Diabetic Retinopathy (DR) is a complication caused by diabetes that can destroy the retina, leading to blurred vision and even blindness. We proposed a deep neural network to classify DR at different stages. The design philosophy of the network is based on classical residual networks. We used a multi-attention mechanism to improve the model's attention for lesion detection. Our proposed network uses a novel down-sampling method that enhances the receptive field size of the model. We introduced Squeeze-to-Excitation blocks to increase the complexity of the model and thus theoretically increase the representational power of the model. The experimental results show that the performance of our model is about 1%-2% higher than other models on the EyePACS test set. The method helps enhance the discrimination ability of the model for non-healthy retinal images and has specific clinical application value.