python - 如何在 PyTorch 中更新神经网络的参数？

Question

假设我想将 PyTorch(继承自 torch.nn.Module 的类的实例)中神经网络的所有参数乘以 0.9。我该怎么做？

Answer 1

让 net 成为神经网络 nn.Module 的实例。然后，将所有参数乘以 0.9:

state_dict = net.state_dict()

for name, param in state_dict.items():
    # Transform the parameter as required.
    transformed_param = param * 0.9

    # Update the parameter.
    param.copy_(transformed_param)

---

如果您只想更新权重而不是每个参数:

state_dict = net.state_dict()

for name, param in state_dict.items():
    # Don't update if this is not a weight.
    if not "weight" in name:
        continue
    
    # Transform the parameter as required.
    transformed_param = param * 0.9

    # Update the parameter.
    param.copy_(transformed_param)