With the wide availability of inertial sensors in smartphones and connected objects, interest in sensor-based activity recognition has risen. Yet, recognizing human actions from inertial data remains a challenging task because of the complexity of human movements and of inter-individual differences in movement execution. Recently, approaches based on deep neural networks have shown success on standardized activity recognition datasets, yet few works investigate systematically how these models generalize to other protocols for data collection. We present a study that evaluates the performance of various deep learning architectures for activity recognition from a single inertial measurement unit, on a recognition task combining data from six publicly available datasets. We found that the best performance on this combined dataset is obtained with an approach combining the continuous wavelet transform and 2D convolutional neural networks.