def main(_):
assert sum([FLAGS.train, FLAGS.predict, FLAGS.eval]) == 1
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.log_dir):
os.makedirs(FLAGS.log_dir)
# config = tf.ConfigProto(gpu_options=tf.GPUOptions(per_process_gpu_memory_fraction=0.5))
# config.gpu_options.allow_growth = True
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
# with tf.Session(config=config) as sess:
with tf.Session() as sess:
dnn = DNN(sess, FLAGS)
if FLAGS.train:
# os.environ["CUDA_VISIBLE_DEVICES"] = '1'
dnn.fit()
elif FLAGS.predict:
dnn.load_network()
samples = np.array(
pd.read_csv('dataset/gen_samples.csv', header=None))
gen_y = samples[:, -1]
predict = dnn.predict(np.delete(samples, -1, 1))
# assert gen_y.shape[0] == predict.shape[0]
# print 'Accuracy: {}%'.format((predict == gen_y).sum() / float(predict.shape[0]) * 100)
samples = samples[gen_y != predict]
pd.DataFrame(samples).to_csv('dataset/gen_samples.csv',
index=False,
header=None)
elif FLAGS.eval:
dnn.load_network()
dnn.eval()
def parse_individual(self, indi):
torch_device = torch.device('cuda')
train_data, test_data, user_num, item_num, train_mat = data_loader.load_dataset(
)
train_dataset = Data(train_data,
item_num,
train_mat,
num_ng=4,
is_training=True) # neg_items=4,default
test_dataset = Data(test_data,
item_num,
train_mat,
num_ng=0,
is_training=False) # 100
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=256,
shuffle=True,
num_workers=4)
dnn = DNN(int(user_num), int(item_num), factor_num=8,
indi=indi) ##################
dnn.cuda()
print(dnn)
#complexity = get_total_params(dnn.cuda(), (220, 30, 30)) ########todo: change the inpupt size
# Loss and optimizer 3.定义损失函数， 使用的是BCEWithLogitsLoss
criterion = nn.BCEWithLogitsLoss()
criterion = criterion.to(torch_device)
# 4.定义迭代优化算法， 使用的是Adam
learning_rate = 0.002 #########################
optimizer = torch.optim.Adam(dnn.parameters(),
lr=learning_rate) ##########
loss_dict = []
num_epochs = train_loader.__len__()
# Train the model 5. 迭代训练
dnn.train()
train_loader.dataset.ng_sample()
for i, data in enumerate(train_loader, 0):
# Convert numpy arrays to torch tensors 5.1 准备tensor的训练数据和标签
user = data[0].cuda()
item = data[1].cuda()
label = data[2].float().cuda()
# Forward pass 5.2 前向传播计算网络结构的输出结果
optimizer.zero_grad()
dnn.zero_grad() ##########
output = dnn(user, item)
# 5.3 计算损失函数
loss = criterion(output, label)
loss.cuda()
# Backward and optimize 5.4 反向传播更新参数
loss.backward()
optimizer.step()
# 可选 5.5 打印训练信息和保存loss
loss_dict.append(loss.item())
if (i + 1) % 5000 == 0:
print('Epoch [{}/{}], Loss: {:.4f}'.format(
i + 1, num_epochs, loss.item()))
# evaluate
dnn.eval()
#test_loss_dict = []
# every user have 99 negative items and one positive items，so batch_size=100
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=99 + 1,
shuffle=False,
num_workers=2)
#test_loader.dataset.ng_sample()
#for i, data in enumerate(test_loader, 0):
#user = data[0].cuda()
#item = data[1].cuda()
#label = data[2].float().cuda()
#output = dnn(user, item)
#loss = criterion(output, label)
#loss = loss.cuda()
#test_loss_dict.append(loss.item())
HR = utils.metricsHR(dnn, test_loader, top_k=10)
NDCG = utils.metricsNDCG(dnn, test_loader, top_k=10)
#mean_test_loss = np.mean(test_loss_dict)
#std_test_loss = np.std(test_loss_dict)
print("HR:{},NDCG:{}".format(HR, NDCG))
return HR, NDCG