persai-lab / edm2021-dmkt Goto Github PK

Is there a reason why the value matrix is initialized randomly at each forward pass in dmkt.py?
Should not the initial value of the value matrix be part of the model parameter ?
Isn't there a risk of having inconsistent results during the test phase ?
Did I miss something ?

self.value_matrix = torch.Tensor(self.num_concepts, self.value_dim).to(self.device)
nn.init.normal_(self.value_matrix, mean=0., std=self.init_std)
self.value_matrix = self.value_matrix.clone().repeat(batch_size, 1, 1)

Thank you