源码：
chain_M = chain_M*mask #update chain_M to include missing regions
decoding_order = torch.argsort((chain_M+0.0001)* (torch.abs(torch.randn(chain_M.shape, device=device)))) #[numbers will be smaller for places where chain_M = 0.0 and higher for places where chain_M = 1.0]
mask_size = E_idx.shape[1]
permutation_matrix_reverse = torch.nn.functional.one_hot(decoding_order, num_classes=mask_size).float()
order_mask_backward = torch.einsum('ij, biq, bjp->bqp',(1-torch.triu(torch.ones(mask_size,mask_size, device=device))), permutation_matrix_reverse, permutation_matrix_reverse)
mask_attend = torch.gather(order_mask_backward, 2, E_idx).unsqueeze(-1)
mask_1D = mask.view([mask.size(0), mask.size(1), 1, 1])
mask_bw = mask_1D * mask_attend
mask_fw = mask_1D * (1. - mask_attend)代码解读：
1. chain_M =
)
