框架全景：深度学习技术栈的演进与格局

第一章 框架全景：深度学习技术栈的演进与格局 / Chapter 1: Framework Panorama — Evolution and Landscape of the Deep Learning Stack

深度学习框架为何走向 PyTorch 与 TensorFlow 双寡头格局？这种格局背后的历史、技术与社区驱动力是什么？What forces—historical, technical, and community-driven—shaped the duopoly of PyTorch and TensorFlow in the deep learning ecosystem?

这并非一个偶然的结果。回望过去二十年，从 Theano 在蒙特利尔大学的实验室里点燃符号图自动微分的火种，到 Google Brain 在 2015 年开源 TensorFlow，再到 Meta AI（前 Facebook AI Research）在 2016 年推出 PyTorch 的前身 Torch 的 Python 化重构——这条演化曲线背后，是研究范式从静态计算图（Static Computational Graph）到动态计算图（Dynamic Computational Graph）的反复博弈，是工业部署对性能、可移植性与可扩展性的极端诉求，也是学术界对调试便捷性与表达力的本能追求。This is no accident. Looking back over the past two decades, from Theano igniting the spark of symbolic-graph automatic differentiation at the University of Montreal, to Google Brain open-sourcing TensorFlow in 2015, to Meta AI's Pythonic refactoring of Torch as PyTorch in 2016—the evolution behind this curve reflects the ongoing contest between static and dynamic computational graphs, the industrial demand for performance/portability/scalability, and academia's instinct for debuggability and expressiveness.

本章将以五个递进的小节展开这一全景图：先回溯第一代框架（Theano、Caffe、Torch7）的遗产与局限；再剖析 TensorFlow 1.x 静态图范式为何能重塑工业级机器学习基础设施；随后追踪 PyTorch 如何以 Pythonic 与动态图为武器完成对学术界与工业界的双重渗透；接着审视 JAX、MindSpore、OneFlow 等新生框架的差异化竞争策略；最后以 2020—2025 年间的多维度生态数据为锚，给出一份可量化的框架生态现状报告。This chapter unfolds this panorama in five progressive sections: first, we trace the legacy and limitations of first-generation frameworks (Theano, Caffe, Torch7); then we dissect why TensorFlow 1.x's static graph paradigm reshaped industrial ML infrastructure; next, we track how PyTorch leveraged Pythonic design and dynamic graphs to penetrate both academia and industry; then we examine the differentiation strategies of emerging frameworks (JAX, MindSpore, OneFlow); finally, we anchor the discussion with multi-dimensional ecosystem data from 2020—2025, delivering a quantifiable report on the current state of the framework ecosystem.

---

1.1 第一代框架的遗产与局限 / Legacy and Limitations of First-Generation Frameworks

要理解当前的双寡头格局，我们必须先回到"前 PyTorch / TensorFlow 时代"。2010 年前后，深度学习仍处于复兴前夜，研究者们使用的工具五花八门——既有 MATLAB 加手写 BP 算法的学术派，也有依赖 cuBLAS 自建 CUDA 扩展的性能派。第一代框架的核心使命，是把神经网络从"手工艺品"变成"可装配的乐高"。To understand today's duopoly, we must revisit the "pre-PyTorch / TensorFlow era." Around 2010, deep learning was on the eve of its revival, and researchers used a wide variety of tools—some preferred MATLAB with hand-written backpropagation, others relied on cuBLAS to build their own CUDA extensions. The core mission of first-generation frameworks was to transform neural networks from "handcrafted artifacts" into "assemblable Lego bricks."

1.1.1 Theano：符号图自动微分的奠基者 / Theano: The Founder of Symbolic-Graph Automatic Differentiation

Theano 由 Yoshua Bengio 团队在 2007 年发布，本质上是一个基于 Theano 的数学表达式编译器——用户先用 NumPy 风格的语法声明计算图，再调用 `theano.function` 将其编译为 C 代码或 CUDA 内核。Theano, released by Yoshua Bengio's group in 2007, is essentially a mathematical expression compiler based on symbolic graphs—users declare the computation graph using NumPy-like syntax, then invoke `theano.function` to compile it into C code or CUDA kernels.

```python import theano import theano.tensor as T

Theano 1.0 风格的符号图定义

x = T.dmatrix('x') y = T.dmatrix('y') z = x + y # 符号变量，非实际数值 f = theano.function([x, y], z) # 编译为可调用函数 print(f([[1, 2], [3, 4]], [[10, 20], [30, 40]])) ```

Theano 的最大贡献是确立了"符号式自动微分（Symbolic Automatic Differentiation）"这一范式：开发者只需声明前向计算图，框架即可通过 reverse-mode AD 自动生成梯度。其后续者 TensorFlow、CNTK、MXNet 在很大程度上都继承了这一思路。Theano's greatest contribution was establishing the "Symbolic Automatic Differentiation" paradigm: developers only need to declare the forward computation graph, and the framework can automatically generate gradients via reverse-mode AD. Its successors—TensorFlow, CNTK, MXNet—largely inherited this idea.

然而 Theano 的局限同样深刻：However, Theano's limitations were equally profound:

1. 编译开销巨大：每一次图结构变更都需要重新编译函数，在调试循环（debug loop）中的体验极差。Massive compilation overhead: every change to the graph structure required recompiling the function, leading to a terrible debugging experience. 2. 错误信息晦涩：Theano 的报错往往出现在 C++ 层，Python 栈几乎无法定位问题。Obscure error messages: Theano's errors often surfaced at the C++ level, making Python stacks nearly useless for debugging. 3. 缺乏对控制流的原生支持：循环与条件分支需要 `theano.scan` 等专门 API，写法冗长且性能欠佳。Lack of native control flow support: loops and conditionals required specialized APIs like `theano.scan`, which were verbose and poorly performing. 4. 生态孤立：Theano 没有与 scikit-learn、pandas 等 Python ML 生态深度集成。Ecosystem isolation: Theano lacked deep integration with the Python ML ecosystem such as scikit-learn and pandas.

Theano 在 2017 年 Bengio 团队宣布终止主开发，其遗产主要由两个继承者吸收——一个是走静态图工业路线的 TensorFlow，另一个是走动态图研究路线的 PyTorch。Theano's main development was discontinued by the Bengio team in 2017, and its legacy was largely absorbed by two inheritors—TensorFlow, which took the static-graph industrial route, and PyTorch, which took the dynamic-graph research route.

1.1.2 Caffe：视觉模型的工业级原型 / Caffe: An Industrial-Grade Prototype for Vision Models

2014 年，贾扬清在伯克利博士期间发布的 Caffe（Convolutional Architecture for Fast Feature Embedding），把卷积神经网络的训练流水线变成了一份配置文件。In 2014, Yangqing Jia released Caffe during his Ph.D. at Berkeley, turning the training pipeline of CNNs into a configuration file.

```protobuf

Caffe 的 prototxt 网络结构定义（节选）

layer { name: "conv1" type: "Convolution" bottom: "data" top: "conv1" convolution_param { num_output: 20 kernel_size: 5 stride: 1 } } layer { name: "pool1" type: "Pooling" bottom: "conv1" top: "pool1" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } ```

Caffe 的优势在于"配置文件 + 预训练权重"的工作流：用户可以通过编辑 prototxt 快速复现 AlexNet、VGG 等论文的实验，几乎所有视觉论文的作者都被迫在 Caffe 上"对齐"实验环境。这种"复现成本极低、训练速度极快"的特性，使 Caffe 在 2014—2017 年间成为视觉研究的事实标准。The advantage of Caffe lay in its "config-file + pretrained-weights" workflow: users could quickly reproduce experiments from AlexNet and VGG papers by editing prototxt, and almost all vision paper authors were forced to "align" their experimental environments with Caffe. This "extremely low reproduction cost + extremely fast training speed" made Caffe the de facto standard for vision research between 2014 and 2017.

但 Caffe 的致命伤也清晰可见：But Caffe's fatal flaws were also evident:

• **缺乏灵活性**：不支持 RNN、注意力机制等动态结构，研究范式一旦突破纯卷积，Caffe 即被抛弃。**Lack of flexibility**: RNNs, attention mechanisms, and other dynamic structures were not supported—once research paradigms moved beyond pure convolutions, Caffe was abandoned.
• **语言壁垒**：核心代码是 C++，新算法贡献门槛极高。**Language barrier**: the core code was C++, making contribution of new algorithms extremely difficult.
• **没有 Python API**：与 NumPy 生态割裂，导致研究效率低下。**No Python API**: its separation from the NumPy ecosystem led to low research efficiency.

Caffe 2（2017）试图引入灵活性，但来得太晚；它最终并入了 PyTorch 的 Caffe2 后端，成为 PyTorch 在移动端推理的雏形。Caffe 2 (2017) tried to introduce flexibility but arrived too late; it eventually merged into PyTorch's Caffe2 backend, becoming the embryo of PyTorch's mobile inference.

1.1.3 Torch7：动态优先的异类 / Torch7: The Dynamic-First Outlier

Torch7 起源于 2002 年的 NYU，其核心数据结构是 LuaJIT FFI 包装的 C 张量。Torch7 originated at NYU in 2002, with its core data structure being LuaJIT FFI-wrapped C tensors.

```lua -- Torch7 中定义一个 2 层 MLP require 'nn' mlp = nn.Sequential() mlp:add(nn.Linear(784, 256)) mlp:add(nn.Tanh()) mlp:add(nn.Linear(256, 10)) mlp:add(nn.LogSoftMax())

-- 即时构建、即时执行（define-by-run） output = mlp:forward(input) gradInput = mlp:backward(input, gradOutput) ```

Torch7 是"先有 PyTorch 范式"的早期实验场。它采用 define-by-run（即时构建）风格，开发者用 Lua 脚本逐层构造网络，每一行的执行立即在 C 张量上发生；反向传播由 `nn.Module:backward()` 隐式触发。Torch7 was an early experimental field for what would later become the PyTorch paradigm. It adopted a define-by-run style (immediate construction), where developers built networks layer by layer in Lua scripts, with each line's execution immediately happening on C tens