深度学习实证研究 (DeepLearning-Empirical-Studies)

前言

本仓库创立于 2026 年 5 月 11 日，也算是我第一个决定长期更新和维护的仓库吧。在经历了持续约大半年的人工智能、计算机视觉、医学图像分割等领域的学习后，每当我学习一些新知识时都会产生一些奇奇怪怪的想法：“如果？那么？...”。因此本仓库主要用于记录作者学习机器学习/深度学习期间遇到的问题与想法。作者希望通过控制实验探究模型的表征现象，旨在从定性（未来也可能会定量）角度分析模型在不同实验条件下的表现差异，并思考其中的原因。期待从中得到启发，也希望作者的专业能力能与本仓库一样不断随着实践而进步。

组织架构

每个独立的实验主题均存放在单独的目录下。代码、实验报告与复现实验说明会保留在仓库中；本地数据集、模型权重、训练日志与中间输出文件默认不会上传至 GitHub。

DeepLearning-Empirical-Studies/
├── Exp_01/                         # 实验一：数据规模、模型复杂度与输入信息量控制实验
│   ├── code/                       # 实验源代码
│   │   ├── train.py                # 训练入口
│   │   ├── model.py                # MLP、AlexNet、ResNet 系列模型定义
│   │   ├── load_data.py            # CIFAR-10/CIFAR-100 数据加载
│   │   ├── utils.py                # 日志、统计与权重保存工具
│   │   └── generate_figures/       # 论文图像生成脚本
│   ├── report/                     # 实验报告
│   │   ├── Report_cn.pdf           # 中文报告
│   │   └── Report_en.pdf           # 英文报告
│   ├── pre_exp.ipynb               # 预实验 notebook
│   ├── requirements.txt            # 当前实验依赖
│   └── README.md                   # 当前实验说明
├── Exp_02/                         # 后续实验主题
├── .gitignore                      # Git 忽略规则
├── LICENSE
└── README.md                       # 仓库主页

数据信息

• 本仓库当前实验主要使用 CIFAR-10 与 CIFAR-100 数据集。
• 数据集默认下载或放置在实验目录外的 data/ 目录中，相关数据文件不会上传至 GitHub。
• 训练过程中产生的模型权重、SwanLab 日志、CSV 输出与论文中间图像默认属于本地产物，也不会随仓库提交。
• 若需要复现实验，请进入对应实验目录查看独立的 README.md 和 requirements.txt。

实验表格

下表收录了本仓库所有的实证探究项目。点击“实验名称”可进入对应目录查看源码、报告与复现实验说明。

实验编号	实验名称	探究内容	实验报告在线预览	完成状态
Exp-01	Data Scale, Model Complexity, and Input Information in Vision Generalization	通过 CIFAR-10/CIFAR-100 控制变量实验，观察训练数据规模、模型复杂度与输入信息模态对模型拟合能力和泛化性能的影响。	arXiv:2606.04409	进行中

联系方式

• 邮箱：zhoulyd@126.com

如果觉得有收获，不妨点击一个 ⭐ 哦~~~

---

DeepLearning-Empirical-Studies

Preface

This repository was created on May 11, 2026. It is also the first repository that I have decided to maintain and update over the long term. After spending more than half a year learning about artificial intelligence, computer vision, medical image segmentation, and related fields, I often come up with small experimental questions whenever I encounter new knowledge: “What if...? Then what would happen...?” Therefore, this repository is mainly used to record the problems, ideas, and empirical explorations that arise during my study of machine learning and deep learning. Through controlled experiments, I hope to investigate representational phenomena in models, compare model behavior under different experimental conditions from a qualitative perspective, and reflect on the possible causes behind these observations. I hope this repository can bring useful inspiration, and that my own professional ability can keep improving together with these experiments.

Repository Structure

Each independent experimental topic is stored in a separate directory. Source code, reports, and reproducibility instructions are kept in the repository, while local datasets, model weights, training logs, and intermediate outputs are excluded from GitHub by default.

DeepLearning-Empirical-Studies/
├── Exp_01/                         # Experiment 1: data scale, model complexity, and input information control
│   ├── code/                       # Source code
│   │   ├── train.py                # Training entry point
│   │   ├── model.py                # MLP, AlexNet, and ResNet model definitions
│   │   ├── load_data.py            # CIFAR-10/CIFAR-100 data loading
│   │   ├── utils.py                # Logging, statistics, and checkpoint utilities
│   │   └── generate_figures/       # Figure generation scripts
│   ├── report/                     # Experiment reports
│   │   ├── Report_cn.pdf           # Chinese report
│   │   └── Report_en.pdf           # English report
│   ├── pre_exp.ipynb               # Preliminary experiment notebook
│   ├── requirements.txt            # Dependencies for this experiment
│   └── README.md                   # Experiment-specific README
├── Exp_02/                         # Future experimental topics
├── .gitignore                      # Git ignore rules
├── LICENSE
└── README.md                       # Repository homepage

Data Information

• The current experiments mainly use CIFAR-10 and CIFAR-100.
• Datasets are downloaded to or placed under a local data/ directory, and dataset files are not uploaded to GitHub.
• Model checkpoints, SwanLab logs, CSV outputs, and intermediate paper figures are treated as local artifacts and are not committed by default.
• To reproduce an experiment, please check the corresponding experiment directory for its own README.md and requirements.txt.

Experiment List

The following table summarizes the empirical studies included in this repository. Click the experiment name to view the source code, reports, and reproduction instructions.

ID	Experiment	Topic	Report Preview	Status
Exp-01	Data Scale, Model Complexity, and Input Information in Vision Generalization	Controlled experiments on CIFAR-10/CIFAR-100 to study how training data scale, model complexity, and input information modalities affect fitting ability and generalization performance.	arXiv:2606.04409	In progress

Contact

• Email: zhoulyd@126.com

If you find this repository helpful, feel free to leave a ⭐.