Frequently Asked Questions

Developing AI models for edge devices addresses well known challenges encapsulated in the acronym BLERP:

• Bandwidth: By running algorithms on-device, you avoid sending sensor data to the cloud, reducing the level of connectivity required.
• Latency: On-device algorithms can be much more responsive, since there’s no round-trip to the cloud required.
• Economics: Reducing connectivity demands can save on both BOM and network communications costs, and removing the need for cloud-hosted AI systems can massively reduce costs.
• Reliability: The simplicity of on-device AI systems removes all of the connectivity overhead, making them inherently more reliable than those in the cloud, essential applications where safety and user experience are critical.

• Privacy: Edge AI allows you to benefit from AI without any data leaving the device, ensuring user privacy.

ML/AI models that are optimized for the edge overcome the issues of BLERP by design.

Edge Impulse isn’t designed to replace machine learning expertise. Instead, it’s designed to assist practitioners in navigating the unique challenges of working with sensor data and resource-constrained hardware. You can think of it as a toolbox of essential items that save you time and allow you to focus on applying your domain expertise. We built the Edge Impulse platform to solve well known edge AI challenges. Here are some of the steps we took with every kind of edge AI developer, engineer, and scientist in mind:

• Our digital signal processing algorithms are battle-tested and based on the latest research. They’re already optimized for high performance on any class of device. No need to rely on others to port Python feature engineering code to C++.

• We provide accurate estimates of on-device performance for models, and help you debug compatibility and performance issues. Instantly understand the performance impact of architecture changes; no need to muck around with benchmarking on a physical device or wait for your embedded team to test models.

• If you have an existing model, we’ll help you select the ideal hardware to run it. If you already know your target hardware, our target-aware hyperparameter optimization framework (EON Tuner) will help you identify the model architectures that meet your latency and memory requirements.

• Our system handles quantization and other optimizations, allowing you to compare the before and after results and decide which to apply.

• We convert any saved model format to C++, so you can benefit from vendor-specific optimizations for ultra fast performance, even if your embedded team is unfamiliar with machine learning.

• We provide embedded-specific model architectures, like FOMO (Fast Objects More Objects) and FOMO-AD (Fast Objects More Objects-Anomaly Detection), which you can’t find anywhere else. Download them as containers to make your own enhancements.

• Our market-leading Edge Optimized Neural (EON) Compiler, runs models with reduced RAM and flash usage, It supports a diverse range of neural networks from TensorFlow or PyTorch, as well as classical ML models from scikit-learn, LightGBM, or XGBoost. This ensures rapid execution of neural network models for real-time or near-real-time scenarios, all the while maintaining accuracy:

   Up to 70% less RAM usage
   Up to 40% less flash usage
   Faster inference times
   Reduced overall power consumption

Edge Impulse is designed to integrate thoughtfully with your machine learning workflow, integrating with all major tools and formats. Our Python SDK makes it easy to optimize models for the edge from wherever you prefer to work: inside Jupyter notebooks and training scripts, PyTorch, TensorFlow, or scikit-learn. You can transform a deep learning model into an edge-optimized C++ library with a single line of Python code.

As a modular toolbox, you can use whichever parts of Edge Impulse you prefer for a given project. Train a model in Python and optimize it for the edge with the Edge Impulse Python SDK, or upload your Numpy or Pandas dataset and run a target-aware hyperparameter sweep that automatically determines the best model architecture for a given hardware target—all from a Python script or notebook.

Edge Impulse does not lock you in: we use standard formats for storing data and models, and you can easily integrate with other MLOps platforms such as Weights & Biases. You can control every feature of Edge Impulse with our REST API, making it easy to automate with systems such as continuous integration as part of your ML pipeline.

If you have your own training code, you can easily wrap it into a container and host it inside Edge Impulse as a custom learning block, benefitting from our target-aware hyperparameter tuning (EON Tuner), and the generous GPU compute included with our enterprise license. You can do the same with custom data transformation and deployment code.

You can do so by following the instructions here.

Edge Impulse supports any edge AI Hardware that can run C++, and more. This includes a variety of MCU-based hardware and Linux-based hardware. For MCU-based hardware, Edge Impulse provides options like: default firmware, examples for integrating your impulse with custom firmware, integrated deployment options, profiling, and extensive hardware testing.

For Linux-based hardware, options include the Edge Impulse for Linux SDK and Command Line Interfacing (CLI) tooling, deployment options as native linux executables, Edge Impulse Model notable as .eim format, and Docker containers, Linux Inferencing SDKs, and profiling. If your hardware is not listed, you can still use Edge Impulse by uploading data in various formats (e.g., CSV, JPG, WAV) and deploying your trained model as a C++ library, which requires a C++ compiler and sufficient flash/RAM on your device.

To try out the Edge Impulse platform, follow our getting started guide.

For startups and enterprises looking to scale edge ML algorithm development from prototype to production, Edge Impulse offers an enterprise version. This includes all of the tools needed to go from data collection to model deployment. For more information, click here.

To access the enterprise version of Edge Impulse, you can try it out with the enterprise free trial.
The enterprise version of Edge Impulse offers a range of features designed to scale edge ML algorithm development from prototype to production. These features include a robust dataset builder to future-proof your data, integrations with all major cloud vendors, dedicated technical support, custom DSP and ML capabilities, and full access to the Edge Impulse APIs to automate your algorithm development.

No. However, Edge Impulse provides tools and processes to help you collect and label data, design features, and automatically train machine learning models. Once your pipeline is set up, it can be fully automated to train your model over time as new data samples are brought to your datasets.

You can build various types of machine learning models with Edge Impulse, including:

• Classification
• Anomaly Detection
• Regression
• Image or Audio Transfer Learning
• Object Detection

Get more information here

To deploy your model to an edge device using Edge Impulse, follow these steps:

Navigate to the Deployment page in your Edge Impulse project.

Choose from a variety of deployment options depending on your target device. These options include deploying as a C++ library, Arduino library, WebAssembly package, Python library, Docker container, or directly to mobile phones/computers.

For deploying as a customizable library or pre-built firmware, select your desired option, configure it if necessary, and then click Build at the bottom of the page.

If deploying to Linux-based devices, you can use Edge Impulse for Linux.

This tool allows you to collect data from any microphone or camera, use SDKs to collect new data from any sensor, and run impulses with full hardware acceleration.

To deploy using a Docker container, select the "Docker container" option within the Deployment section, which packages your signal processing, configurations, and learning blocks into a single container that exposes an HTTP inference server

To run your impulse directly on your computer or mobile phone, select the appropriate option and follow the instructions provided here.

Edge Impulse facilitates multi-impulse deployment, enabling the combination or stacking of multiple models to enhance overall performance and functionality. The deployment block for multi-impulse configurations can be accessed within Edge Impulse Studio. By leveraging this feature, users can seamlessly integrate multiple models to address complex use cases and achieve superior outcomes.

The deployment block can be found here. To add it to your organization, head to this page: Edge Impulse Studio -> Organizations -> Custom blocks -> Deployment blocks.

Please note that the script works with EON compiled projects only and anomaly detection blocks have not been tested.

Edge Impulse is SOC II Compliant. Our commitment to SOC II compliance shows we care about where your sensitive data lives. We implement controls, risk assessment processes, and policies to protect this data, and ultimately your organization and customers. We see it as the first line of cyber security defense in the model development process. Your data is precious and we intend to safeguard it so it can realize its true value in the real world.

Edge Impulse stores data in storage buckets that you link to your organization. You can use AWS S3 buckets, Google Cloud Storage, or any S3-compatible bucket. You need to provide the bucket name, region, endpoint, and access and secret keys to connect your bucket. A green dot indicates a successful connection.

You can import data into your Edge Impulse project by following these steps:

• Collect/import data: You can collect data directly from your sensors or import existing datasets. Edge Impulse provides tools for collecting data from various sensors, such as recording audio, capturing images, or reading sensor values. You can also import data from storage buckets by linking your organization with AWS S3 buckets, Google Cloud Storage, or any S3-compatible bucket.
• Use your mobile phone: You can use your mobile phone to import image data into Edge Impulse. After connecting your phone to your project, you can start capturing images directly through a special UI for collecting images quickly.
• Use the OpenMV Cam H7 Plus: To import data collected with the OpenMV Cam H7 Plus, you can set up your environment, collect images, and then send the dataset to Edge Impulse. This involves using the OpenMV IDE to capture images and then exporting the dataset to your Edge Impulse project.
• Upload data directly: If you have a dataset collected through a different method, you can upload the data directly to Edge Impulse. This can be done visually through the studio or via the CLI. It's recommended to use square images for image datasets.

‍

• JPG, PNG images
• MP4, AVI video files
• WAV audio files
• CBOR/JSON files in the Edge Impulse data acquisition format
• CSV files

Learn more here

To manage datasets in Edge Impulse, follow these steps:

• Link Storage Buckets: First, link your organization with one or several storage buckets. Choose from AWS S3 buckets, Google Cloud Storage, or any S3-compatible bucket. Fill in the form with your bucket name, region, endpoint, access, and secret keys.

• Create a New Dataset: Navigate to the Datasets tab and click on + Add new dataset. Fill out the form and click on Create dataset.

Adding Data to Your Project:

• For Default datasets, navigate in your bucket following the directory structure. Use the "+ New folder" button to add data, and drag and drop your files and folders to upload them to your bucket.

• For Clinical datasets, the clinical view lets you easily manage your data, but to import data, you will need to set up an ingestion pipeline or upload them directly to your bucket.
• Import Data into a Project: Go to the Actions...->Import data into a project, select the project you wish to import to, and configure how to label this data. This step imports the data into the project and optionally creates a new label for each file in the dataset.
• Previewing Data: Edge Impulse has added a data preview feature, allowing you to visualize certain types of data directly within the organization data tab. Supported data types include tables (CSV/Parquet), images, PDFs, audio files (WAV/MP3), and text files (TXT/JSON).

More info on datasets here.

Community Plan:

Free

• Ideal for students and independent developers focused on Edge AI exploration.
• Highly restricted access to compute resources and enterprise grade features and compute time.
• Non-commercial use only

Professional Plan:

• Ideal for individual developers to formalize exploration of Edge AI specific use cases
• Increased feature set over the Community Plan including NVIDIA TAO models
• Commercial use - when Output is used in an entity with less than 15 FTE’s (FTE = Full-time employees and Full-time and Part-time contractors)

See the Pricing Page for a comparison chart.

Enterprise includes additional features, needed for robust edge AI deployment, which are not included in the Pro Plan:  

• Full Suite of Enterprise Features
• Commercial Use
• Ownership of Output
• Unlimited Compute Time Per Job
• Unlimited DSP file size
• Organization Level Tools
• EON Compiler: BYOM or RAM Optimization Mode
• Private Processing Blocks
• FOMO-AD, DALL-E, HRV DSP
• Premium Support offerings