Recently self-supervised learning has been proposed in the field of human activity recognition as a solution to the labelled data availability problem. The idea being that by using pretext tasks such as reconstruction or contrastive predictive coding, useful representations can be learned that then can be used for classification. Those approaches follow the pretrain, freeze and fine-tune procedure. In this paper we will show how a simple change - not freezing the representation - leads to substantial performance gains across pretext tasks. The improvement was found in all four investigated datasets and across all four pretext tasks and is inversely proportional to amount of labelled data. Moreover the effect is present whether the pretext task is carried on the Capture24 dataset or directly in unlabelled data of the target dataset.