link

0. Why save and load models

After training the model with data, an ideal model is obtained , But in practical application, it is impossible to train first and then use , So you have to save the trained model first , Then load it when you need it and use it directly . The essence of a model is a pile of parameters stored in some structure , So there are two ways to save , One way is to save the whole model directly , Then directly load the whole model , But this will consume more memory ; The other is to save only the parameters of the model , When used later, create a new model with the same structure , Then import the saved parameters into the new model .

1. Implementation methods of two cases
（1） Save only the model parameter dictionary （ recommend ）

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

（2） Save the entire model

# preservation 
torch.save(the_model, PATH)
# Read 
the_model = torch.load(PATH)

3. Save only model parameters （ Example ）

pytorch Will put the parameters of the model in a dictionary , And all we have to do is save this dictionary , Then call .
For example, design a single layer LSTM Network of , And then training , After training, save the parameter Dictionary of the model , Save as... Under the same folder rnn.pt file ：

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


#  Save the model 
torch.save(rnn.state_dict(), 'rnn.pt')

After saving, use the trained model to process the data ：

#  Test the saved model 
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)

Here's an explanation , When you save the model rnn.state_dict() Express rnn The parameter Dictionary of this model , When testing the saved model, first load the parameter Dictionary m_state_dict = torch.load(‘rnn.pt’);

Then instantiate one LSTM Antithetic image , Here, we need to ensure that the parameters passed in are consistent with the instantiation rnn Is the same as when the object is passed in , That is, the structure is the same new_m = LSTM(input_size=1, hidden_size=10, num_layers=2).to(device);

Here are the parameters loaded before passing in the new model new_m.load_state_dict(m_state_dict);

Finally, we can use this model to process the data predict = new_m(test_tensor)

4. Save the whole model （ Example ）

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, :])  #  Take the last column as 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


#  Save the model 

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

After saving, use the trained model to process the data ：

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