Money laundering (ML) poses a severe threat to financial stability and social security. Various money laundering detection methods have emerged in the past two decades. Among these methods, some semi-supervised ones based on graph neural networks (GNNs) have achieved impressive performance. However, the homogeneity hypothesis of GNN-based methods does not fit the ML detection scenario, affecting the detection performance. This paper presents a semi-supervised money laundering detection framework based on decoupling training (MLaD2). MLaD2 constructs a transaction relationship network based on node similarity (TRNNS) to model account interactions. Performing on TRNNS, MLaD2 learns the representation of accounts using a GNN. The weighting mechanism of TRNNS can overcome the drawback of the homogeneity hypothesis. Based on the learned account representations, MLaD2 adopts a decoupling training mechanism to build an ML accounts detection model, reducing its dependence on annotated data. The pre-training phase of the decoupling training employs a contrastive self-supervised learning model to learn the intrinsic characteristics of accounts. The fine-tuning phase extracts discriminative features between ML accounts and benign accounts with labeled data. Comprehensive evaluations and comparisons on a real-world ML dataset demonstrate that MLaD2 yields results that surpass existing methods, especially when training with a small scale of labeled samples.