def evaluate_test(self, train_data, val_negative_dict, val_negative,sup_dict):
keep_weight = deepcopy(self.state_dict())
hr_list, ndcg_list = [],[]
idxs = range(train_data.idx)
ts = []
for idx in idxs:
positive_x, supp_x, supp_y, sup_mdsk, positive_mdsk, _,_,_, len_pos = train_data.get_batch(
[idx])
u = train_data.support_fill[idx]['user']
pos_embedding = torch.mul(self.item_embeddings[positive_x],
positive_mdsk.unsqueeze(2).repeat(1, 1, 32)).sum(1)
pos_embedding = torch.mul(pos_embedding, len_pos.unsqueeze(1))
user_embedding = self.generate_layer1(pos_embedding)
user_embedding = user_embedding.unsqueeze(1)
for epoch in range(self.args.local_epoch):
t1 = time()
predicted = torch.bmm(user_embedding, self.item_embeddings[supp_x].permute(0, 2, 1)).squeeze(1)
loss = self.loss_func(predicted, supp_y, user_embedding, self.item_embeddings[supp_x], sup_mdsk,len_pos)
grad_user_embedding = torch.autograd.grad(loss, user_embedding, retain_graph=True)[0]
user_embedding = user_embedding - torch.mul(grad_user_embedding, self.local_lr)
t2 = time()
ts.append(t2-t1)
hr, ndcg = evaluation.metrics_meta(self, user_embedding[0], u, val_negative_dict, val_negative,
self.args.topK,self.item_num,train_data)
hr_list.extend(hr)
ndcg_list.extend(ndcg)
self.load_state_dict(keep_weight)
return hr_list, ndcg_list
def evaluate(self, train_data, val_negative_dict, val_negative):
keep_weight = deepcopy(self.state_dict())
val_hr_list, val_ndcg_list, train_loss, val_loss = [], [], [], []
if self.args.mod=='train':
idxs = random.sample(range(len(train_data.support_fill)), 1000)
else:
idxs = range(train_data.idx)
for idx in idxs:
positive_x, supp_x, supp_y, sup_mdsk, positive_mdsk, _,_,_, len_pos = train_data.get_batch(
[idx])
u = train_data.support_fill[idx]['user']
pos_embedding = torch.mul(self.item_embeddings[positive_x],
positive_mdsk.unsqueeze(2).repeat(1, 1, 32)).sum(1)
pos_embedding = torch.mul(pos_embedding, len_pos.unsqueeze(1))
user_embedding = self.generate_layer1(pos_embedding)
user_embedding = user_embedding.unsqueeze(1)
for epoch in range(self.args.local_epoch):
predicted = torch.bmm(user_embedding, self.item_embeddings[supp_x].permute(0, 2, 1)).squeeze(1)
loss = self.loss_func(predicted, supp_y, user_embedding, self.item_embeddings[supp_x], sup_mdsk,len_pos)
grad_user_embedding = torch.autograd.grad(loss, user_embedding, retain_graph=True)[0]
user_embedding = user_embedding - torch.mul(grad_user_embedding, self.local_lr)
hr, ndcg = evaluation.metrics_meta(self, user_embedding[0], u, val_negative_dict, val_negative,
self.args.topK,self.item_num,train_data)
val_hr_list.extend(hr)
val_ndcg_list.extend(ndcg)
self.load_state_dict(keep_weight)
return np.mean(val_hr_list), np.mean(val_ndcg_list)
def evaluate_test(self, train_data, val_negative_dict, val_negative,
sup_dict):
keep_weight = deepcopy(self.model.state_dict())
hr_list, ndcg_list = [], []
tmp_train_loss = []
ts = []
if self.args.mod == 'train':
idxs = random.sample(range(len(train_data.support_fill)), 1000)
else:
idxs = range(train_data.idx)
for idx in idxs:
positive_x, supp_x, supp_y, sup_mdsk, positive_mdsk, query_x, query_y, query_mdsk, len_pos = train_data.get_batch(
[idx])
u = train_data.support_fill[idx]['user']
user_embedding_init = deepcopy(
self.model.user_embeddings).unsqueeze(0)
for _ in range(self.args.local_epoch):
t1 = time()
if _ == 0:
loss = self.model.forward(user_embedding_init, supp_x,
supp_y, sup_mdsk)
grad_user_embedding = torch.autograd.grad(
loss, user_embedding_init, retain_graph=True)[0]
user_embedding_update = user_embedding_init - torch.mul(
grad_user_embedding, self.model.local_lr)
else:
loss = self.model.forward(user_embedding_update, supp_x,
supp_y, sup_mdsk)
grad_user_embedding = torch.autograd.grad(
loss, user_embedding_update, retain_graph=True)[0]
user_embedding_update = user_embedding_update - torch.mul(
grad_user_embedding, self.model.local_lr)
t2 = time()
tmp_train_loss.append(tensorToScalar(loss))
hr, ndcg = evaluation.metrics_meta(self, user_embedding_update[0],
u, val_negative_dict,
val_negative, self.args.topK,
self.item_num, train_data)
hr_list.extend(hr)
ndcg_list.extend(ndcg)
self.model.load_state_dict(keep_weight)
return hr_list, ndcg_list