Galaxy Morphology Classification Based On Distributed Neural Network

The field of astronomy has entered a data-driven era of "astronomical big data".In the field of Galaxy morphology research,it is a trend to use computer to analyze Galaxy morphology data.In the deep learning model,although convolutional neural network can extract high-dimensional features of Galaxy morphology image data efficiently and automatically,the classification accuracy of shallow model is not high,and the model with too deep and insufficient width will bring instability and poor robustness.In order to solve the problem of instability and insufficient robustness of Galaxy morphology classification model,this thesis selects the concept-resnet-v2 to add Mixed preprocessing technology,proposes the concept mixed classification model,redesigns the stem module,and combines the mixed preprocessing technology,abandons the traditional "serialization" preprocessing method,and uses the "parallelization" scheme,combined with the weight value of each path,It can effectively improve the balance of training samples and improve the stability and robustness of the network model.In this paper,the concept mixed is compared with vgg-16,resnet-50 and concept-resnet-v2.The experimental results show that the concept mixed model can effectively improve the robustness and stability of the model on the premise of ensuring the classification accuracy.In this thesis,aiming at the problem that the effect of accelerating training is not obvious due to the limitation of synchronization fence in distributed training,a DO-SGD algorithm based on synchronization training is proposed considering the difference of computing performance of GPU cluster in production environment.In the data distribution phase,the algorithm takes into account the difference of computing power of different computing nodes,maintains the training schedule of each computing node,adjusts the data distribution strategy in real time according to the state of the computing node,and reduces the time of each iteration in the training process,so as to reduce the time required for the whole training.Experiments show that DO-SGD can reduce the training time by nearly 57% at the expense of 0.47% accuracy when training Galaxy morphology data set.This thesis designs and implements a prototype system of Galaxy morphology classification.The system includes modules of displaying source data information,model building and Galaxy shape prediction.The experimental results show that the prototype system can not only reduce the difficulty of classification model construction,improve the stability of classification model,but also accelerate the training and reduce the training time.