代码与论文中描述似乎不一致

[lrf23]

1.HGNN层，似乎没有用到leakyReLU，但是论文中这里说用到了激活函数。

class HGNNLayer(nn.Module):
	def __init__(self):
		super(HGNNLayer, self).__init__()
		self.act = nn.LeakyReLU(negative_slope=args.leaky)
	
	def forward(self, adj, embeds):
		# lat = self.act(adj.T @ embeds)
		# ret = self.act(adj @ lat)
		lat = (adj.T @ embeds)
		ret = (adj @ lat)
		return ret

2.bpr损失函数，论文中的公式似乎是被注释掉的这行代码，而未被注释的似乎是另外一种形式的损失，好像没在论文里找到，

scoreDiff = pairPredict(ancEmbeds, posEmbeds, negEmbeds)
bprLoss = - (scoreDiff).sigmoid().log().mean()
# bprLoss = t.maximum(t.zeros_like(scoreDiff), 1 - scoreDiff).mean() * 40

3.最终预测使用trnMask的意义是什么，论文中似乎并未提及，可否直接embedding相乘呢

for usr, trnMask in tstLoader:
	i += 1
	usr = usr.long().cuda()
	trnMask = trnMask.cuda()
	usrEmbeds, itmEmbeds = self.model.predict(self.handler.torchBiAdj)
	allPreds = t.mm(usrEmbeds[usr], t.transpose(itmEmbeds, 1, 0)) * (1 - trnMask) - trnMask * 1e8
	_, topLocs = t.topk(allPreds, args.topk)
	recall, ndcg = self.calcRes(topLocs.cpu().numpy(), self.handler.tstLoader.dataset.tstLocs, usr)
	epRecall += recall
	epNdcg += ndcg
	log('Steps %d/%d: recall = %.2f, ndcg = %.2f          ' % (i, steps, recall, ndcg), save=False, oneline=True)