Intelligence at the Edge
26
What if a pre-trained model isn't available? ST also offers
various function packs that include application-oriented code
examples around important use cases, such as computer
vision, sensing, and condition monitoring. For example,
FP-AI-VISION1 is the function pack featuring examples of
computer vision applications based on convolutional neural
networks (CNN), which run on the STM32H747I-DISCO
Discovery Kit connected to the B-CAMS-OMV camera
module bundle. The application examples provided in this
function pack include food recognition, person presence
detection, and people counting.
Using person presence detection as an example,
there are two projects inside the "..\FP-AI-VISION1_
Vx.x.x\Projects\STM32H747I-DISCO\Applications\
PersonDetection\" folder; one is based on MobileNet v1,
and the other is based on MobileNet v2. Each project can
be opened with the user's favorite IDE, STM32CubeIDE,
IAR EWARM, or Keil MDK to explore and modify the
code, as well as to build and run the project on the target
device. There are also several NN models included in
this function pack, both 32-bit floating-point model and
8-bit quantized model, which are built on the open-source
MobileNets. They can be found inside the "..\FP-AI-
VISION1_Vx.x.x\Utilities\AI_resources\" folder.
If a pre-trained model exists and an engineer wants
to develop their own application, they can install
STM32CubeMX from st.com and the X-CUBE-AI expansion
package from STM32CubeMX's embedded software
packages manager. They first select their target STM32
using the MCU selector tool that includes a filter for AI,
which analyzes the NN model to identify the best candidate
MCU choices. After selecting the target device, one creates
the project and enables the X-CUBE-AI component for the
project. The current version of X-CUBE-AI tool can generate
three kinds of projects: System Performance projects that
run on the STM32 MCU and accurately measure the NN
inference CPU load and memory usage; Validation projects
that validate incrementally the results returned by the NN,
stimulated with random or user test data on the desktop
PC and/or STM32 MCU embedded environment; and
Application Template projects that support the building of
AI-based applications.
It is important to ensure there is no accuracy loss after the
optimization and conversion. By choosing the validation
project application, for example, all the AI-related settings
will be automatically configured, including the USART
peripheral that is used for the virtual COM port. Inside the
X-CUBE-AI configuration interface, the network model can
be added and analyzed. For the current version, Keras
and TensorFlow Lite models are natively supported, and
other frameworks are supported through ONNX format.
Users can also choose between the STM32Cube.AI
runtime or TensorFlow Lite Micro runtime.
After performing the analysis, the tool lists the model's
complexity with the number of MACC, and the usage
of both internal and external Flash and RAM memory.
It can also show the topology of the model and the
graph of memory usage. If the memory footprint is too
large, a built-in compression function is provided inside
the tool, which can compress the weight and bias by a
factor of 4 or 8 using a weight-sharing-based algorithm.
The advantage of this approach is to have a quick
compression process, but the result is
not lossless, and the global accuracy
can be impacted.
The validation process helps quantify
the loss. To do this, two execution
modes are provided. The "validation
on desktop" mode runs on the host
PC and allows the comparison of the
original NN model with its generated
X86 C model (this mode is also
available for other applications); while
the "validation on target" mode runs
on the target device and compares
the NN model with
the C model. It
requires a validation
STM32Cube function pack with
AI application for Computer Vision
