将循环层附加到具有不同隐藏大小的 PyTorch LSTM 模型

Question

我正在使用 PyTorch 开发用于序列分析的 BI-LSTM 模型。我正在使用torch.nn.LSTM. 使用该模块，您可以拥有多个层，只需传递一个参数num_layers作为层数（例如，num_layers=2）。然而，它们都是相同的hidden_size，这对我来说部分好，我只是想让它们都一样，hidden_size 但最后一层的大小不同。基本示例如下：

rnn = nn.LSTM(input_size=10, hidden_size=20, num_layers=2)

inp = torch.randn(5, 3, 10)

h0 = torch.randn(2, 3, 20)

c0 = torch.randn(2, 3, 20)

output, (hn, cn) = rnn(inp, (h0, c0))

输出昏暗为 ( 5, 3,20)

一个解决方案（但对我不利）是实现额外的模型，该模型输出我需要的维度并从第一个模型中获取输入，例如：

rnn_two = nn.LSTM(input_size=20, hidden_size=2)

output2, _ = rnn_two(output)

但是，我不想这样做，因为我使用并行化模型DataParallel，所以我需要全部成为一个包。我希望找到类似于 keras 的东西，例如：

rnn.add(LSTM, hidden_size=2)

我检查了LSTM 源代码，但找不到我需要的。

有什么建议？

Answer 1

如果我没记错的话，可以这样做：import torch.nn as nnimport torch.nn.functional as Fclass RnnWith2HiddenSizesModel(nn.Module):&nbsp; &nbsp; def __init__(self):&nbsp; &nbsp; &nbsp; &nbsp; super(RnnWith2HiddenSizesModel, self).__init__()&nbsp; &nbsp; &nbsp; &nbsp; self.rnn = nn.LSTM(input_size=10, hidden_size=20, num_layers=2)&nbsp; &nbsp; &nbsp; &nbsp; self.rnn_two = nn.LSTM(input_size=20, hidden_size=2)&nbsp; &nbsp; def forward(self, inp, hc):&nbsp; &nbsp; &nbsp; &nbsp; output, _ = self.rnn(inp, hc)&nbsp; &nbsp; &nbsp; &nbsp; output2, _ = self.rnn_two(output)&nbsp; &nbsp; &nbsp; &nbsp; return output2inp = torch.randn(5, 3, 10)h0 = torch.randn(2, 3, 20)c0 = torch.randn(2, 3, 20)rnn = RnnWith2HiddenSizesModel()output = RnnWith2HiddenSizesModel()(inp, (h0, c0))tensor([[[-0.0305,&nbsp; 0.0327],&nbsp; &nbsp; &nbsp;[-0.1534, -0.1193],&nbsp; &nbsp; &nbsp;[-0.1393,&nbsp; 0.0474]],&nbsp; &nbsp; [[-0.0370,&nbsp; 0.0519],&nbsp; &nbsp; &nbsp;[-0.2081, -0.0693],&nbsp; &nbsp; &nbsp;[-0.1809,&nbsp; 0.0826]],&nbsp; &nbsp; [[-0.0561,&nbsp; 0.0731],&nbsp; &nbsp; &nbsp;[-0.2307, -0.0229],&nbsp; &nbsp; &nbsp;[-0.1780,&nbsp; 0.0901]],&nbsp; &nbsp; [[-0.0612,&nbsp; 0.0891],&nbsp; &nbsp; &nbsp;[-0.2378,&nbsp; 0.0164],&nbsp; &nbsp; &nbsp;[-0.1760,&nbsp; 0.0929]],&nbsp; &nbsp; [[-0.0660,&nbsp; 0.1023],&nbsp; &nbsp; &nbsp;[-0.2176,&nbsp; 0.0508],&nbsp; &nbsp; &nbsp;[-0.1611,&nbsp; 0.1053]]], grad_fn=<CatBackward>)

Answer 2

尽管@Mikhail Berlinkov 的回答按需要工作这一事实，但它并没有推广到一般情况（甚至在问题中都没有要求），为此我想提出第二个解决方案：import torchimport torch.nn as nnimport torch.nn.functional as Ffrom functools import reduceclass RNNModel(nn.Module):&nbsp; &nbsp; def __init__(self, *models):&nbsp; &nbsp; &nbsp; &nbsp; super(RNNModel, self).__init__()&nbsp; &nbsp; &nbsp; &nbsp; self.models = models&nbsp; &nbsp; def forward(self, inp):&nbsp; &nbsp; &nbsp; &nbsp; return reduce(lambda arg, model: model(arg, None)[0], self.models, inp)并且可以称为：rnn = nn.LSTM(input_size=10, hidden_size=20, num_layers=2)rnn_two = nn.LSTM(input_size=20, hidden_size=2)inp = torch.randn(5, 3, 10)rnn_model = RNNModel(rnn, rnn_two)output = rnn_model(inp)output.shape等于预期（即，5,3,2）