Due to the ambiguity of facial expressions, the subjectivity of annotators, and the low quality of images, noisy labels pose a significant challenge in the field of facial expression recognition. We address the noise label problem based on the principle of contrastive learning. Through CAM visualization, we found that when noisy labels are present, the model can make correct predictions due to its inability to memorize noisy labels from different data augmentation perspectives. Inspired by this observation, we propose the GSMC method, which enforces model consistency between the predictions of the original image and its augmented versions. Specifically, we divide the consistency requirement into two different tasks. The first task uses attention map consistency, emphasizing preventing the model from memorizing noise labels. The second task employs self-distillation learning, where the consistency is achieved through the predictions of student and self-distillation networks, enabling the model to learn more robust label distributions. Our framework, based on ResNet18, achieves a 90.54% accuracy on the RAF-DB dataset, achieving State-Of-The-Art performance. The code will be available at https://github.com/ECNU-Cross-Innovation-Lab/GSMC .