Pytorch 动态层数？

Question

我正在尝试指定动态数量的图层，但我似乎做错了。我的问题是，当我在这里定义 100 层时，我会在前进的步骤中出错。但是当我正确定义图层时它会起作用吗？下面的简化示例

class PredictFromEmbeddParaSmall(LightningModule):
    def __init__(self, hyperparams={'lr': 0.0001}):
        super(PredictFromEmbeddParaSmall, self).__init__()
        #Input is something like tensor.size=[768*100]
        self.TO_ILLUSTRATE = nn.Linear(768, 5)
        self.enc_ref=[]
        for i in range(100):
            self.enc_red.append(nn.Linear(768, 5))
        # gather the layers output sth
        self.dense_simple1 = nn.Linear(5*100, 2)
        self.output = nn.Sigmoid()
    def forward(self, x):
        # first input to enc_red
        x_vecs = []
        for i in range(self.para_count):
            layer = self.enc_red[i]
            # The first dim is the batch size here, output is correct
            processed_slice = x[:, i * 768:(i + 1) * 768]
            # This works and give the out of size 5
            rand = self.TO_ILLUSTRATE(processed_slice)
            #This will fail? Error below
            ret = layer(processed_slice)
            #more things happening we can ignore right now since we fail earlier

我在执行“ret = layer.forward(processed_slice)”时遇到这个错误

RuntimeError: Expected object of device type cuda but got device type cpu for argument #1 'self' in call to _th_addmm

有没有更聪明的编程方法？或者解决错误？

Answer 1

您应该使用 pytorch 中的 ModuleList 而不是列表：https://pytorch.org/docs/master/generated/torch.nn.ModuleList.html。这是因为 Pytorch 必须保留包含模型所有模块的图表，如果您只是将它们添加到列表中，则它们不会在图表中正确索引，从而导致您遇到错误。

你的代码应该是这样的：

class PredictFromEmbeddParaSmall(LightningModule):
    def __init__(self, hyperparams={'lr': 0.0001}):
        super(PredictFromEmbeddParaSmall, self).__init__()
        #Input is something like tensor.size=[768*100]
        self.TO_ILLUSTRATE = nn.Linear(768, 5)
        self.enc_ref=nn.ModuleList()                     # << MODIFIED LINE <<
        for i in range(100):
            self.enc_red.append(nn.Linear(768, 5))
        # gather the layers output sth
        self.dense_simple1 = nn.Linear(5*100, 2)
        self.output = nn.Sigmoid()
    def forward(self, x):
        # first input to enc_red
        x_vecs = []
        for i in range(self.para_count):
            layer = self.enc_red[i]
            # The first dim is the batch size here, output is correct
            processed_slice = x[:, i * 768:(i + 1) * 768]
            # This works and give the out of size 5
            rand = self.TO_ILLUSTRATE(processed_slice)
            #This will fail? Error below
            ret = layer(processed_slice)
            #more things happening we can ignore right now since we fail earlier

那么它应该可以正常工作了！

编辑：替代方法。

除了使用ModuleList，你也可以只使用nn.Sequential，这样你就可以避免在正向传递中使用for循环。这也意味着您将无法访问中间激活，因此如果您需要它们，这不是适合您的解决方案。

class PredictFromEmbeddParaSmall(LightningModule):
    def __init__(self, hyperparams={'lr': 0.0001}):
        super(PredictFromEmbeddParaSmall, self).__init__()
        #Input is something like tensor.size=[768*100]
        self.TO_ILLUSTRATE = nn.Linear(768, 5)
        self.enc_ref=[]
        for i in range(100):
            self.enc_red.append(nn.Linear(768, 5))

        self.enc_red = nn.Seqential(*self.enc_ref)       # << MODIFIED LINE <<
        # gather the layers output sth
        self.dense_simple1 = nn.Linear(5*100, 2)
        self.output = nn.Sigmoid()
    def forward(self, x):
        # first input to enc_red
        x_vecs = []
        out = self.enc_red(x)                            # << MODIFIED LINE <<

Pytorch 动态层数？

Pytorch dynamic amount of Layers?

pytorch

pytorch-lightning