National Repository of Grey Literature 2 records found  Search took 0.00 seconds. 
Conversion of Models between Machine Learning Frameworks for Mobile Platforms
Pavella, Martin ; Zbořil, František (referee) ; Kočí, Radek (advisor)
Machine learning frameworks use various formats to represent and store models of deep neural networks (DNN). One of the most commonly used ones is Open Neural Network Exchange (ONNX). Developing drivers for hardware accelerators on embedded systems is expensive, and ONNX is rarely supported. The necessary software support is typically only implemented for the TensorFlow Lite (TFLite) DNN model format. Currently, the options for conversion of pre-trained ONNX models to TFLite are inadequate and produce suboptimal models. This work focuses on designing and developing a direct converter of ONNX models to TFLite, which produces as optimal models as possible. The resulting program was verified on real models in collaboration with the NXP company. The models produce identical outputs after conversion and their inference speed on target platforms is significantly higher.
Using the library for data serialization in embedded system
Slavov, Jan ; Burian, František (referee) ; Petyovský, Petr (advisor)
This diploma thesis deals with the possibilities of using serialization libraries for communication with embedded systems. Data serialization is a process that converts data objects organized into complex data structures into a stream of bytes. Data serialized in this way can be easily transferred between devices or stored. At the same time, serialization enables the platform to transfer data in a programing language-neutral manner. Also addresses compatibility when updating messages. This work will primarily deal with binary serialization, as it is less time-consuming and the resulting messages are smaller in size. This work will describe work with the following libraries for data serialization: Flatbuffer, Protocol Buffer, Cap'n Proto. These libraries will then be compared with each other and one library will be selected from the results. A demonstration task is designed for this library. It will be implemented in two versions. The first version will use the selected library and the second will use my own method of data serialization. Both approaches will be compared with each other at the end of this thesis.

Interested in being notified about new results for this query?
Subscribe to the RSS feed.