DG results

[DevinCheung]

Hi, the paper is interesting! Just a clarification, in your paper, it is said that the default option in "In Search of Lost Domain Generalization" does not use the BN layer. However, it seems that the default option is that BN layers are used but they are in "eval" setting, i.e. the weights and biases of BN layers could be optimized but the running mean and running variance are fixed. Therefore, results in Table 1 copied from "In Search of Lost Domain Generalization" are obtained with BN layers in "eval" setting, not obtained without BN layers. Please correct me if my understanding is wrong :)