How can we help you today?

How do I convert pre-recorded sound files to CSV so that I can upload them to Qeexo AutoML?

Here is a Python script (link here) that can convert a WAV file to CSV format so that your data becomes compatible with Qeexo AutoML. Depending on the original frequency of the sound file, the script conducts re-sampling to 16kHz, which is the ODR we work with.

If your sound files are in MP3 format, we recommend using ffmpeg to convert them to WAV files first before feeding them to this python script.

Please contact support@qeexo.com if you have more questions on this.

Will the platform support other types of Arm-Cortex-M4-based hardware?

Yes. Qeexo is continuing to add support for other hardware modules. Please stay tuned for our announcement for the next hardware support.

What are the sensors currently supported on Qeexo AutoML?

Qeexo AutoML currently supports IMU sensor, accelerometer sensor, magnetometer sensor, temperature sensor, altimeter/pressure sensor, humidity sensor, IR sensor, and microphone/audio sensor.

What are the machine learning algorithms supported in Qeexo AutoML?

For Multi-class classification, Qeexo AutoML supports Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Gradient Boosting Machine (GBM), Logistic Regression (LR), Random Forest (RF), XGBoost (XGB) and Decision Tree (DT). For Single-class classification, Qeexo AutoML supports Local Outlier Factor (LOF) and Isolation Forest (IF). We are always adding support for more algorithms.

Is the device flashing configurable for a serial connection pin-out from UART?

Custom hardware support is available for Enterprise tier users. For more information, please contact marketing@qeexo.com.

Is the architecture fixed for CNN models? Is it possible to change the layers?

The architecture of CNN is configurable. Users have a choice of configuring several model parameters for CNN as well as ANN (feed-forward neural networks). Please refer to the Qeexo AutoML User Guide.

How many classes can you define at the same time?

You can train up to 30 unique classes at once. With more classes, more data is recommended to ensure good performance.

How do we ensure accuracy when datasets are small?

AutoML performs automatic sensor selection, feature selection, and hyper-parameter optimization to maximize the performance of small datasets. However, if the data does not sufficiently represent the statistical diversity needed for building good machine learning models, more data is recommended. You can see if more data will help with accuracy through the learning curves generated for each model.

How do I preprocess the sensor data? Do I need to create custom code?

Data preprocessing is done automatically on Qeexo AutoML. You don’t need to write custom code.

Does Qeexo AutoML support options for tradeoffs between accuracy and speed?

Yes. Qeexo AutoML’s machine learning pipeline supports the tradeoffs between accuracy, model size, and latency. Soon we will release an option to expose these parameters to the end users so that users will be able to set, for example, the maximum allowed latency, in exchange for lower accuracy.

Does Qeexo AutoML support microcontrollers with cameras for image classification?

Camera-image classification is supported on Qeexo AutoML Vision, which is not yet available to the general public. If you have a specific use case, please contact us at info@qeexo.com.

Does Qeexo AutoML allow for regression output as well as classification?

Regression models are currently being developed and soon will be relased in Qeexo AutoML.

Do I need to develop drivers specific to the hardware sensor modules?

Everything is integrated in the Qeexo AutoML backend for the hardware platforms that we support, so there is no need for driver development. Camera image classification will be supported on Qeexo AutoML-Vision, available at a later time.

Do I need to collect anomaly data when building an anomaly detection model?

The anomaly detection model can be created on Qeexo AutoML as a Single-class classification project. Only data from the “normal” class is needed to train the classifier and anything not recognized within a threshold from the normal class is considered an anomaly.

Can you control the size of the models so they are small enough for target hardware?

When you select the target hardware, Qeexo AutoML becomes aware of its memory constraints, so that the models developed will always fit onto the device. We also use many different kinds of techniques such as feature selection, hyperparameter optimization, and engine setting optimization to make sure models are the smallest possible.

Can Qeexo AutoML support custom hardware modules?

Yes. Qeexo provides custom hardware integration support for Enterprise tier users. For more information, please contact marketing@qeexo.com.

Can Qeexo AutoML ingest customized data?

Qeexo AutoML has built-in data collection and labeling tools and can also accept data uploads via CSVs. Users can convert their data model to Qeexo-AutoML-supported CSV data format to upload and run it through the model-building pipeline. Please refer to the Qeexo AutoML User Guide for the CSV data format. Samples are also available in Downloads from within automl.qeexo.com.

Are you sending all of the sensor data to the cloud then downloading the trained machine learning models to the Edge?

Yes, for Bronze users on AWS, sensor data is sent to the cloud for model-building, and the trained models are downloaded onto the local device. For Enterprise tier users, there is an option to set up Qeexo AutoML on a local server.

Are regularization techniques built into Qeexo AutoML?

Yes. All of the supported algorithms in Qeexo AutoML use regularization. Depending on the type of the algorithm, different regularization techniques are used.

Still need help?