在PyTorch中是否存在干净和可扩展的LSTM实现？

Question

我想自己创建一个LSTM类，但是我不想从头开始重写经典的LSTM函数。

深入研究PyTorch的代码，我只发现至少涉及3-4个具有继承的类的脏实现：

1. https://github.com/pytorch/pytorch/blob/98c24fae6b6400a7d1e13610b20aa05f86f77070/torch/nn/modules/rnn.py#L323
2. https://github.com/pytorch/pytorch/blob/98c24fae6b6400a7d1e13610b20aa05f86f77070/torch/nn/modules/rnn.py#L12
3. 函数/rnn.py#L 297

一个干净的PyTorch实现是否存在于某个地方？任何联系都会有帮助。

例如，我知道LSTM的干净实现存在于TensorFlow中，但我需要派生出PyTorch one。

作为一个明确的例子，我正在搜索的是一个像这一样干净的实现，但是在PyTorch中

Answer 1

我发现最好的实现就是在这里

变种/lstm.py

它甚至实现了四种不同的反复式辍学，这是非常有用的！

如果你把辍学的部分拿走，你就会得到

import math
import torch as th
import torch.nn as nn

class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, bias=True):
        super(LSTM, self).__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.bias = bias
        self.i2h = nn.Linear(input_size, 4 * hidden_size, bias=bias)
        self.h2h = nn.Linear(hidden_size, 4 * hidden_size, bias=bias)
        self.reset_parameters()

    def reset_parameters(self):
        std = 1.0 / math.sqrt(self.hidden_size)
        for w in self.parameters():
            w.data.uniform_(-std, std)

    def forward(self, x, hidden):
        h, c = hidden
        h = h.view(h.size(1), -1)
        c = c.view(c.size(1), -1)
        x = x.view(x.size(1), -1)

        # Linear mappings
        preact = self.i2h(x) + self.h2h(h)

        # activations
        gates = preact[:, :3 * self.hidden_size].sigmoid()
        g_t = preact[:, 3 * self.hidden_size:].tanh()
        i_t = gates[:, :self.hidden_size]
        f_t = gates[:, self.hidden_size:2 * self.hidden_size]
        o_t = gates[:, -self.hidden_size:]

        c_t = th.mul(c, f_t) + th.mul(i_t, g_t)

        h_t = th.mul(o_t, c_t.tanh())

        h_t = h_t.view(1, h_t.size(0), -1)
        c_t = c_t.view(1, c_t.size(0), -1)
        return h_t, (h_t, c_t)

PS:存储库包含更多的LSTM和其他RNN变体：

https://github.com/pytorch/benchmark/tree/master/rnns/benchmarks。

看看它，也许你想到的分机已经在那里了！

编辑：

正如注释中提到的，您可以将上面的LSTM单元格包装为处理顺序输出：

import math
import torch as th
import torch.nn as nn


class LSTMCell(nn.Module):

    def __init__(self, input_size, hidden_size, bias=True):
        # As before

    def reset_parameters(self):
        # As before

    def forward(self, x, hidden):

        if hidden is None:
            hidden = self._init_hidden(x)

        # Rest as before

    @staticmethod
    def _init_hidden(input_):
        h = th.zeros_like(input_.view(1, input_.size(1), -1))
        c = th.zeros_like(input_.view(1, input_.size(1), -1))
        return h, c


class LSTM(nn.Module):

    def __init__(self, input_size, hidden_size, bias=True):
        super().__init__()
        self.lstm_cell = LSTMCell(input_size, hidden_size, bias)

    def forward(self, input_, hidden=None):
        # input_ is of dimensionalty (1, time, input_size, ...)

        outputs = []
        for x in torch.unbind(input_, dim=1):
            hidden = self.lstm_cell(x, hidden)
            outputs.append(hidden[0].clone())

        return torch.stack(outputs, dim=1)

我没有测试代码，因为我使用的是convLSTM实现。如果有什么不对劲，请告诉我。

更新:固定链接。

Answer 2

我做了一个简单而通用的框架来定制LSTMs：https://github.com/daehwannam/pytorch-rnn-util

您可以通过设计LSTM单元并将它们提供给LSTMFrame来实现自定义LSTM。自定义LSTM的一个例子是包中的LayerNormLSTM：

# snippet from rnn_util/seq.py
class LayerNormLSTM(LSTMFrame):
    def __init__(self, input_size, hidden_size, num_layers=1, dropout=0, r_dropout=0, bidirectional=False, layer_norm_enabled=True):
        r_dropout_layer = nn.Dropout(r_dropout)
        rnn_cells = tuple(
            tuple(
                LayerNormLSTMCell(
                    input_size if layer_idx == 0 else hidden_size * (2 if bidirectional else 1),
                    hidden_size,
                    dropout=r_dropout_layer,
                    layer_norm_enabled=layer_norm_enabled)
                for _ in range(2 if bidirectional else 1))
            for layer_idx in range(num_layers))

        super().__init__(rnn_cells, dropout, bidirectional)

LayerNormLSTM有PyTorch标准LSTM的关键选项和附加选项，r_dropout和layer_norm_enabled

# example.py
import torch
import rnn_util


bidirectional = True
num_directions = 2 if bidirectional else 1

rnn = rnn_util.LayerNormLSTM(10, 20, 2, dropout=0.3, r_dropout=0.25,
                             bidirectional=bidirectional, layer_norm_enabled=True)
# rnn = torch.nn.LSTM(10, 20, 2, bidirectional=bidirectional)

input = torch.randn(5, 3, 10)
h0 = torch.randn(2 * num_directions, 3, 20)
c0 = torch.randn(2 * num_directions, 3, 20)
output, (hn, cn) = rnn(input, (h0, c0))

print(output.size())