Pytorch如何保存训练好的模型

1、（1）只保存模型参销糟数字典（推荐）

代码：

#保存

torch.save(the_model.state_dict(), PATH)
#读取
the_model = TheModelClass(*args, **kwargs)
the_model.load_state_dict(torch.load(PATH))

 民拒

2、（2）保存整个级阅迅模型

代码：

#保存

torch.save(the_model, PATH)
#读取
the_model = torch.load(PATH)

1、pytorch会把模型的参数放在一个字典里面，而我们所要做的就是将这个字典保存，然后再调用。比如说设计一个单层LSTM的网络，然后进行训练，训练完之后将模型的参数字典进行保存，保存为同文件夹下面的rnn.pt文件：

代码：

class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, num_layers):
        super(LSTM, self).__init__()
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_size, 1)

    def forward(self, x):
        # Set initial states
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)
         # 2 for bidirection
        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)
        # Forward propagate LSTM
        out, _ = self.lstm(x, (h0, c0))  
        # out: tensor of shape (batch_size, seq_length, hidden_size*2)
        out = self.fc(out)
        return out

rnn = LSTM(input_size=1, hidden_size=10, num_layers=2).to(device)

# optimize all cnn parameters
optimizer = torch.optim.Adam(rnn.parameters(), lr=0.001)  
# the target label is not one-hotted
loss_func = nn.MSELoss()  

for epoch in range(1000):
    output = rnn(train_tensor)  # cnn output`
    loss = loss_func(output, train_labels_tensor)  # cross entropy loss
    optimizer.zero_grad()  # clear gradients for this training step
    loss.backward()  # backpropagation, compute gradients
    optimizer.step()  # apply gradients
    output_sum = output

# 保存模型
torch.save(rnn.state_dict(), 'rnn.pt')

2、保存完之后利用这个训练完的模型对数据进行处理：

代码：

# 测试所保存的模型

m_state_dict = torch.load('rnn.pt')
new_m = LSTM(input_size=1, hidden_size=10, num_layers=2).to(device)
new_m.load_state_dict(m_state_dict)
predict = new_m(test_tensor)

3、这里做一下说明，在保存模型的时候rnn.state_dict()表示rnn这个模型的参数字典，在测试所保存的模型时要先将这个参数字典加载一下:

m_state_dict = torch.load('rnn.pt')；

然后再实例化一个LSTM对像，这里要保证传入的参数跟实例化rnn是传入的对象时一样的，即结构相同:

new_m = LSTM(input_size=1, hidden_size=10, num_layers=2).to(device)；

下面是给这个新的模型传入之前加载的参数:

new_m.load_state_dict(m_state_dict)；

最后就可以利用这个模型处理数据了:

predict = new_m(test_tensor)

1、代码：

class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, num_layers):
        super(LSTM, self).__init__()
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_size, 1)

    def forward(self, x):
        # Set initial states
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)  # 2 for bidirection
        c0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)

        # Forward propagate LSTM
        out, _ = self.lstm(x, (h0, c0))  # out: tensor of shape (batch_size, seq_length, hidden_size*2)
        # print("output_in=", out.shape)
        # print("fc_in_shape=", out[:, -1, :].shape)
        # Decode the hidden state of the last time step
        # out = torch.cat((out[:, 0, :], out[-1, :, :]), axis=0)
        # out = self.fc(out[:, -1, :])  # 取最后一列为out
        out = self.fc(out)
        return out

rnn = LSTM(input_size=1, hidden_size=10, num_layers=2).to(device)
print(rnn)

optimizer = torch.optim.Adam(rnn.parameters(), lr=0.001)  # optimize all cnn parameters
loss_func = nn.MSELoss()  # the target label is not one-hotted

for epoch in range(1000):
    output = rnn(train_tensor)  # cnn output`
    loss = loss_func(output, train_labels_tensor)  # cross entropy loss
    optimizer.zero_grad()  # clear gradients for this training step
    loss.backward()  # backpropagation, compute gradients
    optimizer.step()  # apply gradients
    output_sum = output

# 保存模型

torch.save(rnn, 'rnn1.pt')

2、保存完之后利用这个训练完的模型对数据进行处理：

代码：

new_m = torch.load('rnn1.pt')

predict = new_m(test_tensor)