Static Local Linearization (SLL): zero-intrusive auto-differentiation for discrete programs
Project description
🔷 SLL-Core: Static Local Linearization
离散程序的零侵入可微分化引擎
🤯 问题:为什么离散程序无法自动求导?
在深度学习里,离散决策无处不在:
- 量化:
round(x)、floor(x) - 阈值判断:
sign(x)、x > 0 - 分类选择:
argmax(x)
但这些操作有一个致命特性:梯度几乎处处为零,导致标准反向传播直接失效。
x = torch.tensor([0.5], requires_grad=True)
y = torch.sign(x) # ❌ 梯度为 0,参数永远更新不了
loss = (y - target).pow(2).sum()
loss.backward()
print(x.grad) # tensor([0.]) ← 死了
传统方案的缺点
| 方法 | 是否需要改代码 | 部署有残留 | 梯度质量 | 收敛稳定性 |
|---|---|---|---|---|
| 硬函数直接训练 | ✅ 无需改动 | ✅ 无残留 | ❌ 零梯度,无法训练 | ❌ 完全不收敛 |
| Sigmoid / Softmax 松弛 | ❌ 重写模型 | ❌ 有近似误差 | ⚠️ 梯度消失/爆炸 | ⚠️ 调参困难 |
| Straight-Through Estimator (STE) | ❌ 手写自定义梯度 | ✅ 无残留 | ⚠️ 梯度方向不准 | ⚠️ 容易震荡 |
| 重参数化/Gumbel-Softmax | ❌ 改模型结构 | ❌ 有温度参数残留 | ⚠️ 高方差 | ⚠️ 慢 |
| ⭐ SLL (静态局部线性化) | ✅ 零侵入 | ✅ 严格恢复硬逻辑 | ✅ 常数梯度,无消失 | ✅ 稳定收敛 |
SLL 的核心洞察:不需要在整个定义域上做近似,只在决策边界附近 ε-区间局部线性化,其余区域保持原始硬逻辑。当 ε → 0 时,最优解收敛到原始离散问题的最优解。
⚡ 一句话解决
import torch
import sll
x = torch.tensor([-1.0, 0.0, 1.0], requires_grad=True)
with sll.linearize(eps=1e-2): # ← 就这行
y = torch.sign(x) # 自动可微!
z = torch.round(y * 10)
loss = z.sum()
loss.backward()
print(x.grad) # 梯度正常回传 ✅
离开上下文后,torch.sign 自动恢复原始硬逻辑——训练时可微,部署时零开销。
📊 SLL 为什么更好?
梯度质量对比
| 硬函数 | STE | Sigmoid 松弛 | SLL | |
|---|---|---|---|---|
| 前向输出 | [-1, 0, 1] |
[-1, 0, 1] |
连续值(有误差) | 精确硬输出 |
| 边界附近梯度 | 0 |
1(常数) |
高斯峰(易消失) | 常数 1/(2ε) |
| 远离边界梯度 | 0 |
1 ≈ 0 |
0 |
0(硬逻辑) |
| 是否需要调温度参数 | — | — | 需要调 β |
无需调参 |
可视化对比
上图展示了:
- 左上:SLL 在
|x| > ε时严格等于硬 Sign,在边界附近平滑过渡 - 中上:SLL 梯度在边界区间内为常数,无 Sigmoid 式梯度消失问题
- 右上:SLL Round 在整数点附近线性过渡,远离边界完全等于硬 Round
- 左下:SLL 只在
[-ε, ε]区间做局部线性化,其余区域不受影响 - 中下:
ε越小越接近硬函数,ε越大过渡越平滑 - 右下:SLL 可以稳定收敛,硬函数完全无法优化
🚀 安装
pip install sll-core
要求:Python ≥ 3.8,PyTorch ≥ 1.9.0
🎯 快速开始
方式一:上下文管理器(推荐)
一行包裹,零侵入改造现有代码:
import torch
import sll
model = MyDiscreteModel() # 你的原始模型,无需改动
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
for epoch in range(100):
x = torch.randn(32, 10)
target = torch.randn(32, 1)
with sll.linearize(eps=1e-2): # ← 训练时加这一行
y = model(x)
loss = (y - target).pow(2).sum()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 部署时:直接 model(x),无需 SLL,零开销 ✅
方式二:装饰器
@sll.enable(eps=1e-2)
def quantized_model(x):
quantized = torch.round(x * 2) / 2
return torch.sign(quantized)
x = torch.randn(5, requires_grad=True)
y = quantized_model(x)
y.sum().backward() # 梯度正常计算
方式三:显式调用(不 patch 全局状态)
y = sll.heaviside(x, eps=1e-2)
z = sll.sign(y, eps=1e-2)
📋 支持的可微离散算子
| 算子 | 描述 | 使用示例 |
|---|---|---|
heaviside |
Heaviside 阶跃函数 | sll.heaviside(x) |
sign |
符号函数 | sll.sign(x) / torch.sign(x) |
round |
四舍五入 | sll.round(x) / torch.round(x) |
floor |
向下取整 | sll.floor(x) / torch.floor(x) |
ceil |
向上取整 | sll.ceil(x) / torch.ceil(x) |
threshold |
通用阈值函数 | sll.threshold(x, threshold=0.5) |
argmax |
Soft one-hot 编码 | sll.argmax(x, dim=1) |
🔬 实际应用场景
场景 1:量化感知训练 (QAT)
import torch
import sll
def quantize(x, levels=256):
scale = (levels - 1) / (x.max() - x.min() + 1e-10)
return torch.round((x - x.min()) * scale) / scale + x.min()
x = torch.randn(10, requires_grad=True)
with sll.linearize(eps=1e-3):
y = quantize(x) # 量化操作可微了!
loss = y.sum()
loss.backward()
print("量化梯度:", x.grad) # ✅ 梯度正常回传
场景 2:带硬阈值激活的网络
import torch
import torch.nn as nn
import sll
class DiscreteModel(nn.Module):
def __init__(self):
super().__init__()
self.linear = nn.Linear(10, 5)
def forward(self, x):
x = self.linear(x)
return (x > 0).float() # 硬阈值,原本不可微
model = DiscreteModel()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
# 用 SLL 训练——模型代码完全不用改!
with sll.linearize(eps=1e-2):
y = model(x)
loss = (y - target).pow(2).sum()
optimizer.zero_grad()
loss.backward()
optimizer.step()
🧮 数学原理
SLL 在离散决策边界附近建立局部线性化区间:
- 入口处理:硬边界替换为 ε-局部线性函数
- 可微计算:边界附近使用线性近似,保证处处可微
- 梯度回传:边界附近导数为常数,无梯度消失
- 出口恢复:严格恢复原始硬逻辑,部署零开销
以 Heaviside 阶跃函数为例:
$$ y(x) = \begin{cases} 0.5 + x/(2\epsilon) & 当 |x| ≤ \epsilon \\ H(x) & 其他 \end{cases} $$
其中 H(x) 是原始 Heaviside 函数。当 ε → 0 时,y(x) → H(x),最优解收敛到原始问题最优解。
⚙️ 参数说明
eps:线性化区间半宽,默认1e-3- 输入距离硬边界 ≤
eps:使用线性化近似 - 输入距离硬边界 >
eps:使用原始硬逻辑 eps越小,越接近硬逻辑,梯度区域越窄eps越大,过渡越平滑,近似区域越宽
- 输入距离硬边界 ≤
⚠️ 注意事项
- Tensor 方法建议:
x.sign()等 Tensor 方法 SLL 会尽力拦截,但建议用torch.sign(x)确保一致性 - 比较运算符:Python 比较(如
x > 0)无法被拦截,建议用sll.threshold(x)替代 - 部署阶段:训练完成后直接部署原始代码,无需加载 SLL,零性能损失
- ε 选择:建议从
1e-2开始,根据任务收敛情况微调
📄 许可证
MIT License
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file sll_core-0.1.5.tar.gz.
File metadata
- Download URL: sll_core-0.1.5.tar.gz
- Upload date:
- Size: 335.1 kB
- Tags: Source
- Uploaded using Trusted Publishing? Yes
- Uploaded via: twine/6.1.0 CPython/3.13.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
8ab234416f9d140613e89629ccb507bd5a93c2f03a68799f65420779a8c36ca0
|
|
| MD5 |
a219c5b3b5ae14dabe0fab9d60c7ffe1
|
|
| BLAKE2b-256 |
458b561c5dacce5583461f2a560a757f792f899047af951becfac6008d8d4e98
|
Provenance
The following attestation bundles were made for sll_core-0.1.5.tar.gz:
Publisher:
publish.yml on jacksong-sourse/sll-core
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
sll_core-0.1.5.tar.gz -
Subject digest:
8ab234416f9d140613e89629ccb507bd5a93c2f03a68799f65420779a8c36ca0 - Sigstore transparency entry: 1448990651
- Sigstore integration time:
-
Permalink:
jacksong-sourse/sll-core@504b3b779650750b2f3dafbcbeeadb6370964b78 -
Branch / Tag:
refs/tags/v1.0.5 - Owner: https://github.com/jacksong-sourse
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@504b3b779650750b2f3dafbcbeeadb6370964b78 -
Trigger Event:
push
-
Statement type:
File details
Details for the file sll_core-0.1.5-py3-none-any.whl.
File metadata
- Download URL: sll_core-0.1.5-py3-none-any.whl
- Upload date:
- Size: 9.9 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? Yes
- Uploaded via: twine/6.1.0 CPython/3.13.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
4620510394a8209439eb11ca96c9bdfd3e2291875ebf6e24d2e18f23df56b53c
|
|
| MD5 |
39c99035823d192150f8acce21505133
|
|
| BLAKE2b-256 |
361b9e1a284a011e6fe613a09d022daaa8b5433a136631554a4ba6db7911ea12
|
Provenance
The following attestation bundles were made for sll_core-0.1.5-py3-none-any.whl:
Publisher:
publish.yml on jacksong-sourse/sll-core
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
sll_core-0.1.5-py3-none-any.whl -
Subject digest:
4620510394a8209439eb11ca96c9bdfd3e2291875ebf6e24d2e18f23df56b53c - Sigstore transparency entry: 1448990701
- Sigstore integration time:
-
Permalink:
jacksong-sourse/sll-core@504b3b779650750b2f3dafbcbeeadb6370964b78 -
Branch / Tag:
refs/tags/v1.0.5 - Owner: https://github.com/jacksong-sourse
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@504b3b779650750b2f3dafbcbeeadb6370964b78 -
Trigger Event:
push
-
Statement type: