Questions about the third step of the training

[AXu0511]

Following your training steps, after running the first step, the Dice value of my registration model(UnetFlowModel) was 0.768. Then I run the second step. When I run the third step, line 338 of the code is the test code, and the test result showed that the performance of the alignment model dropped to 0.73. Should I use the pre-trained StyleEncoder when running the third step, and if not, what is the purpose of the second step?