关于相同数据在RNN cell 和 RNN上运行结果不一致的分析

关于相同数据在RNN cell 和 RNN上运行结果不一致的分析

代码

import torch
import torch.nn as nn
import torch.nn.functional as Finput_size = 2  # input_size means the lengths of one-hot encode vector, for example, the code [... 128 dim ...] of 'o' in "hello"
batch_size = 1
seq_len = 3  # it means the length of the whole sequence rather than one-hot encode vector
num_layers = 1
hidden_size = 5data = torch.randn(seq_len, batch_size, input_size)  # (3,2,4)
print(data)
hidden = torch.zeros(batch_size, hidden_size)  # (2,4)
print(hidden)# RNN Cell part
# the vector dimension of input and output for every sample x
Cell = nn.RNNCell(input_size=input_size, hidden_size=hidden_size)for idx, input in enumerate(data):print("=" * 20, idx, "=" * 20)print("input shape:", input.shape)print(input)print(hidden)hidden = Cell(input, hidden)print("hidden shape:", hidden.shape)print(hidden)
print("=" * 20, "=", "=" * 20, "\n")# RNN part
hidden = torch.zeros(num_layers, batch_size, hidden_size)  # (2,4)
RNN = nn.RNN(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers)out, hidden = RNN(data, hidden)
print("data:", data)
print("output size:", out.shape)
print("output:", out)print("hidden size:", hidden.shape)
print("hidden:", hidden)

运行结果

tensor([[[ 1.5129,  0.0149]],[[ 0.5056, -1.7836]],[[ 0.4780,  0.8940]]])
tensor([[0., 0., 0., 0., 0.]])
==================== 0 ====================
input shape: torch.Size([1, 2])
tensor([[1.5129, 0.0149]])
tensor([[0., 0., 0., 0., 0.]])
hidden shape: torch.Size([1, 5])
tensor([[-0.5329,  0.1464, -0.1624, -0.6233, -0.5527]], grad_fn=<TanhBackward>)
==================== 1 ====================
input shape: torch.Size([1, 2])
tensor([[ 0.5056, -1.7836]])
tensor([[-0.5329,  0.1464, -0.1624, -0.6233, -0.5527]], grad_fn=<TanhBackward>)
hidden shape: torch.Size([1, 5])
tensor([[-0.5681,  0.0588, -0.7489,  0.1834, -0.7691]], grad_fn=<TanhBackward>)
==================== 2 ====================
input shape: torch.Size([1, 2])
tensor([[0.4780, 0.8940]])
tensor([[-0.5681,  0.0588, -0.7489,  0.1834, -0.7691]], grad_fn=<TanhBackward>)
hidden shape: torch.Size([1, 5])
tensor([[ 0.1734, -0.2558, -0.4364, -0.6313, -0.1552]], grad_fn=<TanhBackward>)
==================== = ==================== 
data: tensor([[[ 1.5129,  0.0149]],[[ 0.5056, -1.7836]],[[ 0.4780,  0.8940]]])
output size: torch.Size([3, 1, 5])
output: tensor([[[ 0.1714, -0.2880,  0.1836,  0.8961, -0.5190]],[[-0.3229, -0.4838, -0.0166,  0.7877, -0.0175]],[[ 0.7183,  0.1400,  0.0856,  0.6780, -0.1199]]],grad_fn=<StackBackward>)
hidden size: torch.Size([1, 1, 5])
hidden: tensor([[[ 0.7183,  0.1400,  0.0856,  0.6780, -0.1199]]],grad_fn=<StackBackward>)

问题描述

通过输出结果可以看见，两边分别得到三次的h_i的数值是不一样的；

我认为RNN是把RNN cell的循环进行了封装（这里我把RNN的num_layer设置为1，其他参数和RNN cell一致；输入数据一致；h₀初始为0向量），因此参与的计算过程也应该是一样的

那么，为什么两边得到的结果不一致呢？

分析

待解决