Although using ordinary differential equation (ODE) to obtain numerical solutions of ODE parameterized by graph convolution network can speed up training and reduce the risk of over-smoothing, they face the problem of over-fitting in training. For the above problems, this paper proposes a Contrastive Learning for ODE-based collaborative filtering(CL4ODE) method. Firstly, the embedding masking module randomly masks the initial embeddings of users/items. Simultaneously, the initial embeddings are input into the gating module, and different methods are applied to augment the gating output, resulting in various versions of embeddings. Subsequently, the masked embeddings are concatenated with the embeddings augmented and input into the ODE module. Finally, half of the embedding matrix of the ODE output is used to calculate the BPR loss, and the contrast loss is calculated together with the other half. The experimental outcomes prove the significance of our proposal compared with existing recommendation methods.