def doEva(net, d, G):
net.eval()
d = torch.LongTensor(d)
u, i, r = d[:, 0], d[:, 1], d[:, 2]
i_index = i.detach().numpy()
adj_lists = dataloader4graph.graphSage4RecAdjType(G, i_index)
out = net(u, adj_lists)
y_pred = np.array([1 if i >= 0.5 else 0 for i in out])
y_true = r.detach().numpy()
p = precision_score(y_true, y_pred)
r = recall_score(y_true, y_pred)
acc = accuracy_score(y_true, y_pred)
return p, r, acc
def train(epoch=20,
batchSize=1024,
dim=128,
tdim=64,
lr=0.002,
eva_per_epochs=1,
atten_way='base'):
user_set, item_set, train_set, test_set = dataloader4graph.readRecData()
entitys, pairs = dataloader4graph.readGraphData()
G = dataloader4graph.get_graph(pairs)
net = GAFM(max(user_set) + 1, max(entitys) + 1, dim, tdim, atten_way)
criterion = torch.nn.BCELoss()
optimizer = torch.optim.AdamW(net.parameters(), lr=lr)
for e in range(epoch):
net.train()
all_lose = 0
for u, i, r in tqdm(
DataLoader(train_set, batch_size=batchSize, shuffle=True)):
r = torch.FloatTensor(r.detach().numpy())
optimizer.zero_grad()
i_index = i.detach().numpy()
adj_lists = dataloader4graph.graphSage4RecAdjType(G, i_index)
logits = net(u, adj_lists)
loss = criterion(logits, r)
all_lose += loss
loss.backward()
optimizer.step()
print('epoch {}, avg_loss = {:.4f}'.format(
e, all_lose / (len(train_set) // batchSize)))
# 评估模型
if e % eva_per_epochs == 0:
p, r, acc = doEva(net, train_set, G)
print('train: Precision {:.4f} | Recall {:.4f} | accuracy {:.4f}'.
format(p, r, acc))
p, r, acc = doEva(net, test_set, G)
print('test: Precision {:.4f} | Recall {:.4f} | accuracy {:.4f}'.
format(p, r, acc))
return net