with this code and resambles real inputs with a dimension for x of 1x4096x1 and y that is a 1x1.

import numpy as np

x_test = np.random.rand(1, 4096, 1).astype(np.float32)
y_test = np.random.rand(1, 1).astype(np.float32)

np.savez("test_data.npz", x_test=x_test, y_test=y_test)

It seems like if the validation tool internally performs a squeeze. Have you ever encountered this problem ?

[Julian E.]

Hello @Dresult ,

Can you send the model, I think that if you do a stedgeai generate, it is written somewhere, I would like to verify that.

Can you try to generate (batch_size,1,4096,1) and (batch_size,1,1)?

For example:

import numpy as np

x_test = np.random.rand(10,1, 4096, 1).astype(np.float32)
y_test = np.random.rand(10,1, 1).astype(np.float32)

np.savez("test_data.npz", x_test=x_test, y_test=y_test)

It is similar to this question: Solved: Error in Post-Training Quantization: 'Axis 2 is ou... - STMicroelectronics Community

I will need your model to know for sure.

Also the part about the idea that quantizing a small model can make it better may interest you

Have a good day,

Julian

[Dresult]

Thank you very much! With this input shape it works perfectly, however I don't understand, I thought the first dimension was the batch size...

[Julian E.]

I think, to test your data, the code do something like: prediction = model(X_test[i]).

So in your first code, X_test[0] would be of shape (4096,1) and not (1,4096,1).

Have a good day,

Julian