RNN、LSTM、GRU等模型使用GPU时的常见错误

错误信息：

RuntimeError: Input and hidden tensors are not at the same device, found input tensor at cuda:0 and hidden tensor at cpu

错误提示说，输入的张量在cuda中，但模型的隐藏层在CPU中，但我们检查模型，发现模型已经被放入到了cuda中了

# 实例化模型
model = LSTM(input_dim=input_dim, hidden_dim=hidden_dim, output_dim=output_dim, num_layers=num_layers)

# GPU加速
if torch.cuda.is_available():
	device = 'cuda:0'
    model = model.to(device)
    trainX = trainX.to(device)
    trainY = trainY.to(device)
    testX = testX.to(device)
    testY = testY.to(device)

究竟问题出在哪里了呢？

这时我们检查一下模型类

input_dim = 5      # 数据的特征数
hidden_dim = 32    # 隐藏层的神经元个数
num_layers = 2     # LSTM的层数
output_dim = 1     # 预测值的特征数
                   
class LSTM(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_layers, output_dim):
        super(LSTM, self).__init__()
        # Hidden dimensions
        self.hidden_dim = hidden_dim

        # Number of hidden layers
        self.num_layers = num_layers

        # Building your LSTM
        # batch_first=True causes input/output tensors to be of shape (batch_dim, seq_dim, feature_dim)
        self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, batch_first=True)

        # Readout layer 在LSTM后再加一个全连接层，因为是回归问题，所以不能在线性层后加激活函数
        self.fc = nn.Linear(hidden_dim, output_dim) 

    def forward(self, x):
        # Initialize hidden state with zeros   
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).requires_grad_() 
        # 这里x.size(0)就是batch_size

        # Initialize cell state
        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim).requires_grad_()

        # One time step
        # We need to detach as we are doing truncated backpropagation through time (BPTT)
        # If we don't, we'll backprop all the way to the start even after going through another batch
        out, (hn, cn) = self.lstm(x, (h0.detach(), c0.detach()))

        out = self.fc(out) 

        return out

发现在forward函数中，出现了h0和c0两个变量，这是给LSTM最开始的时步用的，这两个变量却是在CPU中，但它们不是模型的参数，即便后面使用了model = model.to(device)命令，也无法将这两个变量放入到GPU中。

将上述程序改成

if torch.cuda.is_available():
	device = 'cuda:0'
    
trainX = trainX.to(device)
trainY = trainY.to(device)
testX = testX.to(device)
testY = testY.to(device)

input_dim = 6      # 数据的特征数
hidden_dim = 32    # 隐藏层的神经元个数
num_layers = 2     # LSTM的层数
output_dim = 1     # 预测值的特征数
                   
class LSTM(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_layers, output_dim):
        super(LSTM, self).__init__()
        # Hidden dimensions
        self.hidden_dim = hidden_dim

        # Number of hidden layers
        self.num_layers = num_layers

        # Building your LSTM
        # batch_first=True causes input/output tensors to be of shape (batch_dim, seq_dim, feature_dim)
        self.lstm = nn.LSTM(input_dim, hidden_dim, num_layers, batch_first=True)

        # Readout layer 在LSTM后再加一个全连接层，因为是回归问题，所以不能在线性层后加激活函数
        self.fc = nn.Linear(hidden_dim, output_dim) 

    def forward(self, x):
        # Initialize hidden state with zeros   
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim, device=x.device).requires_grad_() 
        # 这里x.size(0)就是batch_size

        # Initialize cell state
        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_dim, device=x.device).requires_grad_()

        # One time step
        # We need to detach as we are doing truncated backpropagation through time (BPTT)
        # If we don't, we'll backprop all the way to the start even after going through another batch
        out, (hn, cn) = self.lstm(x, (h0.detach(), c0.detach()))

        out = self.fc(out) 

        return out

程序可以正常运行。