Titulo Estágio
mModelsAnalyser - Standardization and analysis of machine learning models on mobile scenarios
Áreas de especialidade
Engenharia de Software
Sistemas Inteligentes
Local do Estágio
Fraunhofer AICOS Headquarters (Rua Alfredo Allen, 455/461 4200-135 Porto) or Remote Work.
Enquadramento
Abstract
Most of the deep learning exploration and experimentation process is done in a desktop environment, and when transformed to be applied on a smartphone scenario, the model’s inference time and performance change, making it unreliable sometimes. This change comes with the need for the models to be adapted (1) in order to match the target deployment environment and to match the requirements for that specific scenario.
Objetivo
Objectives and Expected Results
The main goal for this thesis is to produce a mobile app able to test and compare multiple DL models and assure that they are reliable concerning the model’s achieved performance, the model’s inference time (for real-time or non real-time scenarios) and the device’s resources. Once these metrics are computed for each model that is being tested, the app should be able to recommend a specific model or a list of models based on one of the parameters of a set of parameters previously descried, assuring that the selected model is compliant with the requirements that the solution may have (e.g. real time solution).
The computation of such metrics can be done using similar benchmark tools such as those described in (2)(3).
For each tested model the app must generate a report chart that would allow to support the choice of a given model between the tested ones.
The solution must be able to test different models’ architectures and be compliant with multiple input and output data structures, in order for the application to be used in different domain scenarios.
Innovation Aspects
This comparison tool will allow a fast and extensive comparison of multiple DL models when applied to a smartphone use scenario and ease the effort when adapting and deploying previously developed models to this specific context.
Plano de Trabalhos - Semestre 1
Workplan
T1 – Literature review.
T2 – Study and evaluation of possible models to be used in order to develop the application.
T3 – Intermediate Report writing.
Plano de Trabalhos - Semestre 2
Workplan
T1 – Literature review (cont.).
T2 – Study and evaluation of possible models to be used in order to develop the application (cont.).
T3 – Research and implement the application flow to handle data and run the models.
T4 – Implement the model reporting system: chart generation and model selection.
T5 – Thesis writing.
Condições
The internship will be hosted at Fraunhofer AICOS Headquarters (Rua Alfredo Allen, 455/461 4200-135 Porto) or by Remote Work (to be discussed with the student).
The student will have:
- Support and supervision provided by AICOS' Researcher Pedro Faria;
- Access to models or the inference results of the models developed in a desktop environment;
- Access to the respective quantized models to be integrated and tested in the Android Application.
Observações
Candidate Profile
- Basic competences in Android Development.
- Good programming experience in kotlin and/or C++.
- Basic knowledge of machine/deep learning concepts.
- Experience with machine/deep learning is a plus.
References
(1) Andrey Ignatov, Radu Timofte, Andrei Kulik, Seungsoo Yang, Ke Wang, Felix Baum, Max Wu, Lirong Xu, & Luc Van Gool. (2019). AI Benchmark: All About Deep Learning on Smartphones in 2019.
(2)https://www.tensorflow.org/lite/performance/measurement
(3)https://github.com/XiaoMi/mobile-ai-bench
Orientador
Pedro Miguel Martins de Lemos da Cunha Faria
pedro.faria@fraunhofer.pt 📩