model.train() and model.eval() gives different results

I trained a model and did inference, the model predicts the same bounding boxes for all images with model.eval(). However, the model gives different results with model.train(). Why does the model perform so strange with model.train() and model.eval()? Can someone helps me?