A semi-supervised learning method based on the distance metric loss framework was proposed in order to solve the problems of different types of loss functions and inconsistent loss scales in the training process of semi-supervised learning methods, which make it difficult to adjust the loss weights, inconsistent optimization directions and insufficient generalization ability. A unify loss framework function was proposed from the perspective of distance metric loss, and the adjustment of loss weights between different loss functions in semi-supervised tasks was achieved. Adaptive similarity weights were introduced for the target region problem of embedding vectors in the loss framework in order to avoid the conflict of optimization directions of traditional metric learning loss functions and improve the generalization performance of the model. CNN13 and ResNet18 networks were used to construct semi-supervised learning models on CIFAR-10, CIFAR-100, SVHN, STL-10 standard image dataset and medical pneumonia dataset Pneumonia Chest X-ray, respectively, for comparison with commonly used semi-supervised methods in order to validate the effectiveness of the method. Results show that the method has the optimal classification accuracy under the condition of the same number of labels.