Modify model a little, Recall/Precison/mAP is very low

[dingxueli]

I modify model a little , only remove the second convolution's relu layer in resnet block, after training, the recall/precison/mAP is very low.
If I we don't restore the model, train the model from zero, the recall / precison / mAP is very low,too. But I restore the model and train, the recall/precison/mAP is OK.
The dataset is VOC2007 + VOC2012, thank you!