在配置文件中有参数是每隔5个epoch进行一次AP的测试。我在微调模型,训练到100epoch时会测试一下当前的AP然后保存这个最后的模型。然后用这个模型,搭配完全相同的配置文件,用测试脚本进行测试时却发现AP大幅下降,几乎减少了一半。希望有大佬能解答一下
In the configuration file, there is a parameter set to test AP every 5 epochs. I am fine-tuning the model, and when training reaches 100 epochs, it tests the current AP and saves this final model. However, when I use this model with the exact same configuration file and run the test script, I notice a significant drop in AP—almost halved. I hope an expert can explain this.