ESP-DL, recently introduced by Espressif Systems, provides APIs for neural network (NN) inference, image processing, maths operations and deep learning (DL) models. With ESP-DL, developers can use Espressif's SoCs for neural network inference easily and quickly.
Thus far, Espressif’s data show that, in comparison to the company’s ESP32 wireless MCU, when ESP-DL runs on an ESP32-S3 chip the acceleration of a 16-bit detection model is increased by 4,5 times, or 6,25 times in the case of a face recognition model. In addition, the 8-bit face recognition model is 2,5 times faster than the 16-bit model on the ESP32-S3. Furthermore, Espressif offers a tool that can help developers make their own model compatible with the company’s platform.
Since ESP-DL does not need any peripherals, it can be used as a project component. For example, if it is added to the esp-who/components/directory, ESP-DL can be used as a submodule of ESP-WHO, which contains several project-level examples of image-based applications. The accompanying figure below shows what ESP-DL consists of and how it is implemented as a project component.
ESP-DL provides tools that can help developers convert their own model (possibly developed with a third-party platform such as TensorFlow, PyTorch, MXNet, etc.) into an 8-bit or 16-bit model and evaluate the performance of the quantised model.
ESP-DL also provides certain easy-to-use models such as human face detection, human face recognition, cat face detection, etc., in the so-called Model Zoo. These models can be used out-of-the-box. Other models, such as for colour detection and hand-pose recognition, will be released soon.
Customising a model
In the GitHub tutorial entitled ‘How to Customize a Model Step by Step’ you can find all relevant instructions with a runnable example which will help you implement your model quickly and easily.
Variety of APIs
ESP-DL contains sufficient APIs for developers to build their own model, such as neural networks (NN), image processing and matrix operations. Meanwhile, developers can customise layers and provide feedback if they have any special requirements regarding the API utility.
Software and hardware boost
ESP-DL implements quantised computation and brings about a more efficient kind of software by optimising the assembly and architecture of the C/C++ code. It is worth mentioning that the ESP32-S3, with its vector instructions, high-speed SPI interface and configurable cache memory, achieves a much faster acceleration in AI applications.
|Tel:||+27 11 781 2029|
|Fax:||086 555 7808|
|Articles:||More information and articles about iCorp Technologies|
© Technews Publishing (Pty) Ltd | All Rights Reserved