引言

本着“凡我不能创造的,我就不能理解”的思想,本系列文章会基于纯Python以及NumPy从零创建自己的深度学习框架,该框架类似PyTorch能实现自动求导。

要深入理解深度学习,从零开始创建的经验非常重要,从自己可以理解的角度出发,尽量不使用外部完备的框架前提下,实现我们想要的模型。本系列文章的宗旨就是通过这样的过程,让大家切实掌握深度学习底层实现,而不是仅做一个调包侠。

在前面的文章中,我们实现了多层、双向RNN。但是这几天一直在思考,这种实现方式是不是有问题。因为RNN的实现关乎后面ELMo和seq2seq,所以不得不重视。

双向RNN的实现方式

以两层双向RNN为例。我们之前实现的方式类似如下图所示:

这两张图片来自于:https://github.com/pytorch/pytorch/issues/4930#issuecomment-361851298

就是正向RNN和反向RNN可以看成是两个独立的两层RNN网络,最终拼接了它们的输出。但是总感觉双向RNN不会这么简单,带着这个疑问去拜读了双向RNN的论文1,得到下面的这张图片:

如果采用这种方式的话,那么两层双向RNN的实现应该像下图这样:

即第一层BRNN的输出同时考虑了正向和方向输出,将它们拼接在一起,作为第二层BRNN的输入。

但是这时遇到了一个问题,如果这样实现的话,那么输出的维度会怎样呢?BRNN中每层参数的维度会产生怎样的变化呢?

遇事不决找Torch,我们摸着PyTorch过河。

带着这个问题,我们去看PyTorch的文档,并查阅资料,梳理一下PyTorch实现的RNN(GRU、LSTM)中各种输入、输出、隐藏状态的维度。

理解RNN中的各种维度

以RNN为例,为什么不以最复杂的LSTM为例呢?因为LSTM参数过多,相比RNN太过复杂,不太容易理解。柿子要挑软的捏,我们理解了RNN,再去理解GRU或LSTM就会简单多了。

此图片参考了https://stackoverflow.com/a/48305882

从上图可以看出,在一个堆叠了ll l层的RNN中,output包含了最后一层RNN输出的所有隐藏状态;h_n包含了最后一个时间步上所有层的输出。

我们知道了它们的构成方式,下面看一下它们和上图中另外两个参数 inputh_0在不同类型的RNN中维度如何2

  • input RNN的输入序列。若batch_first=False,则其大小为(seq_len, batch, input_size);若batch_first=True,则其大小为(batch, seq_len, input_size)
  • h_0 RNN的初始隐藏状态,可以为空。大小为(num_layers * num_directions, batch, input_size)
  • output RNN最后一层所有时间步的输出。若batch_first=False,则其大小为(seq_len, batch, num_directions * hidden_size);若batch_first=True,则其大小为(batch, seq_len, num_directions * hidden_size)
  • h_nRNN中所有层最后一个时间步的隐藏状态。其大小为(num_layers * num_directions, batch, hidden_size)。不受batch_first的影响,其批次维度表现和batch_first=False一样。后面以代码实现的角度解释下为何这样,不代表官方的意图。

其中seq_len表示输入序列长度;batch表示批次大小;input_size表示输入的特征数量;num_layers 表示层数;num_directions表示方向个数,单向RNN时为1,双向RNN时为2hidden_size表示隐藏状态的特征数。

下面我们进行验证,首先看一下初始参数:

# 输入大小INPUT_SIZE = 2# 序列长度SEQ_LENGTH = 5# 隐藏大小HIDDEN_SIZE = 3# 批大小BATCH_SIZE = 4

以及输入:

inputs = Tensor.randn(BATCH_SIZE, SEQ_LENGTH, INPUT_SIZE)

简单RNN

简单RNN就是单向单层RNN:

rnn = nn.RNN(input_size=INPUT_SIZE, hidden_size=HIDDEN_SIZE, num_layers=1, batch_first=True)output, h_n = rnn(inputs)print(f'Input Shape: {inputs.shape} ')print(f'Output Shape: {output.shape} ')print(f'Hidden Shape: {h_n.shape} ')

inputs维度是我们预先定理好的,注意这里batch_first=True,所以inputs的第一个维度是批大小。

output来自最后一层所有时间步的输出,时间步长度为5,包含整个批次内4条数据,每条数据的输出维度为3,可以理解为3分类问题。

h_n来自单层最后一个时间步的隐藏状态,包含整个批次内4条数据,每条数据的输出维度为3

Input Shape: (4, 5, 2) Output Shape: (4, 5, 3) Hidden Shape: (1, 4, 3) 

堆叠RNN

如果将层数改成3,我们就得到了3层RNN堆叠在一起的架构,来看下此时outputh_n的维度会发生怎样的变化。

rnn = nn.RNN(input_size=INPUT_SIZE, hidden_size=HIDDEN_SIZE, num_layers=3, batch_first=True)output, h_n = rnn(inputs)print(f'Input Shape: {inputs.shape} ')print(f'Output Shape: {output.shape} ')print(f'Hidden Shape: {h_n.shape} ')
Input Shape: (4, 5, 2) Output Shape: (4, 5, 3) Hidden Shape: (3, 4, 3) 

output来自最后一层所有时间步的输出,时间步长度为5,包含整个批次内4条数据,每条数据的输出维度为3。其维度保持不变。

h_n来自所有三层最后一个时间步的隐藏状态,包含整个批次内4条数据,每条数据的输出维度为3。可以看到,其输出的第一个维度大小由1变成了3,因为包含了3层的结果。

双向RNN

传入bidirectional=True,并将层数改回单层。

rnn = nn.RNN(input_size=INPUT_SIZE, hidden_size=HIDDEN_SIZE, num_layers=1, batch_first=True, bidirectional=True)output, h_n = rnn(inputs)print(f'Input Shape: {inputs.shape} ')print(f'Output Shape: {output.shape} ')print(f'Hidden Shape: {h_n.shape} ')
Input Shape: (4, 5, 2) Output Shape: (4, 5, 6) Hidden Shape: (2, 4, 3) 

output来自最后一层所有时间步的输出,时间步长度为5,包含整个批次内4条数据,每条数据的输出维度为3,由于是双向,包含了两个方向上的结果,在此维度上进行堆叠,所以由3变成了6

h_n最后一个时间步的隐藏状态,包含整个批次内4条数据,每条数据的输出维度为3。第一个维度由1变成了2,因为在此维度上堆叠了双向的结果。

它们都包含了双向的结果,那如果想分别得到每个方向上的结果,要怎么做呢?

  • 对于output。若batch_first=True,将output按照out.reshape(shape=(batch, seq_len, num_directions, hidden_size))进行变形,正向和反向的维度值为别为01
  • 对于h_n,按照h_n.reshape(shape=(num_layers, num_directions, batch, hidden_size)),正向和反向的维度值为别为01

我们来对output进行拆分:

# batch_first=Trueoutput_reshaped = output.reshape((BATCH_SIZE, SEQ_LENGTH, 2, HIDDEN_SIZE))print("Shape of the output after directions are separated: ", output_reshaped.shape)# 分别获取正向和反向的输出output_forward = output_reshaped[:, :, 0, :]output_backward = output_reshaped[:, :, 1, :]print("Forward output Shape: ", output_forward.shape)print("Backward output Shape: ", output_backward.shape)
Shape of the output after directions are separated:(4, 5, 2, 3)Forward output Shape:(4, 5, 3)Backward output Shape:(4, 5, 3)

h_n进行拆分:

# 1: 层数 2: 方向数h_n_reshaped = h_n.reshape((1, 2, BATCH_SIZE, HIDDEN_SIZE))print("Shape of the hidden after directions are separated: ", h_n_reshaped.shape)h_n_forward = h_n_reshaped[:, 0, :, :]h_n_backward = h_n_reshaped[:, 1, :, :]print("Forward h_n Shape: ", h_n_forward.shape)print("Backward h_n Shape: ", h_n_backward.shape)
Shape of the hidden after directions are separated:(1, 2, 4, 3)Forward h_n Shape:(1, 4, 3)Backward h_n Shape:(1, 4, 3)

堆叠双向RNN

设置bidirectional=True,并将层数设成3层。

rnn = nn.RNN(input_size=INPUT_SIZE, hidden_size=HIDDEN_SIZE, num_layers=3, batch_first=True, bidirectional=True)output, h_n = rnn(inputs)print(f'Input Shape: {inputs.shape} ')print(f'Output Shape: {output.shape} ')print(f'Hidden Shape: {h_n.shape} ')
Input Shape: (4, 5, 2) Output Shape: (4, 5, 6) Hidden Shape: (6, 4, 3) 

output来自最后一层所有时间步的输出,时间步长度为5,包含整个批次内4条数据,每条数据的输出维度为3,由于是双向,包含了两个方向上的结果,在此维度上进行堆叠,所以由3变成了6

h_n来自所有三层最后一个时间步的隐藏状态,包含整个批次内4条数据,每条数据的输出维度为3。第一个维度由变成了6,因为三层输出在此维度上堆叠了双向的结果。

如果我们也对它们按方向进行拆分的话。

首先对output拆分:

# batch_first=Trueoutput_reshaped = output.reshape((BATCH_SIZE, SEQ_LENGTH, 2, HIDDEN_SIZE))print("Shape of the output after directions are separated: ", output_reshaped.shape)# 分别获取正向和反向的输出output_forward = output_reshaped[:, :, 0, :]output_backward = output_reshaped[:, :, 1, :]print("Forward output Shape: ", output_forward.shape)print("Backward output Shape: ", output_backward.shape)
Shape of the output after directions are separated:(4, 5, 2, 3)Forward output Shape:(4, 5, 3)Backward output Shape:(4, 5, 3)

其次对h_out拆分:

# 3: 层数 2: 方向数h_n_reshaped = h_n.reshape((3, 2, BATCH_SIZE, HIDDEN_SIZE))print("Shape of the hidden after directions are separated: ", h_n_reshaped.shape)h_n_forward = h_n_reshaped[:, 0, :, :]h_n_backward = h_n_reshaped[:, 1, :, :]print("Forward h_n Shape: ", h_n_forward.shape)print("Backward h_n Shape: ", h_n_backward.shape)
Shape of the hidden after directions are separated:(3, 2, 4, 3)Forward h_n Shape:(3, 4, 3)Backward h_n Shape:(3, 4, 3)

重构双向RNN的实现

我们按照对每层输出状态进行拼接的方式来重构多层双向RNN。

这里有一个问题是,由于我们对隐藏状态进行了拼接, 其维度变成了(n_steps, batch_size, num_directions * hidden_size)

受到了PyTorch官网启发:

  • ~RNN.weight_ih_l[k] – the learnable input-hidden weights of the k-th layer, of shape (hidden_size, input_size) for k = 0. Otherwise, the shape is (hidden_size, num_directions * hidden_size)
  • ~RNN.weight_hh_l[k] – the learnable hidden-hidden weights of the k-th layer, of shape (hidden_size, hidden_size)

所以,我们相应地改变输入到隐藏状态的维度:(hidden_size, num_directions * hidden_size)

我们说 h_n的输出维度不受batch_first的影响,其批次维度表现和batch_first=False一样。这是因为在实现时,为了统一,将input的时间步放到了第1个维度,将批大小放到中间,input就像batch_first=False一样,而隐藏状态的方式和它保持一致即可。

if self.batch_first:batch_size, n_steps, _ = input.shapeinput = input.transpose((1, 0, 2))# 将batch放到中间维度

下面看具体实现:

RNNCellBase

class RNNCellBase(Module):def reset_parameters(self) -> None:stdv = 1.0 / math.sqrt(self.hidden_size) if self.hidden_size > 0 else 0for weight in self.parameters():init.uniform_(weight, -stdv, stdv)def __init__(self, input_size, hidden_size: int, num_chunks: int, bias: bool = True, num_directions=1, reset_parameters=True, device=None, dtype=None) -> None:'''RNN单时间步的抽象:param input_size: 输入x的特征数:param hidden_size: 隐藏状态的特征数:param bias: 线性层是否包含偏置:param nonlinearity: 非线性激活函数 tanh | relu (mode = RNN)'''factory_kwargs = {'device': device, 'dtype': dtype}super(RNNCellBase, self).__init__()self.input_size = input_sizeself.hidden_size = hidden_size# 输入x的线性变换self.input_trans = Linear(num_directions * input_size, num_chunks * hidden_size, bias=bias, **factory_kwargs)# 隐藏状态的线性变换self.hidden_trans = Linear(hidden_size, num_chunks * hidden_size, bias=bias, **factory_kwargs)if reset_parameters:self.reset_parameters()def extra_repr(self) -> str:s = 'input_size={input_size}, hidden_size={hidden_size}'if 'bias' in self.__dict__ and self.bias is not True:s += ', bias={bias}'if 'nonlinearity' in self.__dict__ and self.nonlinearity != "tanh":s += ', nonlinearity={nonlinearity}'return s.format(**self.__dict__)

RNNCell

class RNNCell(RNNCellBase):def __init__(self, input_size, hidden_size: int, bias: bool = True, nonlinearity: str = 'tanh', num_directions=1, reset_parameters=True, device=None, dtype=None):factory_kwargs = {'device': device, 'dtype': dtype, 'reset_parameters': reset_parameters}super(RNNCell, self).__init__(input_size, hidden_size, num_chunks=1, bias=bias, num_directions=num_directions,**factory_kwargs)if nonlinearity == 'tanh':self.activation = F.tanhelse:self.activation = F.reludef forward(self, x: Tensor, h: Tensor, c: Tensor = None) -> Tuple[Tensor, None]:h_next = self.activation(self.input_trans(x) + self.hidden_trans(h))return h_next, None

RNNCellforward中也返回了一个元组,元组中第二个元素代表了c_next,为了兼容LSTM的实现。

RNNBase

class RNNBase(Module):def __init__(self, cell: RNNCellBase, input_size: int, hidden_size: int, batch_first: bool = False, num_layers: int = 1, bidirectional: bool = False, bias: bool = True, dropout: float = 0, reset_parameters=True, device=None, dtype=None) -> None:''' :param input_size:输入x的特征数 :param hidden_size: 隐藏状态的特征数 :param batch_first: 批次维度是否在前面 :param num_layers: 层数 :param bidirectional: 是否为双向 :param bias: 线性层是否包含偏置 :param dropout: 用于多层堆叠RNN,默认为0代表不使用dropout :param reset_parameters: 是否执行reset_parameters :param device: :param dtype: '''super(RNNBase, self).__init__()factory_kwargs = {'device': device, 'dtype': dtype, 'reset_parameters': reset_parameters}self.num_layers = num_layersself.hidden_size = hidden_sizeself.input_size = input_sizeself.batch_first = batch_firstself.bidirectional = bidirectionalself.bias = biasself.num_directions = 2 if self.bidirectional else 1# 支持多层self.cells = ModuleList([cell(input_size, hidden_size, bias, **factory_kwargs)] +[cell(hidden_size, hidden_size, bias, num_directions=self.num_directions,**factory_kwargs) for _ in range(num_layers - 1)])if self.bidirectional:# 支持双向self.back_cells = copy.deepcopy(self.cells)self.dropout = dropoutif dropout != 0:# Dropout层self.dropout_layer = Dropout(dropout)def _one_directional_op(self, input, n_steps, cell, h, c) -> Tuple[Tensor, Tensor, Tensor]:hs = []# 沿着input时间步进行遍历for t in range(n_steps):inp = input[t]h, c = cell(inp, h, c)hs.append(h)return h, c, F.stack(hs)def _handle_hidden_state(self, input, state):assert input.ndim == 3# 必须传入批数据,最小批大小为1if self.batch_first:batch_size, n_steps, _ = input.shapeinput = input.transpose((1, 0, 2))# 将batch放到中间维度else:n_steps, batch_size, _ = input.shapeif state is None:h = Tensor.zeros((self.num_layers * self.num_directions, batch_size, self.hidden_size), dtype=input.dtype, device=input.device)else:h = state# 得到每层的状态hs = list(F.unbind(h))# 按层数拆分hreturn hs, [None] * len(hs), input, n_steps, batch_sizedef forward(self, input: Tensor, state: Tensor) -> Tuple[Tensor, Tensor, Tensor]:'''RNN的前向传播:param input: 形状 [n_steps, batch_size, input_size] 若batch_first=False:param state: (隐藏状态,单元状态)元组, 每个元素形状 [num_layers, batch_size, hidden_size]:return:num_directions = 2 if self.bidirectional else 1output: (n_steps, batch_size, num_directions * hidden_size)若batch_first=False 或(batch_size, n_steps, num_directions * hidden_size)若batch_first=True包含每个时间步最后一层(多层RNN)的输出h_th_n: (num_directions * num_layers, batch_size, hidden_size) 包含最终隐藏状态c_n: (num_directions * num_layers, batch_size, hidden_size) 包含最终单元状态(LSTM);非LSTM为None'''hs, cs, input, n_steps, batch_size = self._handle_hidden_state(input, state)# 正向得到的h_n,反向得到的h_n,正向得到的c_n,反向得到的c_nh_n_f, h_n_b, c_n_f, c_n_b = [], [], [], []for layer in range(self.num_layers):h, c, hs_f = self._one_directional_op(input, n_steps, self.cells[layer], hs[layer], cs[layer])h_n_f.append(h)# 保存最后一个时间步的隐藏状态c_n_f.append(c)if self.bidirectional:h, c, hs_b = self._one_directional_op(F.flip(input, 0), n_steps, self.back_cells[layer],hs[layer + self.num_layers], cs[layer + self.num_layers])hs_b = F.flip(hs_b, 0)# 将输出时间步维度逆序,使得时间步t=0上,是看了整个序列的结果。# 拼接两个方向上的输入h_n_b.append(h)c_n_b.append(c)input = F.cat([hs_f, hs_b], 2)# (n_steps, batch_size, num_directions * hidden_size)else:input = hs_f# (n_steps, batch_size, num_directions * hidden_size)# 在第1层之后,最后一层之前需要经过dropoutif self.dropout and layer != self.num_layers - 1:input = self.dropout_layer(input)output = input# (n_steps, batch_size, num_directions * hidden_size) 最后一层最后计算的输入,就是它的输出c_n = Noneif self.bidirectional:h_n = F.cat([F.stack(h_n_f), F.stack(h_n_b)], 0)if c is not None:c_n = F.cat([F.stack(c_n_f), F.stack(c_n_b)], 0)else:h_n = F.stack(h_n_f)if c is not None:c_n = F.stack(c_n_f)if self.batch_first:output = output.transpose((1, 0, 2))return output, h_n, c_ndef extra_repr(self) -> str:s = 'input_size={input_size}, hidden_size={hidden_size}'if self.num_layers != 1:s += ', num_layers={num_layers}'if self.bias is not True:s += ', bias={bias}'if self.batch_first is not False:s += ', batch_first={batch_first}'if self.dropout:s += ', dropout={dropout}'if self.bidirectional is not False:s += ', bidirectional={bidirectional}'return s.format(**self.__dict__)

同样,做了兼容LSTM的实现,会多了一些if判断。

RNN

class RNN(RNNBase):def __init__(self, *args, **kwargs) -> None:''':param input_size:输入x的特征数:param hidden_size: 隐藏状态的特征数:param batch_first::param num_layers: 层数:param bidirectional: 是否为双向:param bias: 线性层是否包含偏置:param dropout: 用于多层堆叠RNN,默认为0代表不使用dropout:param nonlinearity: 非线性激活函数 tanh | relu'''super(RNN, self).__init__(RNNCell, *args, **kwargs)def forward(self, input: Tensor, state: Tensor = None) -> Tuple[Tensor, Tensor]:output, h_n, _ = super().forward(input, state)return output, h_n

因为基类RNNBaseforward会返回output,h_n,c_n,所以RNN这里重写了forward方法,仅返回outputh_n

通过这种方式实现GRURNN非常类似。

GRU

class GRU(RNNBase):def __init__(self, *args, **kwargs):''':param input_size:输入x的特征数:param hidden_size: 隐藏状态的特征数:param batch_first::param num_layers: 层数:param bidirectional: 是否为双向:param bias: 线性层是否包含偏置:param dropout: 用于多层堆叠RNN,默认为0代表不使用dropout'''super(GRU, self).__init__(GRUCell, *args, **kwargs)def forward(self, input: Tensor, state: Tensor = None) -> Tuple[Tensor, Tensor]:output, h_n, _ = super().forward(input, state)return output, h_n

实例测试

同样的配置下:

embedding_dim = 128hidden_dim = 128batch_size = 32num_epoch = 10n_layers = 2dropout = 0.2model = RNN(len(vocab), embedding_dim, hidden_dim, num_class, n_layers, dropout, bidirectional=True, mode=mode)

两层双向RNN可以得到75%的准确率。

Training Epoch 0: 94it [01:16,1.23it/s]Loss: 220.78Training Epoch 1: 94it [01:16,1.24it/s]Loss: 151.85Training Epoch 2: 94it [01:14,1.26it/s]Loss: 125.62Training Epoch 3: 94it [01:15,1.25it/s]Loss: 110.55Training Epoch 4: 94it [01:14,1.27it/s]Loss: 100.75Training Epoch 5: 94it [01:13,1.28it/s]Loss: 94.12Training Epoch 6: 94it [01:12,1.29it/s]Loss: 88.64Training Epoch 7: 94it [01:12,1.29it/s]Loss: 84.51Training Epoch 8: 94it [01:13,1.28it/s]Loss: 80.83Training Epoch 9: 94it [01:13,1.27it/s]Loss: 78.12Testing: 29it [00:06,4.79it/s]Acc: 0.75Cost:749.8793613910675

完整代码

https://github.com/nlp-greyfoss/metagrad

References


  1. Bidirectional recurrent neural networks↩︎

  2. Pytorch [Basics] — Intro to RNN ↩︎