The pre-training model of the published coco data set is not problematic, and the test results are poor. 