Getting started with MEMS Studio

macOS platforms

To install MEMS Studio, simply open the DMG package and drag MEMS Studio to your Applications folder.

Quick Setup page

The Quick Setup page allows the user to configure basic sensor configurations such as the output data rate, full scale, and others according to the device datasheet. The content of the page is adjusted according to the connected evaluation board, sensors, and firmware.

If more evaluation modes are available like sensor fusion low-power, FIFO, TDM, and others, they can be selected on this page. The available sensor evaluation pages are adjusted according to the selected mode. If applicable, the control can align its status to the current register value.

The Easy Configuration button configures the sensor to a predefined configuration to stream sensor data. The Reset Device Configuration button restores the default sensor configuration. The Reset Timestamp button resets the timestamp value to 0 and clears all application data buffers, which means it will also clear all the charts.

Registers Map page

The Registers Map displays the list of device registers divided by register pages, if any.

Every register item has a Read, Write, and Default(restore default value) button enabled according to the datasheet and a bit view populated according to the register value.

Using the dedicated button ?, any register item can be expanded to explore the register value bit map description.

Custom Register Action allows direct read and write transactions to a single register by specifying the register address, page and value.

Save to File page

The user can use this page to save a stream of sensor output data in a .csv file for postprocessing. This can be done by selecting the data to be stored.

The Browse button is used to select the folder and insert the file name. The acquisition period can be set to a fixed time in [ms] if desired. Then data to be saved can be selected from the list. The acquisition can be controlled using the Start and Stop buttons. If the Logging timeout option is used, the acquisition is going to run as long as the timeout period. A label can be assigned to the log using the Set label option.

If the sensors are configured according to different output data rates, the user can select the Data log period source from among the available data types using a dedicated selection option. Refer to the following figure.

If the Timestamp is selected as the datalog period source, every new data from any sensor will create a new line in the log file. For other sensors that do not have new data at that moment, the line is filled with previous data. A new data flag is used to distinguish which data points are new in that particular line. The new data flag is a bit field with following meaning.

Table 6. New data flag values

Sensor	Bit	Value
Pressure Sensor	1	0x001
Temperature Sensor	2	0x002
Humidity Sensor	3	0x004
Accelerometer	4	0x008
Gyroscope	5	0x010
Magnetometer	6	0x020
Interrupt	7	0x040
MLC	8	0x080
FSM	9	0x100
ISPU	10	0x200
VAFE/QVAR	11	0x400
Accelerometer	12	0x800

When using Datalog2 (High Speed Datalog) firmware, data are primarily stored in Datalog2 native binary format. Using the Convert to CSV switch, binary format can be automatically converted to CSV format at the end of the logging session. After conversion to a CSV file, an additional column called data_changed is added to indicate changes. This value is a bit field that shows which data in the current row is new. The first bit (LSB) indicates new data in the first sensor data column, the second bit indicates new data in the second sensor data column, and so on.

Data Table tool

The Data Table tool displays the output values measured by the sensor connected to the evaluation board. This view provides an overview of all samples received with a table data update rate of 0.5 Hz to reduce overall CPU/ GPU consumption. This functionality is disabled in the case of very high data rates.

Data Monitor tool

The Data Monitor tool displays the latest output values measured by the sensor connected to the evaluation board. This view provides an overview of the last samples received with an update rate of 10 Hz to reduce overall CPU/GPU consumption in the case of high data rates.

MLC Monitor, FSM Monitor, ISPU Monitor

The MLC Monitor, FSM Monitor, and ISPU Monitor pages can be used to display output from the machine learning core (MLC) or finite state machine (FSM) including the label assigned to each output value and output from the intelligent sensor processing unit (ISPU). This feature provides a clear and user-friendly overview of the current states and classifications detected by the sensors, making it easier to monitor and interpret the results in real time. The textual values are taken from the JSON file with the sensor configuration, which is loaded from the Load/Save Configuration or Example Configurations pages.

Bar Charts tool

The Bar Charts tool displays the data measured by the sensor in a bar chart format.

The height of each bar is determined by the amplitude of the signal measured by the sensor along the corresponding axis. The full scale of the graph depends on the configuration and can be changed through either the Quick Setup or the Registers Map tab.

Line Charts tool

The Line Charts tool shows the evolution of the output over time. This tool generates a time graph of data samples from the connected sensors.

Each waveform can be shown or hidden by clicking on the corresponding colored box in the top bar of the plot.

The user can click on the minus and plus at the bottom of the graph object to manually scale the plot horizontally. Furthermore, using a dedicated scroll bar at the bottom of the page, the user can display the older samples.

Hovering over the right side of a graph object, available options and settings are displayed such as:

• Fit: updates the Y-axis scale to fit all waveforms
• Auto: automatic update of the Y-axis scale
• Full: sets the full-scale value according to the device configuration
• Save: exports a screenshot of the chart to an image file in .png format
• Copy: saves the chart as an image in the system clipboard
• Help: displays the available keyboard shortcuts for navigation in the chart

Scatter Plot tool

The Scatter Plot tool shows the graphical representation of the magnetometer data using Cartesian coordinates. In general, a scatter plot is used to display values for typically two variable sets of data.

The plot shows three lines according to the sensor axes (X-Y, Y-Z, X-Z). The three lines can be enabled and disabled independently by clicking on the corresponding colored box at the top of the graph. By clicking on the Clear button, the user can reset all data in the plot.

Compass tool

The Compass tool shows an example of a compass application, which can be implemented using a 6-axis module (3-axis accelerometer and 3-axis magnetometer).

The algorithm uses the magnetometer data to measure the Earth’s magnetic field, and the accelerometer data to compensate for the inclination of the board. Rotating the board, the application shows the heading of the compass.

The performance of the compass is related to the configuration used. So, the application shows the current configuration and the recommended configuration (accelerometer and magnetometer ODR).

Before using the compass demo, the system must be calibrated by moving the board randomly for a few seconds; the quality of the calibration step is indicated by the compass object border color according to the legend on the right side.

Interrupt tool

The Interrupt tool allows evaluating the interrupt generation features of the MEMS sensor.

On this page, it is possible to configure the characteristics of the inertial events that must be recognized by the device and to visualize (in real time) the evolution of the interrupt lines.

The application provides access to the interrupt registers, which allow the configuration of the interrupt sources of the device.

The user can configure the interrupts using the dedicated combo box in order to load the recommended configurations in the device.

Inclinometer tool

The Inclinometer tool represents the angle between the accelerometer axis and the horizontal plane. This tool is available if the sensor in use integrates an accelerometer.

FFT tool

The FFT tool displays the fast Fourier transform of the sensor data. The page shows the frequency-domain plot for each axis of the selected sensor.

The tool provides various options such as:

• DC nulling: removes the DC component from the signal
• Magnitude: calculates Euler’s formula of the vector composed of the axis components
• Magnitude (scale): Y-axis scale
• Frequency (scale): X-axis scale
• Style: style of chart (waveform or bars)
• Window: window function applied to the signal before FFT evaluation
• Length: length of the window used to evaluate FFT (number of samples from which FFT is computed)
• Zero padding: number of zero samples added to the captured signal
• Show spectrogram: enables or disables spectrogram visualization
• Fixed refresh rate: updates the FFT chart without overlapping windows
• Fixed overlap: updates the FFT chart, overlapping windows according to the slider value of the window overlap
• Slider value: percentage of the samples saved for the next FFT computation

6D tool

The 6D tool provides an example of how to use the 6D position function.

On this page, it is possible to configure the interrupt for 6D recognition by clicking on the 6D Enabled toggle button.

After loading the configuration, the selected 3D model is oriented depending on the 6D position detected.

The window also shows (in real time) the evolution of the interrupt line.

The user can change the register configuration by setting different values in the interrupt registers as shown in the tool.

3D Plot tool

The 3D Plot tool shows the graphical representation of the magnetometer data in 3D space.

Zooming in and out of the chart is possible using the mouse wheel. Changing the view angle is possible by clicking and panning in the chart.

Features Demo page

The Features Demo page allows the user to evaluate demo modes available for a connected sensor when this feature is supported. The side menu on the right displays indicators to notify the user about event detection and set parameters, if applicable.

Communication and Power Settings tool

The Comm. & Power Settings tool allows the user to select a communication interface like I²C, SPI 3-wire, SPI 4-wire, I3C if available, read the current consumption, set the VDD and VDDIO values and select between Interrupt and Polling Data Reading Mode. This functionalities are available only on the Professional MEMS tool board (STEVAL-MKI109D, STEVAL-MKI109V3).

FIFO tool

The FIFO tool can be used to test the first-in first-out data buffer embedded in the device when this feature is supported by the sensor (see the device datasheet for more details).

By using the combo box available on the page, the FIFO can be configured in all the supported modes (for example, bypass, FIFO, stream, stream-to-FIFO).

The application shows the values of the X, Y, Z data stored in the deep FIFO buffer, indicating both numerical data and the corresponding graph.

Sensor fusion tool

Some sensors can be equipped with the sensor fusion low-power (SFLP) feature for generating game rotation vector, gravity vector, and gyroscope bias data based on the accelerometer and gyroscope data processing. If this feature is available, the evaluation can be activated on the Quick Setup page.

If sensor fusion mode is selected, the available sensor evaluation pages are adjusted according to this mode. Game rotation vector, quaternions, gravity vector, and gyroscope bias signals are added in Data Table, Data Monitor, Bar Charts, Line Charts and Save to File. The 3D Model page is available in this mode. A 3D model is displayed on this page and the orientation of the model changes based on the game rotation vector data so the model follows the sensor movement. Different 3D models can be selected. The initial position of the model can be set by using the Reset Model button.

Load/Save Configuration page

The Load/Save Configuration page allows the user to save and load the current register configuration. Clicking on the Save button stores the register values in a .json file. The saved configuration can be loaded at any time by clicking on the Load button.

Example Configurations page

The Example Configurations page is available for sensors equipped with FSM, MLC, or ISPU features. This page downloads content directly from the STMicroelectronics GitHub repository, providing an up-to-date list of example configurations for embedded features. Users can select a configuration and upload it directly to the sensor for immediate testing.

Save To File

The user can use this page to save a stream of sensor and library output data in a .csv file. This can be done by selecting the data to be stored. The Browse button is used to select the folder and insert the file name, then the Start and Stop buttons define the acquisition period. Additionally, the user can define a recording timeout using the dedicated toggle button and text field to interrupt data logging after a defined [ms] interval.

Data Table

The Data Table tool displays the output values from the sensors and libraries. This view provides an overview of all values received with a table data update rate of 0.5 Hz to reduce overall CPU/GPU consumption.

Data Monitor

The Data Monitor tool displays the latest output values from the sensors and libraries. This view provides an overview of the last samples received with an update rate of 10 Hz to reduce overall CPU/ GPU consumption.

Data visualization pages

The available data visualization pages depend on the selected library. Line charts, bar charts, and a scatter plot (for magnetometer data) are always available to visualize sensor data. Additional pages for visualization of library outputs like 3D plot, 3D model, status indicator and table, and others are displayed based on the library functionality.

Time Statistics

The Time Statistics page displays the library elapsed time in line charts and statistical information line average, min, max values.

Data Injection

The Data Injection page allows injecting previously acquired sensor data into the library input instead of real-time sensor data. The library processes the data and sends the results in the same way as working with real-time sensor data. All the other features for the evaluation of the library are available and active.

The data log file with previously acquired data can be opened by clicking on the Browse button. The data must be acquired with the minimum sampling rate frequency given by the library and the library evaluation firmware. The data log with the higher sampling rate is automatically decimated. Input data are displayed in the table.

Data injection is enabled by switching the firmware to Offline mode. Data injection is controlled by the Start and Stop buttons. Injection line by line can be done using the Step button. The active line is highlighted in the table. A wraparound to move from the last sample to the first sample can be activated using the Repeat toggle switch.

FSM parameters

• Sensor selector: displays the selected device to configure
• State machine selector: displays the selected state machine to configure. This selection is applied to both the configuration page and debug page if available.
• FSM ODR: this setting determines the processing / output data rate of the FSM samples.
• Converter: this tool helps to generate the values to be set in the threshold resources in the Variable Data section. It converts a float32 value into a float16 format and vice versa.
• FSM Start Address: determines the first address written in the FSM registers while configuring the state machines

Configuration

This tab lets the user add commands and conditions to the selected state machine and visualize diagrams while editing state parameters. A dedicated tool bar illustrates the actions that can be performed.

The right side of the configuration tab shows the active parameters and provides the capability to edit them in order to update the values of the resources used in the Variable Data section. This section shows the active and enabled resources, depending on the list of instructions that have been set by the user in the FSM state diagram on the left side.

The FSM configuration tool bar provides the capability to perform actions on all device state machines or on a single state machine. Some actions are available only if the target device is connected, such as writing or reading data to/from the connected device, while others might work in a different way such as loading an FSM configuration.

Loading an FSM configuration decodes configuration data on a host pc if no target device is connected or writes a configuration on the device and then decodes device data if the target device is connected to the MEMS Studio application in order to allow the user to inspect configuration data, ensuring the best reliability of data shown in the GUI.

• Load all state machines from file: loads multiple state machines from .json file and populates the GUI
• Save all state machine as: exports state machines configuration to a dedicated .json file
• Load device configuration from file: imports all state machine configurations, writes the configuration to the device, and updates the GUI according to the written instructions
• Save device configuration as: exports the instructions of the device configuration for configuring all state machines to a .json file
• Reset state machine: clears the selected state machine GUI data without changing the device configuration
• Read FSM configuration from sensor: updates all state machine GUI data according to the current FSM device configuration
• Write FSM configuration to sensor: writes all current state machine GUI data to the device
• Load state machine: loads one state machine from .fsm (legacy) or .json file and populates the GUI
• Save current FSM: exports the current state machine configuration to a .json file
• Reset all state machines: clears all state machine GUI data
• Read state machine from examples: displays the available FSM examples for the selected device with their descriptions and allows the user to load a configuration to the GUI

Testing

This tab shows the available sensor data compatible with the FSM configuration, interrupt data plots, and FSM output register values. A dedicated blinking LED graphic is toggled every time an FSM interrupt is detected during the parsing of the FSM status register data and once it is activated, its state is not changed for ~300 msec.

This page helps the user to check the program functionalities after the FSM configuration is completed.

Debug

The Debug function allows debugging the selected state machine sample by sample after loading an input data pattern.

The main purpose of this tab is to check the state machine functionality in order to validate the program using recorded log files, assuming the device is configured as it was before starting to log data in the file.

During data injection, the Program Pointer and the Reset Pointer are updated according to the data that are read from the device.

The Program Pointer value contains the current address of the program and it is used to highlight the corresponding state in the diagram. The Reset Pointer value contains the address of the state or condition related to the CONTREL command or return condition of the program.

The tool provides the user with several data injection interactive controls to inject one or multiple samples at a time and write output data to a dedicated table.

Every command before a condition and every condition has a radio button to configure a break point for pausing the data injection process if the program pointer address is equal to the address of the item with the enabled break point indicator. Whenever a break point is activated, the background color of the current state changes according to the selected color theme of the application.

Data injection controls

• Run: injects all the next samples
• Pause: pauses the data injection process
• Stop: interrupts the data injection process and exits debug mode.
• Next: injects the next sample and updates the table with output data
• Next step: injects a defined number of samples according to the Step lengthsettings
• Export log: exports the content of the table with the results to a .csv file

The data table is updated at every step if the Results at each step switch is enabled, otherwise it is updated only when the data injection is paused or stopped. The Read OUTS registers switch enables or disables reading sensor output registers. The Read INT registers switch enables or disables reading interrupts status registers. Detect INT indicates the number of detected interrupts in the processed log file.

Data patterns

The Data patterns tab allows managing the data patterns to be used for the machine learning processing. The data patterns that are possible to load must have the same data format as the log files generated by MEMS Studio, Unico-GUI, or Unicleo-GUI.

A data pattern(s) is selected using the Browse button. Multiple files can be selected. A class label must be assigned to each data pattern, which is then used for machine learning processing. The class label is confirmed by the Load button. Data patterns can be removed from the list using the trash bin button in the tables.

AFS tool

The AFS tool offers the possibility to estimate suitable window length, filters, and features for classification done by the decision tree.

Automatic window length calculation

The window length is calculated based on the lowest frequency present in the signal across all data classes.

Automatic filter selection

There are two methods that can be used for automatic filter selection. Basic search is a peak-based method that identifies the maxima of the signal in the frequency spectrum. This method is less computationally expensive, but it is not accurate on flat signals without significant peaks. Exhaustive search is an entropy-based method that searches all combinations of frequency bands (on a logarithmic scale) and selects the frequency of interest that minimizes the entropy against the other classes.

Automatic feature selection

The automatic feature selection algorithm considers filters generated by the automatic filter selection and the features are computed for different filters and raw data. It uses an ensemble method for feature importance and cross correlation to remove highly correlated features. The following statistical features are supported for both single axis and norm: mean, variance, energy, peak to peak, minimum, maximum. The automatic feature selection algorithm selects the top features by computing the rank of each feature according to an ensemble method and averaging across all ensemble methods. There are five different ensemble methods available to determine feature importance: ANOVA, Ada Boost (ADA), Sequential Feature Selection (SFS), Random Forest (RF), Recursive Feature Elimination (RFE). By default, ANOVA, Ada Boost, Random Forest, and Recursive Feature Elimination are used as methods for ranking important features. Including Sequential Feature Selection may improve the overall performance but increases the overall runtime.

Data analysis takes into account the machine learning core ODR and input specification parameters from the ARFF generation tab. The window length is also taken from this tab in case the window length calculation is not enabled.

After the desired parameter selection, the data analysis is executed using the Run button.

The generated window length, filter, and features can be transferred to the MLC configuration in the ARFF generation tab using the Apply to Settings button.

ARFF generation

Attribute-Relation File Format (ARFF) is a file format used for storing data in machine learning applications. It contains a list of instances, where each instance represents a single window processed by the MLC.

The ARFF generation tab allows configuring sensor and MLC parameters. The MLC processes the data at the Machine Learning Core ODR rate in consequential data windows. The number of samples in each window can be configured by the Window length option. The MLC processes the sensor data specified in the Inputs section. The full scale and output data rate of the selected sensors can be defined. Filters offer preprocessing on the machine learning core input signals. The Features section contains a list of features that are calculated from the input data (filtered or not filtered). After setting all the parameters, the ARFF file can be generated using the Generate ARFF File button. Statistical parameters computed from the inputs are then used for classification in the decision tree.

Decision tree generation

The Decision tree generation tab allows configuring a decision tree using selected classes to process and evaluating decision tree classification performance. The Number of decision trees can be specified.

A new decision tree algorithm was introduced in MEMS Studio v2.0.0. The new algorithm has both depth-first and best-first tree building implementations, implements Breiman's cost-complexity pruning, and identifies optimal tree size using cross validation.

The maximum depth of the tree (Max depth), minimum samples required in a leaf node (Min leaf samples), and Percentage of data to use for training can be configured.

The previous algorithm can be enabled using the Use legacy decision tree train algo switch. In this case, the Maximum number of nodes and Confidence factor, which control decision tree pruning, can be set.

A decision tree is generated using the Generate Decision Tree button. A file with an externally generated decision tree definition can be loaded using the Import button. After decision tree generation or loading, statistical parameters of the decision tree like accuracy, number of instances, confusion matrix, and others are displayed. The resulting decision tree can be analyzed using the Inspect tree info button for a textual content or using Show Decision Tree for a graphical representation of the same.

If the user imports a decision tree, a feature mapping is needed to map every feature selected during arff generation with feature listed into decision tree file every feature selected during arff generation must be mapped, listing each feature in the decision tree file. Note that this mapping is applied for every decision tree imported or generated.

Config generation

The Config generation tab allows configuring the metaclassifiers and generating the .json file containing the machine learning core configuration.

The generated MLC configuration can be stored also in .h file using the Export As button.

In the case of a board with a selected sensor connected, the generated configuration can be uploaded to the sensor using the Load Config into Sensor button.

Decision tree output viewer

The Decision tree output viewer tab allows testing the generated MLC configuration by monitoring individual MLC outputs in the chart.

Debug

The Debug tab allows checking the MLC functionality in order to validate the configuration using previously recorded log files, assuming the device is configured as it was before starting to log data in the file.

Data injection control

• Run: injects all the next samples
• Pause: pauses the data injection process with the possibility to resume it
• Stop: interrupts the data injection process and exits debug mode.
• Next: injects the next sample and updates the table with output data
• Next step: injects a defined number of samples according to the Step lengthsettings
• Export log: exports the content of the table with the results to a .csv file

The data table is updated at every step if the Results at each step switch is enabled, otherwise it is updated only when the data injection is paused or stopped. The Number of decision tree set settings specifies which MLC output registers will be monitored.

The debug page provides the user with a dedicated converter tool to simplify data conversion during log data processing. The tool can be opened using the Show Convertor button. The debug tool works with raw data in LSB and it converts sensor data internally into LSB using the sensitivity evaluated on the basis of a given configuration file if no LSB data was given.

The MLC in the hardware configures conditions in the units of measurement, which is why users might need a converter tool to easily understand if a certain condition of a decision tree is verified for a certain LSB data. For example if a decision tree condition has a threshold of 0.5 g, the corresponding LSB threshold value will be 1024 using a sensitivity of 0.488.

Configuration

The Configuration tab allows fast-configuring the pedometer. In this window, it is possible to visualize in real time the evolution of both the accelerometer signal and the two interrupt lines and to read the pedometer step count output.

The user can enable and configure the pedometer with its default configuration using a dedicated button.

The GUI also allows directly accessing the pedometer-related registers in order to set a user-defined pedometer configuration.

Debug

The Debug tab is used to load a data pattern into the device and run the pedometer logic on the pattern itself (offline postprocessing).

On the right side of the GUI, a three-button toolbar is available for loading a data pattern in the GUI. Pedometer configuration can be loaded from a .ucf or .json file. It is also possible to fast-configure the pedometer in its default configuration by clicking on the Pedometer Easy Configuration button.

The device enters debug mode right after the data pattern has been loaded.

On the top of the GUI, a toolbar is available to control the data injection, which can be applied with the maximum level of flexibility:

• Run: injects all samples
• Pause: pauses the data injection process with the possibility to resume it
• Stop: interrupts the data injection process and exits debug mode
• Next: injects the next sample and updates the table with the output data
• Next step: injects a defined number of samples according to the Step lengthsettings
• Export log: exports the content of the table with the results to a .csv file

The data table is updated at every step if the Results at each step switch is enabled, otherwise it is updated only when the data injection is paused or stopped.

Once the debug session is completed, another debug session can be started (the data pattern must be reloaded).

Regression Tool

The Regression Tool tab allows finding an optimal configuration on the basis of a predefined dataset (in this context, a dataset is a collection of data patterns with a reference number of steps for each pattern).

On the right side of the GUI, the initialization view contains the two initialization parameters:

• Error type: the error to be minimized by the tool (that is, mean, mean plus standard deviation, mean plus two times the standard deviation)
• Complexity level: the level of depth (in terms of iterations) of the regression. The execution time of a low-complexity level regression is lower than a high-complexity level, but the parameter space is less deeply explored.

At the top of the GUI, the run view contains the textbox for the input data path and the output file path.

Using dedicated Start and Stop buttons, the user can run the regression. A progress bar is available in order to notify the user about the progress of the regression.

Once the regression has been completed, the tool generates a folder under the output folder, named [data]_[hour]_Pedometer, containing two files:

• pedometer.ucf, containing the optimal pedometer configuration generated by the tool
• regression_tool.config, containing some metainformation used by the Regression Tool itself and statistical data about the pedometer performance on the input dataset

Under the output folder, another regression_tool.config file is generated. This file is automatically used by the Regression Tool in order to start a new regression analysis from the optimal configuration found in the latest regression executed. If the user's intention is to run a new regression from scratch (starting from the default pedometer configuration), the user should click on the Clear Regression History button.

Load NN Model / Generate page

To start configuring the ISPU Model Converter, the Workspace directory and the Output directory must be selected using the corresponding Browse buttons.

All GUI settings set during neural network processing can be restored to the default values by clicking on the Clear Settings button.

By clicking on the Terminal button, accessible from all pages, the user is able to inspect additional information related to the execution of commands requested in the available pages.

A drag and drop box allows importing NN files with supported formats such as Keras, ONNX, and TFLite (.h5, .onnx, .tflite).

By clicking on the [Display NN graph] button, the user is able to see visual graph of the chosen neural network model. To be able to access this function, the user needs to install the Netron application on the device and add it to the MEMS Studio tool path in the “External Tools” page.

The Model history section stores files previously uploaded by the user, displaying their directory path and the date of upload. Users can reselect files if they still exist in the same directory by clicking the Select button or delete them from the list by clicking on Delete. Deleting a file from the history only removes it from the graphical view, not from the container directory.

Once the workspace directory, output directory, and file to process are set, the Run Generate button becomes active. Users are able to start the generation of reports, .c and .h files by clicking on this button. Once the process is completed, the Inspect Report button becomes active, allowing users to view the textual report generated.

An additional button Open Output Directory is shown to navigate to the output directory.

Analyze page

Once the workspace directory, output directory, and file to process are set, the Run Analyze button becomes active and the user is able to start the analysis of the neural network model. Once the process is completed, the Inspect Report button becomes active, allowing users to view the textual report generated.

Users can view a summary of the report with the following data representation:

• Table: displays structured data extracted from the Analyze report, providing a detailed overview of key information

• Radar Chart: shows data metrics in a radar chart format, offering a graphical representation of different parameters.

Considering (as an example) the LSM6DSO16IS sensor, the list of maximum values for radar chart parameters is as follows:

• Total data ram: 8 KiB (activation + input + output size)
• Activation: 8 KiB
• Input: 8 KiB
• Output: 8 KiB
• Total code ram: 32 KiB (weight + code)
• Weight: 32 KiB

The percentage of total ram used by the model is divided into two parts: data ram (8 KiB), which is the sum of activation, input and output size, and code ram (32 KiB), which includes code size and weight size.

Validate page

On the Validate page, the user is able to run the original and converted models and compare the results by clicking on the Run Validate button. The converted model runs on the host PC and on the Target by connecting the sensor and choosing the mode to run On Host/ On Target. Once the process is completed, the Inspect Report button becomes active, allowing users to view the textual report generated. The processed data results are available in three different representations:

• Table: displays real values extracted from the Validate report, providing detailed information about various neural network metrics and parameters
• Radar Chart: displays normalized values in a radar chart format, offering a graphical representation of the key parameter.
  Note: All the parameters in the radar chart are normalized in order to present data in a range [0, 1]. They are not on the same scale, so, to be able to show them in a radar chart, they are normalized according to their maximum value.

Converter models have more similar results to the original model if the parameters are close to the maximum.

• Confusion Matrix: if available in the report, displays the performance of a classification model, showing true positive, true negative, false positive, and false negative values, assuming the original model as the reference
• Select the output: this combo box appears if the model contains multiple outputs, then the user is able see the result in a table based on the selected output

To be able to "Run Validate" on the Target, after selecting the mode On Target, the user needs to import the UCF file. In order to generate the ISPU UCF file  using MEMS Studio, the ISPU-toolchain path should be set in the “External Tools” page and a valid validation template should be selected. MEMS Studio embeds a template and link to download it. Once the Generate Validate UCF request is completed, a new .ucf file is created in the output directory in the subdirectory ispu_validation_out /make/bin/ispu.ucf

By default, the validation of the selected neural network is executed using random data. The user can select custom input and output to quantize using custom data.

Benchmark page

On the Benchmark page, the user can compare different neural network models through numerical and visual results. This functionality allows for the comparison of the same model in different validation modes (On Host / On Target).

A drag and drop box allows importing neural network report files with the supported format as text (.txt).

By clicking on the Export results button, the user is able to extract the results of the entire benchmark table as a .csv file.

By clicking on the Show Chart button, the user is able to see multiple layers and compare the parameter visually.

The radar chart viewer displays normalized values in a radar chart format, offering a graphical representation of the key parameter in multiple layers of the selected model.