| In recent years,with the continuous development and progress of deep learning,a large number of unprecedented achievements have been obtained in this field,and deep learning algorithms have also been widely used in various industries.As the scale of the network is getting bigger and bigger,the requirement of computing power is getting higher and higher.However,currently mainstream deep learning frameworks such as Tensorflow,MXNet,and Caffe can only perform some server-level GPU optimization.There are various difficulties in effectively deploying deep neural networks on some devices with insufficient resources.Therefore,related research work has been carried out in support of dedicated chips.On the basis of summarizing and introducing the research status of related work at home and abroad,this tensis proposes a feasible plan for deploying TFlite model on the dedicated chip Deep Eye2000 based on the TVM framework.The main content of this article is as follows:(1)Establish a TFlite model library with rich types and high operator coverage,with a total of 291 models,which lays the foundation for transplanting,improving and optimizing the front-end analysis interface of TVM’s TFlite model.(2)According to the design of the front-end analysis interface of TVM,referring to the front-end analysis interface of other deep learning frameworks of TVM and the TFlite front-end analysis interface of TVM0.8,the TFlite front-end analysis interface of TVM used in the research project was transplanted and realized.The degree of support of the model in is 33%.(3)In order to verify the support degree of TVM’s TFlite front-end analysis interface to the model,two verification methods for verifying the accuracy loss after model analysis are designed and implemented.rate verification.According to the verification method,a total of test cases with more than 6,000 lines of code were constructed for all models.In the process of repeatedly verifying the support degree of the interface to the model with the test cases,the TFlite front-end analysis interface of TVM was improved and optimized,and the 229 The support of TFlite models increases the support rate of models in the model library to 78.7%.(4)Apply the optimized TVM’s TFlite front-end analysis interface to the tool chain TVM of the dedicated chip Deep Eye2000,design and implement a deep learning model deployment system,and successfully deploy the model to Deep Eye2000 through the system. |
