Advances in the area of ubiquitous computing have paved the way for a plethora of mobile applications operating in smart devices that aim to assist various aspects of everyday life. In the context of human-environment interaction in smart homes, remote monitoring of older adults, or general improvement of daily habits, the estimation of human activities in real-time and in the appropriate context is becoming of increasing importance. At the same time, the ever-increasing capabilities of smartwatches and their premium wrist-attached placement makes them an exciting tool for solving the above problem even for tasks of high manual dexterity. In this work, we propose a hierarchical framework on smartwatch IMU data to learn activities of daily living in varying granularity: from high-level, generic descriptions to low-level detailed activities. We employ a swarm of CNN-LSTM-based classifiers across N hierarchy levels, trained with a set of features tailored to each task addressed. Our proposed hierarchical method achieves tangible improvement in classification accuracy and up to 4x speed-up in inference times compared to the standard non-hierarchical approach. We also contribute a new dataset of smartwatch IMU data used to build and evaluate our method, so that future research in the community can benefit from it.