5

我有一个具有以下结构的神经网络：

class myNetwork(nn.Module):
    def __init__(self):
        super(myNetwork, self).__init__()
        self.bigru = nn.GRU(input_size=2, hidden_size=100, batch_first=True, bidirectional=True)
        self.fc1 = nn.Linear(200, 32)
        torch.nn.init.xavier_uniform_(self.fc1.weight)
        self.fc2 = nn.Linear(32, 2)
        torch.nn.init.xavier_uniform_(self.fc2.weight)

我需要通过重置神经网络的参数将模型恢复到未学习状态。我可以nn.Linear使用以下方法对图层执行此操作：

def reset_weights(self):
    torch.nn.init.xavier_uniform_(self.fc1.weight)
    torch.nn.init.xavier_uniform_(self.fc2.weight)

但是，要重置nn.GRU图层的权重，我找不到任何这样的片段。

我的问题是如何重置nn.GRU图层？任何其他重置网络的方式也可以。任何帮助表示赞赏。

4

3 回答 3

4

您可以reset_parameters在图层上使用方法。正如这里给出的

for layer in model.children():
   if hasattr(layer, 'reset_parameters'):
       layer.reset_parameters()

或者另一种方法是先保存模型，然后重新加载模块状态。使用torch.save并torch.load 查看文档以获取更多信息或保存和加载模型

于 2020-08-28T06:11:36.040 回答
1

这是带有运行示例的代码：

def lp_norm(mdl: nn.Module, p: int = 2) -> Tensor:
    lp_norms = [w.norm(p) for name, w in mdl.named_parameters()]
    return sum(lp_norms)

def reset_all_weights(model: nn.Module) -> None:
    """
    refs:
        - https://discuss.pytorch.org/t/how-to-re-set-alll-parameters-in-a-network/20819/6
        - https://stackoverflow.com/questions/63627997/reset-parameters-of-a-neural-network-in-pytorch
        - https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    """

    @torch.no_grad()
    def weight_reset(m: nn.Module):
        # - check if the current module has reset_parameters & if it's callabed called it on m
        reset_parameters = getattr(m, "reset_parameters", None)
        if callable(reset_parameters):
            m.reset_parameters()

    # Applies fn recursively to every submodule see: https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    model.apply(fn=weight_reset)


def reset_all_linear_layer_weights(model: nn.Module) -> nn.Module:
    """
    Resets all weights recursively for linear layers.

    ref:
        - https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    """

    @torch.no_grad()
    def init_weights(m):
        if type(m) == nn.Linear:
            m.weight.fill_(1.0)

    # Applies fn recursively to every submodule see: https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    model.apply(init_weights)


def reset_all_weights_with_specific_layer_type(model: nn.Module, modules_type2reset) -> nn.Module:
    """
    Resets all weights recursively for linear layers.

    ref:
        - https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    """

    @torch.no_grad()
    def init_weights(m):
        if type(m) == modules_type2reset:
            # if type(m) == torch.nn.BatchNorm2d:
            #     m.weight.fill_(1.0)
            m.reset_parameters()

    # Applies fn recursively to every submodule see: https://pytorch.org/docs/stable/generated/torch.nn.Module.html
    model.apply(init_weights)


# -- tests

def reset_params_test():
    import torchvision.models as models
    from uutils.torch_uu import lp_norm

    resnet18 = models.resnet18(pretrained=True)
    resnet18_random = models.resnet18(pretrained=False)

    print(f'{lp_norm(resnet18)=}')
    print(f'{lp_norm(resnet18_random)=}')
    print(f'{lp_norm(resnet18)=}')
    reset_all_weights(resnet18)
    print(f'{lp_norm(resnet18)=}')


if __name__ == '__main__':
    reset_params_test()
    print('Done! \a\n')

输出：

lp_norm(resnet18)=tensor(517.5472, grad_fn=<AddBackward0>)
lp_norm(resnet18_random)=tensor(668.3687, grad_fn=<AddBackward0>)
lp_norm(resnet18)=tensor(517.5472, grad_fn=<AddBackward0>)
lp_norm(resnet18)=tensor(476.0836, grad_fn=<AddBackward0>)
Done!

我假设这是可行的，因为我为预训练的网络计算了两次范数，并且在调用重置之前两次都是相同的。

虽然我很不高兴，但我必须承认它并不接近随机网络的规范，但我认为这已经足够好了。

同样：https ://discuss.pytorch.org/t/how-to-re-set-alll-parameters-in-a-network/20819/11

于 2021-11-09T22:05:34.493 回答
0

pytorch的新手，我想知道这是否可以解决:)

假设来自 torch.nn.module 的模型固有属性，

将其重置为零：

dic = Model.state_dict()
for k in dic:
    dic[k] *= 0
Model.load_state_dict(dic)
del(dic)

随机重置它

dic = Model.state_dict()
for k in dic:
    dic[k] = torch.randn(dic[k].size())  
Model.load_state_dict(dic)
del(dic)
于 2021-07-19T15:26:34.107 回答