论文介绍
题目:xLSTM: Extended Long Short-Term Memory
论文地址:https://arxiv.org/pdf/2405.04517
QQ深度学习交流群:994264161
扫描下方二维码,加入深度学习论文指南星球!
加入即可获得,模块缝合、制作、写作技巧,学会“结构”创新、“创新点”创新,从三区到顶会,小论文大论文,毕业一站式服务
创新点
引入指数门控(Exponential Gating):
改进传统LSTM的存储决策问题,增加指数激活函数的门控机制(输入门和遗忘门)。
提供了适当的归一化和稳定化技术,避免指数函数引起的数值溢出。
新的记忆结构:
sLSTM:采用标量记忆单元,引入新的记忆混合方法,并支持多头机制。
mLSTM:基于矩阵的记忆结构,支持完全并行化,并使用协方差更新规则来增强存储容量。
提出两种新的LSTM变体:
残差结构的改进:
将改进的LSTM集成到残差模块中,形成xLSTM模块,并通过堆叠这些模块构建完整的xLSTM架构。
提供两种残差块结构:用于sLSTM的后投影块和用于mLSTM的前投影块。
改进并行化能力:
mLSTM摒弃了传统LSTM中的隐藏-隐藏连接,增强了模型的并行性。
CUDA优化使得模型在GPU上的计算更加高效。
在存储容量和序列处理方面的性能提升:
mLSTM的矩阵记忆结构显著提高了模型的存储容量,可更好地处理稀有词预测和长序列上下文。
sLSTM的新记忆混合方法提升了模型的状态跟踪能力。
在语言建模中的竞争表现:
通过在大型数据集上的训练和对比实验,证明了xLSTM在验证困惑度(perplexity)和上下文外推能力上优于现有的Transformer和状态空间模型。
方法
整体架构
xLSTM 是一种基于残差模块构建的深层神经网络架构,通过堆叠 sLSTM(标量记忆单元,支持记忆混合)和 mLSTM(矩阵记忆单元,支持完全并行化)模块组成。其核心创新包括引入指数门控机制和协方差更新规则,以增强模型的存储能力和长序列处理能力。整个架构采用预归一化和残差连接来保证深层网络的稳定性,并通过线性计算复杂度实现高效的语言建模和长序列任务处理。
1. 基本组件:xLSTM模块
xLSTM模块由以下两种主要变体组成:
a. sLSTM(Scalar LSTM)模块:
记忆单元: 标量记忆(scalar memory)。
门控机制: 引入指数门控(Exponential Gating),允许输入门和遗忘门使用指数激活函数。
记忆混合(Memory Mixing): 支持多头机制,每个头内的记忆单元可以混合,但不同头之间不混合。
残差块: 后投影残差块(Post Up-Projection Block),即非线性变换发生在低维空间后,再线性映射到高维空间。
b. mLSTM(Matrix LSTM)模块:
记忆单元: 矩阵记忆(matrix memory),提升存储容量。
更新规则: 协方差更新(Covariance Update Rule),允许存储和检索键值对(key-value pairs)。
并行性: 摒弃了隐藏状态间的连接,支持完全并行化计算。
残差块: 前投影残差块(Pre Up-Projection Block),即先将输入映射到高维空间,进行非线性处理后再映射回原始空间。
2. 架构构建:xLSTM Blocks
通过堆叠多个xLSTM模块,构成整个架构的基础单元:
每个模块以预归一化(Pre-LayerNorm)和残差连接为基础,确保深层网络的稳定性。
模块内部采用sLSTM和mLSTM的混合组合,例如:
xLSTM[7:1]
表示 7 个 mLSTM 块和 1 个 sLSTM 块的组合。
3. 完整架构:xLSTM Architecture
通过堆叠多个xLSTM模块(如48个块),形成完整的深层模型架构。
支持预训练和下游任务的微调,针对语言建模任务进行优化。
即插即用模块作用
xLSTM 作为一个即插即用模块:
改进存储能力:
sLSTM 的指数门控机制允许对新信息进行更灵活的存储和更新,从而增强存储稀有事件和长时间依赖信息的能力。
增强表达能力:
通过引入记忆混合机制,sLSTM 可更好地整合多头或多单元的记忆,提升模型的上下文理解能力。
扩展性和兼容性:
sLSTM 可作为即插即用模块,嵌入到现有的深度学习架构(如 Transformer 或其他 RNN)中,增强其在状态跟踪和记忆管理方面的能力。
稳定性与易用性:
sLSTM 通过指数门控的归一化和稳定化技术,避免了数值溢出问题,使得模块在训练和推理过程中更加稳定和易用。
消融实验结果
位置:表 2 顶部
内容: 分析从传统 LSTM 到 xLSTM 的各个改进步骤对性能的影响。
说明:
每一步的改进显著降低了验证集困惑度(Perplexity)。
指数门控和矩阵记忆的引入对性能提升尤为关键。
将传统多层 LSTM 转换为 xLSTM,通过以下步骤逐步改进:
结果:
引入残差结构(ResNet Backbone)。
增加后投影模块(Up-Projection Backbone)。
添加指数门控(Exponential Gating)。
引入矩阵记忆(Matrix Memory)。
位置:表 2 底部
内容: 分析不同的门控机制对模型性能的影响。
说明:
没有门控的模型性能最差。
使用输入门和遗忘门的组合时模型表现最好,验证了门控机制对模型存储能力和表达能力的重要性。
比较了以下门控机制的影响:
结果:
无门控(No Gates)。
单独使用遗忘门(Forget Gate)。
单独使用输入门(Input Gate)。
同时使用输入门和遗忘门,并结合指数激活函数。
即插即用模块
import torch
import torch.nn as nn
from typing import Tuple, Optional, List
class sLSTMCell(nn.Module):
def __init__(self, input_size: int, hidden_size: int, bias: bool = True) -> None:
super().__init__()
# Store the input and hidden size
self.input_size = input_size
self.hidden_size = hidden_size
self.bias = bias
# Combine the Weights and Recurrent weights into a single matrix
self.W = nn.Parameter(
nn.init.xavier_uniform_(
torch.randn(self.input_size + self.hidden_size, 4 * self.hidden_size)
),
requires_grad=True,
)
# Combine the Bias into a single matrix
if self.bias:
self.B = nn.Parameter(
(torch.zeros(4 * self.hidden_size)), requires_grad=True
)
def forward(
self,
x: torch.Tensor,
internal_state: Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor],
) -> Tuple[
torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]
]:
# Unpack the internal state
h, c, n, m = internal_state # (batch_size, hidden_size)
# Combine the weights and the input
combined = torch.cat((x, h), dim=1) # (batch_size, input_size + hidden_size)
# Calculate the linear transformation
gates = torch.matmul(combined, self.W) # (batch_size, 4 * hidden_size)
# Add the bias if included
if self.bias:
gates += self.B
# Split the gates into the input, forget, output and stabilization gates
z_tilda, i_tilda, f_tilda, o_tilda = torch.split(gates, self.hidden_size, dim=1)
# Calculate the activation of the states
z_t = torch.tanh(z_tilda) # (batch_size, hidden_size)
# Exponential activation of the input gate
i_t = torch.exp(i_tilda) # (batch_size, hidden_size)
# Exponential activation of the forget gate
f_t = torch.sigmoid(f_tilda) # (batch_size, hidden_size)
# Sigmoid activation of the output gate
o_t = torch.sigmoid(o_tilda) # (batch_size, input_size)
# Calculate the stabilization state
m_t = torch.max(torch.log(f_t) + m, torch.log(i_t)) # (batch_size, hidden_size)
# Calculate the input stabilization state
i_prime = torch.exp(i_tilda - m_t) # (batch_size, hidden_size)
# Calculate the new internal states
c_t = f_t * c + i_prime * z_t # (batch_size, hidden_size)
n_t = f_t * n + i_prime # (batch_size, hidden_size)
# Calculate the stabilized hidden state
h_tilda = c_t / n_t # (batch_size, hidden_size)
# Calculate the new hidden state
h_t = o_t * h_tilda # (batch_size, hidden_size)
return h_t, (
h_t,
c_t,
n_t,
m_t,
) # (batch_size, hidden_size), (batch_size, hidden_size), (batch_size, hidden_size), (batch_size, hidden_size)
def init_hidden(
self, batch_size: int, **kwargs
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
return (
torch.zeros(batch_size, self.hidden_size, **kwargs),
torch.zeros(batch_size, self.hidden_size, **kwargs),
torch.zeros(batch_size, self.hidden_size, **kwargs),
torch.zeros(batch_size, self.hidden_size, **kwargs),
)
class sLSTM(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
num_layers: int,
bias: bool = True,
batch_first: bool = False,
) -> None:
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.bias = bias
self.batch_first = batch_first
self.cells = nn.ModuleList(
[
sLSTMCell(input_size if layer == 0 else hidden_size, hidden_size, bias)
for layer in range(num_layers)
]
)
def forward(
self,
x: torch.Tensor,
hidden_states: Optional[
List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]
] = None,
) -> Tuple[
torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]
]:
# Permute the input tensor if batch_first is True
if self.batch_first:
x = x.permute(1, 0, 2)
# Initialize the hidden states if not provided
if hidden_states is None:
hidden_states = self.init_hidden(x.size(1), device=x.device, dtype=x.dtype)
else:
# Check if the hidden states are of the correct length
if len(hidden_states) != self.num_layers:
raise ValueError(
f"Expected hidden states of length {self.num_layers}, but got {len(hidden_states)}"
)
if any(state[0].size(0) != x.size(1) for state in hidden_states):
raise ValueError(
f"Expected hidden states of batch size {x.size(1)}, but got {hidden_states[0][0].size(0)}"
)
H, C, N, M = [], [], [], []
for layer, cell in enumerate(self.cells):
lh, lc, ln, lm = [], [], [], []
for t in range(x.size(0)):
h_t, hidden_states[layer] = (
cell(x[t], hidden_states[layer])
if layer == 0
else cell(H[layer - 1][t], hidden_states[layer])
)
lh.append(h_t)
lc.append(hidden_states[layer][0])
ln.append(hidden_states[layer][1])
lm.append(hidden_states[layer][2])
H.append(torch.stack(lh, dim=0))
C.append(torch.stack(lc, dim=0))
N.append(torch.stack(ln, dim=0))
M.append(torch.stack(lm, dim=0))
H = torch.stack(H, dim=0)
C = torch.stack(C, dim=0)
N = torch.stack(N, dim=0)
M = torch.stack(M, dim=0)
return H[-1], (H, C, N, M)
def init_hidden(
self, batch_size: int, **kwargs
) -> List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]]:
return [cell.init_hidden(batch_size, **kwargs) for cell in self.cells]
if __name__ == '__main__':
# 定义输入张量的参数
input_size = 128
hidden_size = 128
num_layers = 2
seq_length = 10
batch_size = 32
dropout = 0.1
# 初始化 mLSTM 模块
block = sLSTM(input_size, hidden_size, num_layers)
# 随机生成输入张量
input_seq = torch.rand(batch_size, seq_length, input_size)
# 运行前向传递
output, hidden_state = block(input_seq)
# 输出输入张量和输出张量的形状
print(" sLSTM.Input size:", input_seq.size())
print("sLSTM.Output size:", output.size())
class mLSTMCell(nn.Module):
def __init__(self, input_size: int, hidden_size: int, bias: bool = True) -> None:
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.bias = bias
# Initialize weights and biases
self.W_i = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
self.W_f = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
self.W_o = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
self.W_q = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
self.W_k = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
self.W_v = nn.Parameter(
nn.init.xavier_uniform_(torch.zeros(input_size, hidden_size)),
requires_grad=True,
)
if self.bias:
self.B_i = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
self.B_f = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
self.B_o = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
self.B_q = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
self.B_k = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
self.B_v = nn.Parameter(torch.zeros(hidden_size), requires_grad=True)
def forward(
self,
x: torch.Tensor,
internal_state: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
# Get the internal state
C, n, m = internal_state
# Calculate the input, forget, output, query, key and value gates
i_tilda = (
torch.matmul(x, self.W_i) + self.B_i
if self.bias
else torch.matmul(x, self.W_i)
)
f_tilda = (
torch.matmul(x, self.W_f) + self.B_f
if self.bias
else torch.matmul(x, self.W_f)
)
o_tilda = (
torch.matmul(x, self.W_o) + self.B_o
if self.bias
else torch.matmul(x, self.W_o)
)
q_t = (
torch.matmul(x, self.W_q) + self.B_q
if self.bias
else torch.matmul(x, self.W_q)
)
k_t = (
torch.matmul(x, self.W_k) / torch.sqrt(torch.tensor(self.hidden_size))
+ self.B_k
if self.bias
else torch.matmul(x, self.W_k) / torch.sqrt(torch.tensor(self.hidden_size))
)
v_t = (
torch.matmul(x, self.W_v) + self.B_v
if self.bias
else torch.matmul(x, self.W_v)
)
# Exponential activation of the input gate
i_t = torch.exp(i_tilda)
f_t = torch.sigmoid(f_tilda)
o_t = torch.sigmoid(o_tilda)
# Stabilization state
m_t = torch.max(torch.log(f_t) + m, torch.log(i_t))
i_prime = torch.exp(i_tilda - m_t)
# Covarieance matrix and normalization state
C_t = f_t.unsqueeze(-1) * C + i_prime.unsqueeze(-1) * torch.einsum(
"bi, bk -> bik", v_t, k_t
)
n_t = f_t * n + i_prime * k_t
normalize_inner = torch.diagonal(torch.matmul(n_t, q_t.T))
divisor = torch.max(
torch.abs(normalize_inner), torch.ones_like(normalize_inner)
)
h_tilda = torch.einsum("bkj,bj -> bk", C_t, q_t) / divisor.view(-1, 1)
h_t = o_t * h_tilda
return h_t, (C_t, n_t, m_t)
def init_hidden(
self, batch_size: int, **kwargs
) -> Tuple[torch.Tensor, torch.Tensor]:
return (
torch.zeros(batch_size, self.hidden_size, self.hidden_size, **kwargs),
torch.zeros(batch_size, self.hidden_size, **kwargs),
torch.zeros(batch_size, self.hidden_size, **kwargs),
)
class mLSTM(nn.Module):
def __init__(
self,
input_size: int,
hidden_size: int,
num_layers: int,
bias: bool = True,
batch_first: bool = False,
) -> None:
super().__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.bias = bias
self.batch_first = batch_first
self.cells = nn.ModuleList(
[
mLSTMCell(input_size if layer == 0 else hidden_size, hidden_size, bias)
for layer in range(num_layers)
]
)
def forward(
self,
x: torch.Tensor,
hidden_states: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None,
) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
# Permute the input tensor if batch_first is True
if self.batch_first:
x = x.permute(1, 0, 2)
if hidden_states is None:
hidden_states = self.init_hidden(x.size(1), device=x.device, dtype=x.dtype)
else:
# Check if the hidden states are of the correct length
if len(hidden_states) != self.num_layers:
raise ValueError(
f"Expected hidden states of length {self.num_layers}, but got {len(hidden_states)}"
)
if any(state[0].size(0) != x.size(1) for state in hidden_states):
raise ValueError(
f"Expected hidden states of batch size {x.size(1)}, but got {hidden_states[0][0].size(0)}"
)
H, C, N, M = [], [], [], []
for layer, cell in enumerate(self.cells):
lh, lc, ln, lm = [], [], [], []
for t in range(x.size(0)):
h_t, hidden_states[layer] = (
cell(x[t], hidden_states[layer])
if layer == 0
else cell(H[layer - 1][t], hidden_states[layer])
)
lh.append(h_t)
lc.append(hidden_states[layer][0])
ln.append(hidden_states[layer][1])
lm.append(hidden_states[layer][2])
H.append(torch.stack(lh, dim=0))
C.append(torch.stack(lc, dim=0))
N.append(torch.stack(ln, dim=0))
M.append(torch.stack(lm, dim=0))
H = torch.stack(H, dim=0)
C = torch.stack(C, dim=0)
N = torch.stack(N, dim=0)
M = torch.stack(M, dim=0)
return H[-1], (H, C, N, M)
def init_hidden(
self, batch_size: int, **kwargs
) -> List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
return [cell.init_hidden(batch_size, **kwargs) for cell in self.cells]
if __name__ == '__main__':
dropout = 0.1
block = mLSTM(128, 128, 2)
input_seq = torch.rand(32, 10, 128)
output, hidden_state = block(input_seq)
print(input_seq.size()) print(output.size())
便捷下载方式
浏览打开网址:https://github.com/ai-dawang/PlugNPlay-Modules
更多分析可见原文