05 Fakultät Informatik, Elektrotechnik und Informationstechnik
Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/6
Browse
126 results
Search Results
Item Open Access Neural Networks on Microsoft HoloLens 2(2021) Lazar, LéonThe goal of the present Bachelor thesis is to enable comparing different approaches of integrating Neural Networks in HoloLens 2 applications in a quantitative and qualitative manner by defining highly diagnostic criteria. Moreover, multiple different approaches to accomplish the integration are proposed, implemented and evaluated using the aforementioned criteria. Finally, the work gives an expressive overview of all working approaches. The basic requirements are that Neural Networks trained by TensorFlow/Keras can be used and executed directly on the HoloLens 2 without requiring an internet connection. Furthermore, the Neural Networks have to be integrable in Mixed/Augmented Reality applications. In total four approaches are proposed: TensorFlow.js, Unity Barracuda, TensorFlow.NET, and Windows Machine Learning which is an already existing approach. For each working approach a benchmarking application is developed which runs a common reference model on a test datatset to measure inference time and accuracy. Moreover, a small proof of concept application is developed in order to show that the approach also works with real Augmented Reality applications. The application uses a MobileNetV2 model to classify image frames coming from the webcam and displays the results to the user. All the feasible approaches are evaluated using the aforementioned evaluation criteria which include ease of implementation, performance, accuracy, compatibility with Machine Learning frameworks and pre-trained models, and integrability with 3D frameworks. The Barracuda, TensorFlow.js and WinML approaches turned out to be feasible. Barracuda, which only can be integrated in Unity applications, is the most performant framework since it can make use of GPU inference. After that follows TensorFlow.js which can be integrated in JavaScript Augmented Reality frameworks such as A-Frame. Windows ML can currently only use CPU inference on the HoloLens 2 and is therefore the slowest one. It can be integrated in Unity projects with some difficulties as well as plain Win32 and UWP apps. Barracuda and Windows Machine Learning are also integrated in a biomechanical visualization application based on Unity for performing simulations. The results of this thesis make the different approaches for integrating Neural Networks on the HoloLens 2 comparable. Now an informed decision which approach is the best for a specific application can be made. Furthermore, the work shows that the use of Barracuda or TensorFlow.js on the HoloLens 2 is feasible and superior compared to the existing WinML approach.Item Open Access Webanwendung für Multiphysik-Simulationen mit opendihu(2020) Tompert, MatthiasOpendihu ist ein Software-Framework zum Lösen von Multi-Physik-Problemen mit Hilfe der Finiten-Elemente-Methode. Die Anwendungen von Opendihu sind hauptsächlich im Bereich der Skelett-Muskel-Simulationen. Das Erstellen einer Simulation in Opendihu erfolgt über eine C++-Datei, in welcher verschachtelte Löserstrukturen angegeben werden und über eine Python-Datei in welcher die Parameter der verwendeten Löser konfiguriert werden. Das Bearbeiten vorhandener Simulationen und das Erstellen neuer Simulationen mit Hilfe dieser Schnittstelle erfordern gute Kenntnisse über den Sourcecode, beziehungsweise die Struktur von Opendihu. Daher wäre es Sinnvoll Opendihu um eine Nutzerfreundlichere und auch für Einsteiger geeignete Nutzerschnittstelle zu erweitern. Im Rahmen dieser Arbeit habe Ich daher eine grafische Benutzeroberfläche für Opendihu implementiert, welche die Löserstruktur und die Parameter der einzelnen Löser einer Simulation visualisiert. Außerdem ist es mit der Anwendung möglich vorhandene Simulationen zu ändern und neue Simulationen mit Hilfe eines Baukastensystems zu erstellen. Diese Bachelorarbeit erläutert den Aufbau dieser Anwendung und erforscht mit Hilfe einer Nutzerstudie ob die entstandene Benutzerschnittstelle einen Mehrwert gegenüber der bereits vorhandenen Schnittstelle bietet. Das Bearbeiten und Erstellen neuer Simulationen mit Hilfe der Anwendung wurde von den Teilnehmern der Studie im Durchschnitt als einfacher empfunden, als das Bearbeiten und Erstellen neuer Simulationen mit Hilfe der bereits vorhandenen Schnittstelle. Die entstandene Anwendung bietet also einen Mehrwert beim Bearbeiten und Erstellen von Opendihu-Simulationen. Besonders beim Erstellen neuer Simulationen wurde das Baukastensystem als hilfreich bewertet.Item Open Access Optimization of diffusive load-balancing for short-range molecular dynamics(2020) Hauser, SimonIn recent years, multi-core processors have become more and more important for manufacturers, which means that developers now have to think more about how to distribute a single application sensibly over several processes. This is where load balancing comes in, allowing us to move load from an overloaded process to an underloaded process. One way of load balancing is diffusive load balancing, which is a method of moving load in the local neighborhood and therefore no global communication is needed. The advantage of this is that processes that have completed the local communication and thus the load-balancing process can continue with the next calculations. This form of load balancing is found in librepa, a library that deals with the balancing of linked-cell grids and can be used in the simulation software ESPResSo. In the course of this thesis the library has been extended with the First and Second Order Diffusion. Furthermore, a feature was added that allows to keep the initial structure of the grid constant, which means that the neighborhood of each process does not change. This feature is necessary for the Second Order Diffusion. A comparison between the methods shows that both First and Second Order Diffusion distribute the load better in the system than librepa's default and prior to this work only diffusive variant. Furthermore, we show that there is no significant overhead in using the Preserving Structure Diffusion. With the use of flow iteration the imbalance values of First and Second Order Diffusion can be improved even further.Item Open Access Hochperformante Auflösung kleiner Referenzen in verteilten Systemen(2023) Waimer, JoelBereits mit dem Aufkommen erster Filesharing-Systeme wurde die Entwicklung effektiver Verfahren zum Auffinden von mittels global eindeutiger Referenzen bezeichneter Datenobjekte in Peer-to-Peer-Systemen intensiv diskutiert und vorangetrieben, mit dem Ergebnis zahlreicher konkret ausgearbeiteter Lösungsansätze. Unbeachtet geblieben ist dabei jedoch der, für Filesharing-Systeme nicht lohnenswerte, für kleinere verteilte Datenspeichersysteme aber durchaus vorteilhafte Einsatz kleiner Referenzen, in der, aufgrund neuer Fortschritte im Bereich der Speicherdichte, damit einhergehenden dünnen Besetzung dieser kleinen Adressräume, durch welche allerdings die den Verfahren zueigenen Garantien bezüglich der benötigten Anzahl an Schritten zur Auflösung einer Referenz innerhalb des Systems stark verzerrt werden und sich die je Auflösung nötige Laufzeit vergrößert. Diese Arbeit beleuchtet zunächst die Grundlagen der mit der Auflösung von Datenreferenzen in verteilten Speichersystemen einhergehenden Problematiken, beschreibt die beiden Verfahren Chord [SMK+01] und Koorde [KK03] und misst anschließend deren Leistungsfähigkeit in dünn besetzten Adressräumen, unter der Verwendung kleiner Referenzen; mit den Messungen kann schließlich die Vermutung eines negativen Einflusses der dünnen Besetztheit des Adressraums auf die benötigte Laufzeit je Auflösung bestätigt werden. Eingegangen wurde hierbei auch auf mögliche Gegenmaßnahmen zur Verbesserung der Leistungsfähigkeit der beiden Verfahren, wobei hier die Verbindung der beiden untersuchten Verfahren mit einer Abwandlung des beim Distance-Halving-Netzwerk [NW03] eingesetzten Initialisierungsverfahrens zu einer nahezu gleichmäßigen Aufteilung des Adressraumes auf die einzelnen Knoten hier großes Potential besitzt, da so einer Entartung der Pfadlänge je Auflösung entgegengewirkt werden kann; zudem zeigten sich in den Messungen stark ungünstige Auswirkungen einer naiven, iterativen Implementierung des Chord-Verfahrens gegenüber Koorde.Item Open Access Weighted independent colorful sets in large vertex colored conflict graphs for timetriggered flow scheduling(2022) Rönsch, LorenzThe need for time sensitive communication on networks is increasing more and more, especially due to Industrial internet of things and Industry 4.0. With the appearance of graphics-based network participants in time-critical networks, such as VR glasses, the absolute amount of traffic that needs to be scheduled over the network increases strongly. The most common method to realize real time communication is using the IEEE Time-Sensitive Network (TSN) and the Time-Aware Shaper (TAS). However, the TSN schedule calculation is not standardized. There are several approaches, such as SMT solver, integer linear programming and calculating a conflict graph to calculate time-triggered flow schedules. But none of them are tackling the problem of maximizing traffic. In our work, we extend the time-triggered flow scheduling problem to include the component of maximum traffic. For this purpose, we modify an existing heuristic, called Greedy Flow Heap Heuristic, so that we can adapt the scheduling to our problem. The results that our version provides compared to the original heuristic are very promising. On all our evaluation data, we achieved an average improvement of 81.91% in terms of maximum network traffic. We also developed an alternative non-deterministic approach based on a genetic algorithm. In our work we investigate different variants of the algorithm with the goal to provide better results with different adaptations of the algorithm. In our repair version, we manage to beat our benchmark algorithm the Greedy Flow Heap Heuristic on every circle based conflict graph.Item Open Access Parameter-dependent self-learning optimization(2022) Abu El Komboz, TareqManually developing optimization algorithms is a time-consuming task requiring expert knowledge. Therefore, it makes a lot of sense to automate the design process of such algorithms. Additionally, learned optimization algorithms reduce the number of a priori assumptions made about the characteristics of the underlying objective function. Numerous works discuss possibilities for learning optimization algorithms. This field of study is called learn-to-optimize. In this bachelor’s thesis, we concentrate on the reinforcement learning perspective. Consequently, optimization algorithms are represented as policies. The comparison of learned algorithms to current state-of-the-art algorithms for particular applications reveals that learned algorithms manage to perform better concerning convergence speed and final objective function value. However, most existing approaches only consider fixed sets of parameters to be optimized. Because of this, it is challenging to adapt the learned optimization algorithm to other objective functions. More importantly, it is impossible to optimize when explicit constraints on so-called “free” optimization parameters are given. We investigated the learn-to-optimize approach under various optimization parameter sets and conditions on “free” parameters to solve this problem. Furthermore, we studied the performance of learned optimizers in high-dimensional setups.Item Open Access Future-proof scheduling heuristics for TTEthernet(2024) Haas, RicoTime-triggered networks run on a premade schedule to ensure that all messages arrive at their destination on time and without being dropped. They are deployed in safety-critical environments like airplanes, cars, but also factories. Online scheduling algorithms can create schedules for such networks in mere seconds while also being able to modify an existing schedule. With an algorithm of this kind, new devices can be plugged into its time-triggered network and start interacting with it in seconds, without manual setup. We propose modifications for improving two already existing online scheduling algorithms: Hierarchical Heuristic Scheduling (H2S) and Cost-Efficient Lazy Forwarding (CELF). We call these modifications "relaxed jitter", "backward rifts", and "gap closing". Our modifications achieved about 550 MBit/s more aggregated throughput while only increasing solving time by 2.5ms in a network with 25 switches.Item Open Access Distributed Deep Reinforcement Learning for Learn-to-optimize(2023) Mayer, PaulIn the context of increasingly complex applications, e.g., robust performance tuning in Integrated Circuit Design, conventional optimization methods have difficulties in achieving satisfactory results while keeping to a limited time budget. Therefore, learning optimization algorithms becomes more and more interesting, replacing the established way of hand-crafting or tweaking algorithms. Learned algorithms reduce the amount of assumptions and expert knowledge necessary to create state-of-the-art solvers by decreasing the need of hand-crafting heuristics and hyper-parameter tuning. First advancements using Reinforcement Learning have shown great success in outperforming typical zeroth- and first-order optimization algorithms, especially with respect to generalization capabilities. However, training still is very time consuming. Especially challenging is training models on functions with free parameters. Changing these parameters (that could represent, e.g., conditions in a real world example) affects the underlying objective function. Robust solutions therefore depend on thorough sampling, which tends to be the bottleneck considering time consumption. In this thesis we identified the runtime bottleneck of the Reinforcement Learning Algorithm and were able to decrease runtime drastically by distributing data collection. Additionally, we studied the effects of combining sampling strategies in regards to generalization capabilities of the learned algorithm.Item Open Access Investigation of self-learned zeroth-order optimization algorithms(2022) Schüttler, KilianDesigning optimization algorithms manually is a laborious process. In Addition, many optimization algorithms rely on hand-crafted heuristics and perform poorly in applications for which they are not specifically designed. Thus, automating the algorithm design process is very appealing. Moreover, learned algorithms minimize the amount of a priori assumptions and do not rely on hyperparameters after training. Several works exist that present methods to learn an optimization algorithm. In this project, we focus on the reinforcement learning perspective. Therefore, any particular optimization algorithm is represented as a policy. Evaluation of the existing methods shows, learned algorithms outperform existing algorithms in terms of convergence speed and final objective value on particular training tasks. However, the inner mechanisms of learned algorithms largely remain a mystery. A first work has discovered that learned first-order algorithms show a set of intuitive mechanisms that are tuned to the training task. We aim to explore the inner workings of learned zeroth-order algorithms and compare our discoveries to previous works. To address this issue, we study properties of learned zeroth-order algorithms to understand the relationship between what is learned and the quantitative and qualitative properties, e.g., curvature or convexity of the objective function. Furthermore, we study the generalization in relation to these properties. Moreover, we explore the feasibility of finetuning a learned zeroth-order optimization algorithm to a related objective function. Finally we provide guidelines for training and application of learned zeroth-order optimization algorithms.Item Open Access Advanced object localization pipeline for robot manipulation(2020) Schäfer, TimThe thesis objective is to develop a robust tracking pipeline for a Baxter robot system. The tracking pipeline includes the applications of robot tracking and object tracking, which localizes the surrounding objects for robot manipulation. By combining, extending, and comparing state-of-theart approaches for object tracking and image segmentation, this thesis estimates the initial poses of the objects such that their pose can be used to auto-initialize object tracking algorithms. The pipeline is able to handle unfamiliar and dynamic environments. The system was implemented in simulation and can be applied to a real robotic system. The resulting pipeline uses an RGB image, a depth image, and the 3D object models as input. The outputs are the tracked object poses in real time. A dataset was generated to train the instance segmentation network. Furthermore, the pipeline was evaluated with several conditions to test the robustness of the tracking. A comparison to other 6D pose estimation approaches is provided in the results. The code of the pipeline is available on GitHub: https://github.com/timschaeferde/rai_baxter