Human activity recognition (HAR) tasks are an active area of research within the field of pervasive and ubiquitous computing. As sensors become more ubiquitous in mobile devices, new applications where human activitiy recognition is needed  appear at a fast pace. Thus there is need for a generic HAR classifier that can provide accurate predictions in diverse HAR problems.  In this paper, we propose an universally applicable classification method for HAR tasks that is capable of reaching near cutting edge results in various HAR tasks without expert knowledge. We present an end-to-end classification method based on raw accelerometer signals. A novel pre-processing step using horizontal, vertical, as well as gravitational components is developed and combined with an attentive neural network. The network uses an adjustable compression phase to allow sequences of various lengths to be trained and classified. Our findings are evaluated using three different publicly available datasets and four different tasks with 8 transportation modes, 6 different on-body locations, 6 different activities, spanning sequences between 1 and 60 seconds. We reach state-of-the-art results in datasets such as RealWorld HAR with 94.3\% on-body localization and 90.6\% motion state F1-scores.