关于Compositional Attention的问题

[pistachio2]

为什么训练和测试的时候，inputs2 = torch.randn(int(mat.shape[0]), 7)都是randn,这样对于一个特定的图片应该赋予相同的参数，随机的参数对于每一张特定的图片产生的参数是一样的吗？这样的随机参数向量的意义是什么？

Why are inputs2 = torch.randn(int(mat.shape[0]), 7) both randn during training and testing, so that the same parameters should be assigned to a specific picture, and random parameters are for each specific picture. Are the parameters generated by the pictures the same? What is the significance of such a random parameter vector?