Abstract
The iterative training required to improve the performance of machine learning (ML)models becomes inefficient when relying on manual retraining and redeployment. This
master thesis presents the development and evaluation of a machine learning operations
(MLOps) pipeline designed for the continuous training and mobile deployment of a food
recognition model. In addition, this work provides valuable insights into the effectiveness of data augmentation and addresses the challenges associated with deploying and
maintaining ML models in production environments.
The pipeline integrates Apache Airflow and MLflow in a local environment. A mobile
Android application collects training data and another mobile application deploys the
model in an offline-enabled order system, where the food recognition model is used to
add items to an order. Additionally, data augmentation techniques such as geometric
and color transformations are applied to improve model performance and increase the
volume of training data.
The performance of the pipeline is evaluated using metrics such as training time
and resource usage, while the model performance on augmented data is evaluated using accuracy, precision, recall and F1-score. The performance evaluation reveals that
while the local pipeline setup is sufficient for sequential execution, it lacks the scalability required for concurrent task execution. This suggests the need for a cloud-based
infrastructure to effectively handle increasing workloads. Data augmentation shows a
significant improvement in model performance, with the best-performing augmentation.
technique being a combination of geometric and color transformations
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Marc Kurz (Supervisor) |