python - RNN 中的隐藏大小与输入大小

Question

前提1:

关于 RNN 层中的神经元 - 我的理解是，在“每个时间步，每个神经元都接收输入向量 x (t) 和前一个时间步的输出向量 y (t –1)”^[1]:

前提2:

据我了解，在 Pytorch 的 GRU 层中，input_size 和 hidden_​​size 含义如下:

• input_size – The number of expected features in the input x
• hidden_size – The number of features in the hidden state h

自然地，hidden_​​size 应该表示 GRU 层中神经元的数量。

我的问题:

给定以下 GRU 层:

# assume that hidden_size = 3

class Encoder(nn.Module):
    def __init__(self, src_dictionary_size, hidden_size):
        super(Encoder, self).__init__()
        self.embedding = nn.Embedding(src_dictionary_size, hidden_size)
        self.gru = nn.GRU(input_size = hidden_size, hidden_size = hidden_size)

假设hidden_​​size为3，我的理解是上面的GRU层将有3个神经元，每个神经元在每个时间步同时接受大小为3的输入向量。

我的问题是:为什么 hidden_​​size 和 input_size 的参数必须相等？ IE。为什么 3 个神经元中的每一个都不能接受大小为 5 的输入向量？

举个例子:以下两种情况都会导致尺寸不匹配:

self.gru = nn.GRU(input_size = hidden_size, hidden_size = hidden_size-1)
self.gru = nn.GRU(input_size = hidden_size, hidden_size = hidden_size+1)

[1] 杰龙，奥雷利安。使用 Scikit-Learn 和 TensorFlow 进行机器学习实践(第 388 页)。奥莱利媒体。 Kindle版。

[3] https://pytorch.org/docs/stable/nn.html#torch.nn.GRU

<小时/>

添加完整代码以实现可重复性:

import torch
import torch.nn as nn

class Encoder(nn.Module):
    def __init__(self, src_dictionary_size, hidden_size):
        super(Encoder, self).__init__()
        self.hidden_size = hidden_size
        self.embedding = nn.Embedding(src_dictionary_size, hidden_size)
        self.gru = nn.GRU(input_size = hidden_size, hidden_size = hidden_size-1)

    def forward(self, pad_seqs, seq_lengths, hidden):
        """
        Args:
          pad_seqs of shape (max_seq_length, batch_size, 1): Padded source sequences.
          seq_lengths: List of sequence lengths.
          hidden of shape (1, batch_size, hidden_size): Initial states of the GRU.

        Returns:
          outputs of shape (max_seq_length, batch_size, hidden_size): Padded outputs of GRU at every step.
          hidden of shape (1, batch_size, hidden_size): Updated states of the GRU.
        """
        embedded_sqs = self.embedding(pad_seqs).squeeze(2)
        packed_sqs = pack_padded_sequence(embedded_sqs, seq_lengths)
        packed_output, h_n = self.gru(packed_sqs, hidden)
        output, input_sizes = pad_packed_sequence(packed_output)

        return output, h_n

    def init_hidden(self, batch_size=1):
        return torch.zeros(1, batch_size, self.hidden_size)

def test_Encoder_shapes():
    hidden_size = 5
    encoder = Encoder(src_dictionary_size=5, hidden_size=hidden_size)

    # maximum word count
    max_seq_length = 4

    # num sentences
    batch_size = 2
    hidden = encoder.init_hidden(batch_size=batch_size)

    # these are padded sequences (sentences of words). There are 2 sentences (i.e. 2 batches) with a maximum of 4 words.
    pad_seqs = torch.tensor([
        [1, 2],
        [2, 3],
        [3, 0],
        [4, 0]
    ]).view(max_seq_length, batch_size, 1)

    outputs, new_hidden = encoder.forward(pad_seqs=pad_seqs, seq_lengths=[4, 2], hidden=hidden)
    assert outputs.shape == torch.Size([4, batch_size, hidden_size]), f"Bad outputs.shape: {outputs.shape}"
    assert new_hidden.shape == torch.Size([1, batch_size, hidden_size]), f"Bad new_hidden.shape: {new_hidden.shape}"
    print('Success')

test_Encoder_shapes()

Answer 1

我刚刚解决了这个问题，这个错误是我自己造成的。

结论:input_size 和 hidden_​​size 的大小可能不同，这没有固有的问题。问题中的前提陈述正确。

上面(完整)代码的问题是 GRU 的初始隐藏状态没有正确的维度。初始隐藏状态必须与后续隐藏状态具有相同的维度。就我而言，初始隐藏状态的形状为 (1,2,5) 而不是 (1,2,4)。在前者中，5表示嵌入向量的维数。 4 表示 GRU 中的hidden_​​size(神经元数量)。正确的代码如下:

import torch
import torch.nn as nn

class Encoder(nn.Module):
    def __init__(self, src_dictionary_size, input_size, hidden_size):
        super(Encoder, self).__init__()
        self.hidden_size = hidden_size
        self.embedding = nn.Embedding(src_dictionary_size, input_size)
        self.gru = nn.GRU(input_size = input_size, hidden_size = hidden_size)

    def forward(self, pad_seqs, seq_lengths, hidden):
        """
        Args:
          pad_seqs of shape (max_seq_length, batch_size, 1): Padded source sequences.
          seq_lengths: List of sequence lengths.
          hidden of shape (1, batch_size, hidden_size): Initial states of the GRU.

        Returns:
          outputs of shape (max_seq_length, batch_size, hidden_size): Padded outputs of GRU at every step.
          hidden of shape (1, batch_size, hidden_size): Updated states of the GRU.
        """
        embedded_sqs = self.embedding(pad_seqs).squeeze(2)
        packed_sqs = pack_padded_sequence(embedded_sqs, seq_lengths)
        packed_output, h_n = self.gru(packed_sqs, hidden)
        output, input_sizes = pad_packed_sequence(packed_output)

        return output, h_n

    def init_hidden(self, batch_size=1):
        return torch.zeros(1, batch_size, self.hidden_size)

def test_Encoder_shapes():
    hidden_size = 4
    embedding_size = 5
    encoder = Encoder(src_dictionary_size=5, input_size = embedding_size, hidden_size = hidden_size)
    print(encoder)

    max_seq_length = 4
    batch_size = 2
    hidden = encoder.init_hidden(batch_size=batch_size)
    pad_seqs = torch.tensor([
        [1, 2],
        [2, 3],
        [3, 0],
        [4, 0]
    ]).view(max_seq_length, batch_size, 1)

    outputs, new_hidden = encoder.forward(pad_seqs=pad_seqs, seq_lengths=[4, 2], hidden=hidden)
    assert outputs.shape == torch.Size([4, batch_size, hidden_size]), f"Bad outputs.shape: {outputs.shape}"
    assert new_hidden.shape == torch.Size([1, batch_size, hidden_size]), f"Bad new_hidden.shape: {new_hidden.shape}"
    print('Success')

test_Encoder_shapes()