想将model1的entities_emb,也当作第model2的entities_emb,这样在更新model2参数的时候,能更新entities_emb。
首先在定义模型时,模型二的参数中加入model1.entities_emb
model1 = Model1(n1, n2, norm=distance_norm, dim=embedding_dim).cuda()
model2 = Model(n3, n4, device, model1.entities_emb, args).to(
device)
然后各自训练:
for epoch in range(epochs):
optimizer1.zero_grad()
loss1 = criterion()
loss1.backward(retain_graph=True)
optimizer1.step()
optimizer2.zero_grad()
loss2 = criterion()
loss2.backward(retain_graph=True)
optimizer2.step()
思路就是这样了,成功实现联合训练