MNIST and Quick Draw on STM32F429ZI

[NAghl]

Hello Everybody,

I just wanted to share with you two projects I have done with CUBEMX.AI.

ALL networks are trained with TensorFlow 2.0.

Tensoflow2.0 saved networks(Keras Format) receives an error when converting in CUBEMX so I changed some options in python code to make it compatible with CUBEMX.AI.

Links to github:

https://github.com/nimaaghli/STM32AI_MNIST

https://github.com/nimaaghli/STM32AI_QuickDraw

Link to video demonstrations:

https://www.youtube.com/watch?v=8oBEGNwsj4A

https://www.youtube.com/watch?v=_YbzxUagpbw

[hirotakaster]

I have a same problem. I think maybe CubeMX/Artificial Inteligence is not compatible to Tensorflow 2.0 yet.

Windows 10

CubeMX 5.4

Artificial Inteligence 5.0.0

TensorFlow 2.0(Keras) model is here(very simple).

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

dense_66 (Dense) (None, 40) 680

_________________________________________________________________

dense_67 (Dense) (None, 20) 820

_________________________________________________________________

dense_68 (Dense) (None, 3) 63

=================================================================

Total params: 1,563

Trainable params: 1,563

Non-trainable params: 0

_________________________________________________________________

But I got a INVALID MODEL error when I try "Analyze" on CubeMX.

[NAghl]

Hi,

Please add following to each layer. It would fix the problem.

kernel_initializer='random_uniform'

for example :

model.add(tf.keras.layers.Dense(256, activation='relu',kernel_initializer='random_uniform'))

[hirotakaster]

Hi,

Analyze works!! thank you.

[Fernando1]

Hi, I asked a question about how to make predictions. I will see your code to see if I can answer. I want to send the 28x28 image by uart, since I am using a generic plate with an STM32F4VE. Thanks!

[Fernando1]

Hello again. I was seeing the code generated by STM32Cube.AI and generated a code for verification. Is it possible to generate code for prediction? Looking at your code, I see that it is different from the one generated by STM32Cube.AI, could you explain how to generate it? And if you did it manually, could you explain it to me? Greetings.

[NAghl]

Hi,

I have actually used STM32Cube.AI to generate the files. I do prediction inside app_x-cube-ai.c (generated by CubeAI) using MX_X_CUBE_AI_Process() method. That is where I pass the image information in float values(Pixel values are normizled from 0-255 to 0-1)