The invention discloses a target detection model
pruning method, and relates to the technical field of model
pruning, and the method comprises the following steps: constructing and training a target detection model, carrying out the
pruning training of the target detection model, obtaining a structured stripe pruning model, and carrying out the re-training of the model through a
mixed precision training method, so as to complete the pruning. According to the method, training and dynamic sparsification
processing are performed on the model before pruning of the target detection model, so that the target detection model can reach the expected accuracy before pruning, the model performance is improved, the subsequent pruning
processing efficiency is improved, pruning training is performed on the target detection model in combination with network pruning and stripe pruning, and the training efficiency is improved. According to the method, a
convolution kernel skeleton matrix is obtained, stripes, smaller than a pruning threshold, in the corresponding
convolution kernel skeleton matrix are sheared according to a preset pruning proportion, so that structuring of the
convolution kernel is achieved, meanwhile, a structured stripe pruning model is trained again through a
mixed precision training method to complete final model pruning
processing, and the
compression ratio and precision loss of the model can be balanced.