循环神经网络RNN、RNNCell、GRUCell

python 复制代码
###################### RNNCell的使用 ####################################

#初始化完全一样
# input_size - The number of expected features in the input x
# hidden_size - The number of features in the hidden state h
# num_layers - Number of recurrent layers, default 1

#但是forward是不一样的
#ht = rnncell(xt,ht_1)
x = torch.randn(10, 1, 100)
rnnCell = nn.RNNCell(input_size=100, hidden_size=30)
ht = torch.zeros(1, 30)
out = []
for xt in x:
    ht = rnnCell(xt, ht)
    out.append(ht)
print('输出总时间步:', len(out),',每个时间步输出的形状:',out[0].shape)


#####################lstm使用#########################
# 多层RNN的例子
rnn = nn.LSTM(input_size=100, hidden_size=20, num_layers=4)
print(rnn)
x = torch.randn(10, 3, 100)
out, (h, c) = rnn(x)
print(out.shape, h.shape, c.shape)
#输出
#torch.Size([10, 3, 20]) torch.Size([4, 3, 20]) torch.Size([4, 3, 20])

x = torch.rand(3,3,4)
id = [0,1]
c = x[id].transpose(0,1)
print(x)
print(c)
print(c.shape)
#实例:正弦曲线预测

input_size = 1
hidden_size = 256
output_size = 1
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.rnn = nn.LSTM(input_size=input_size, hidden_size=hidden_size, num_layers=2)
        self.linear = nn.Linear(hidden_size, output_size)

    def forward(self, x):
        _, (h, c) = self.rnn(x)
        h = h[-1]
        h = self.linear(h)
        return h


model = Net()
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)


num_time_steps =200
start = 0 #在0-3之间随机初始化
time_steps = np.linspace(start, start+20, num_time_steps)
data = np.sin(time_steps)
data = data.reshape(num_time_steps)

horizon_len = 10 #用于估计的时间序列长度
list = [1, 2, 3]


for iter in range(500):
    batch_size = 32
    s = random.sample(range(num_time_steps-horizon_len-1), batch_size)
    train_data_x = []
    train_data_y = []
    for s_i in s:
        train_data_x.append(data[s_i:s_i+horizon_len])
        train_data_y.append(data[s_i + horizon_len])
    train_data_x = torch.tensor(train_data_x, dtype=torch.float32).T
    train_data_x = train_data_x.unsqueeze(2).to(torch.float32)
    train_data_y = torch.tensor(train_data_y, dtype=torch.float64).reshape(-1, batch_size, 1).to(torch.float32)


    output = model(train_data_x)

    loss = criterion(output, train_data_y)
    model.zero_grad()
    loss.backward()
    optimizer.step()

    if iter % 50 == 0:
        print("Iteration: {} loss {}".format(iter, loss.item()))


num_time_steps =200
start = 0
time_steps = np.linspace(start, start+20, num=num_time_steps)
data = np.sin(time_steps)

predictions1 = []

h = torch.zeros(1, 1, hidden_size)
c = torch.zeros(1, 1, hidden_size)

input = deque(data[0:horizon_len], maxlen=horizon_len)
predictions1.extend(data[0:horizon_len])
for i in range(horizon_len, len(data)*5):
    input_tensor = torch.tensor(input).to(torch.float32)
    input_tensor = input_tensor.view(horizon_len, 1, 1)
    pred= model(input_tensor)
    input.append(pred.detach().numpy().ravel()[0])
    predictions1.append(pred.detach().numpy().ravel()[0])

plt.figure()
plt.plot(data)
plt.plot(predictions1)
plt.legend(['sin wave','pred1'])
plt.show()
相关推荐
AI人工智能+1 小时前
炫光活体检测技术:通过光学技术实现高效、安全的身份验证,有效防御多种伪造手段。
人工智能·深度学习·人脸识别·活体检测
东方佑2 小时前
打破常规:“无注意力”神经网络为何依然有效?
人工智能·深度学习·神经网络
Francek Chen2 小时前
【深度学习计算机视觉】03:目标检测和边界框
人工智能·pytorch·深度学习·目标检测·计算机视觉·边界框
九章云极AladdinEdu2 小时前
AI集群全链路监控:从GPU微架构指标到业务Metric关联
人工智能·pytorch·深度学习·架构·开源·gpu算力
惯导马工2 小时前
【论文导读】IDOL: Inertial Deep Orientation-Estimation and Localization
深度学习·算法
爱学习的茄子2 小时前
Function Call:让AI从文本生成走向智能交互的技术革命
前端·深度学习·openai
CoovallyAIHub3 小时前
基于YOLO集成模型的无人机多光谱风电部件缺陷检测
深度学习·算法·计算机视觉
CoovallyAIHub3 小时前
几十个像素的小目标,为何难倒无人机?LCW-YOLO让无人机小目标检测不再卡顿
深度学习·算法·计算机视觉
IMER SIMPLE4 小时前
人工智能-python-深度学习-经典网络模型-LeNets5
人工智能·python·深度学习
却道天凉_好个秋4 小时前
深度学习(五):过拟合、欠拟合与代价函数
人工智能·深度学习·过拟合·欠拟合·代价函数