Skip to content
View lartpang's full-sized avatar
😶‍🌫️
Learning and Working
😶‍🌫️
Learning and Working

Block or report lartpang

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
lartpang/README.md

Hi 👋, I'm lartpang

🧑‍🤝‍🧑 Me

$$ \textbf{life} = \int_{birth}^{now} \mathbf{happy}(time) + \mathbf{sad}(time) d(time) $$

A Python and PyTorch developer, deep-learning worker and open-source activist.

Created by the awesome tool. 😊

📝 Recent Writing

  • 神经网络 | 从线性结构到可学习非线性 - Mon, 08 Dec 2025: CNN、Transformer、ONN(Operational Neural Network)和KAN
  • 储层计算 (Reservoir Computing) 概述 - Sat, 06 Dec 2025: 储层计算(RC)通过固定非线性储层与可训练线性读出的解耦设计,克服了传统递归神经网络训练中的梯度问题。其核心在于利用高维动力系统将输入信号映射到线性可分空间,仅需训练输出层权重。数学证明表明,当储层权重矩阵的谱半径满足特定条件时,系统具备回声状态属性和衰退记忆特性,确保状态收敛并遗忘久远历史。RC架构从随机连接演进到结构化拓扑(如简单环、带跳跃环),并发展出深度堆叠等变体,显著提升了计算效率与性能。这一范式为时间序列建模提供了高效解决方案。
  • 告别乱码:OpenCV 中文路径(Unicode)读写的解决方案 - Mon, 03 Nov 2025: 本文针对OpenCV中文路径读取失败问题,提出了一种基于C++17标准库的跨平台解决方案。核心思路是:使用std::filesystem处理中文路径,利用std::fstream进行二进制文件读写,最后通过OpenCV的imdecode和imencode函数实现图像编解码。
  • 生成模型 | DDPM -> Imrpoved DDPM -> DDIM - Sun, 24 Aug 2025: 本文介绍了三种扩散模型变体:DDPM、Improved DDPM和DDIM。DDPM通过迭代去噪过程生成样本,但采样速度较慢。Improved DDPM改进了噪声调度策略,采用余弦形式的调整,并引入混合损失函数以优化训练。DDIM则通过非马尔可夫链设计,在保持相同训练目标的同时,显著加快采样速度。这三种方法在扩散模型的噪声处理、损失函数设计和采样效率上各有创新,推动了扩散模型在生成任务中的性能提升。
  • 生成模型 | 扩散模型损失函数公式推导 - Sat, 23 Aug 2025: 本文推导了扩散模型的损失函数,通过引入前向分布简化计算,最终将损失分解为三部分:$L_T$(可忽略的常量)、$L_{t-1}$(KL散度项)和$L_0$(重构误差)。
  • 生成模型 | 扩散模型公式推导 - Sat, 23 Aug 2025: 本文介绍了扩散模型的前向加噪和反向去噪过程。前向过程通过马尔科夫链逐步将数据$x_0$转化为高斯噪声$x_T$,其中噪声强度由预设参数$�eta_t$控制。反向过程则利用神经网络从噪声$x_T$逐步恢复原始数据$x_0$。
  • ICCV 2025 | Reverse Convolution and Its Applications to Image Restoration - Sun, 17 Aug 2025: 本文提出了一种新颖的深度可分离反向卷积算子(reverse convolution),通过建立并求解正则化最小二乘优化问题,实现了对depthwise卷积的有效反转。该算子采用FFT推导闭式解,并详细研究了核初始化、padding策略等实现细节。基于此构建的reverse卷积块结合了层归一化、1×1卷积和GELU激活,形成类Transformer结构,可直接替换现有网络中的常规卷积层,构建ConverseNet。
  • TCSVT 2023 | StructToken - Rethinking Semantic Segmentation with Structural Prior - Sun, 17 Aug 2025: 一种新的语义分割范式,通过结构化token直接构建语义掩码并逐步细化,而非传统逐像素分类方法。作者设计了三种交互结构(CSE、SSE和静态卷积)来捕获特征图中的结构信息,并通过堆叠处理单元实现mask细化。
  • torchvision 中 deform_conv2d 操作的经验性解析 - Sun, 17 Aug 2025: 详细解析了torchvision中可变形卷积(deform_conv2d)的实现原理和使用方法。
  • 一次由默认参数引起的思考 - Sun, 17 Aug 2025: 本文探讨了依赖版本更新导致代码输出不一致的问题。作者在迁移代码时发现,由于Pillow图像处理库从6.2.1升级到7.2.0,其默认插值策略改变导致resize()函数输出结果不同。文章分析了默认参数的利弊,指出其虽提升开发效率但存在潜在风险。作者建议采取两种应对策略:一是固定依赖版本确保稳定性;二是对关键参数进行显式配置。最后强调开发应以程序稳定运行为首要目标,盲目追求新版本可能得不偿失,并提醒开发者需谨慎对待工具依赖的版本管理。

View the archives @ csdn@p_lart.

📽️ Some Projects

Name Stars Description
Hands-on-Docker (中文) stars 一份详尽的 Docker 使用指南。
Awesome-Class-Activation-Map stars An awesome list of papers and tools about the class activation map (CAM) technology.
PyTorchTricks stars Some tricks of pytorch…
MethodsCmp stars A Simple Toolkit for Counting the FLOPs/MACs, Parameters and FPS of Pytorch-based Methods.
PySODEvalToolkit stars A Python-based salient object detection and video object segmentation evaluation toolbox.
PySODMetrics stars A simple and efficient implementation of SOD metrcis.
PyLoss stars Some loss functions for deeplearning.
OpticalFlowBasedVOS stars A simple and efficient codebase for the optical flow based video object segmentation.
CoSaliencyProj stars A project for co-saliency detection. Some codes are borrowed from ICNet. Thanks to ICNet Intra-saliency Correlation Network for Co-Saliency Detection (NIPS2020)
RunIt stars A simple program scheduler for your code on different devices.
RegisterIt stars Register it: A more flexible register for the DeepLearning project.
mssim.pytorch stars A better pytorch-based implementation for the mean structural similarity. Differentiable simpler SSIM and MS-SSIM.
tta.pytorch stars Test-Time Augmentation library for Pytorch.
YuQueTools stars A simple tool to download your own articles from yuque.
ManageMyAttachments stars Manage the attachments of your own obsidian vault.

💾 Some Datasets

Name Description
lartpang/OVCamo Open-Vocabulary Camouflaged Object Segmentation
yooweey/AugmentedIRSTD1kTestset Augmented Testset of the IRSTD-1k dataset

Pinned Loading

  1. awesome-segmentation-saliency-dataset awesome-segmentation-saliency-dataset Public

    A collection of some datasets for segmentation / saliency detection. Welcome to PR...:smile:

    602 98

  2. PyTorchTricks PyTorchTricks Public

    Some tricks of pytorch... ⭐

    1.2k 126

  3. SAMs-CDConcepts-Eval SAMs-CDConcepts-Eval Public

    Inspiring the Next Generation of Segment Anything Models: Comprehensively Evaluate SAM and SAM 2 with Diverse Prompts Towards Context-Dependent Concepts under Different Scenes

    Python 574 31

  4. PySODMetrics PySODMetrics Public

    PySODMetrics: A Simple and Efficient Implementation of Grayscale/Binary Segmentation Metrics

    Python 370 26

  5. OVCamo OVCamo Public

    (ECCV 2024) Open-Vocabulary Camouflaged Object Segmentation

    Python 212 7

  6. PyIRSTDMetrics PyIRSTDMetrics Public

    [NeurIPS 2025 (D&B)] Rethinking Evaluation of Infrared Small Target Detection

    Python 265 18