lingerskill

    1

    Interactive deep learning paper analysis focusing on model architecture, tensor flow, and code-to-paper alignment.

    Free

    Security scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    lingerskill

    Example session with this skill installed

    使用lingerskill给我讲解这篇关于PatchTST的论文,重点看它的数据处理和架构。

    • Read your context and instructions
    • Compiled the lingerskill

    我已读取 PatchTST 论文及相关代码。该模型通过对时间序列进行分块(Patching)和独立通道处理,试图解决长期时序预测中的局部语义丢失问题。

    阶段 01:研究问题与架构导读

    PatchTST 的核心设计理念是 Channel-independence(通道独立)和 Patching(分块)。与传统的逐点(Point-wise)建模不同,它将增强局部语义信息并降低计算复杂度。

    整体架构数据流:
    原始序列 (B, L, C) → 通道拆分 (B*C, L, 1) → 分块 (B*C, N, P) → Transformer Encoder → 预测头 → 逆归一化 → 输出 (B, H, C)
    [P:PDF p.4, Fig.1]

    核心设计点

    • Channel-independence:每个变量独立建模,权重在通道间共享。
    • Patching:将连续时间点聚合为 Patch,作为 Transformer 的输入 Token,减少计算量并捕捉局部趋势。

    关键结论:PatchTST 证明了简单的通道独立设计在时序预测中往往优于复杂的跨变量建模。
    待解决问题:原文未详细说明在极端异构数据下通道共享权重的负面影响。

    下一步建议:进入“阶段 02:数据管线与分块逻辑”,详细拆解 Patch 的 Tensor 变换过程。

    Next steps

    • 输入“继续”以查看 Patching 的具体张量形状变化与代码实现。
    • 询问特定的 Patch 长度 (P) 或步长 (S) 对模型性能的影响。
    • 要求查看 Transformer Encoder 在分块数据上的注意力计算方式。

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Analyzing deep learning research papers, specifically in time-series forecasting, is often a disjointed process of jumping between PDF text, complex formulas, and separate code repositories. This skill bridges that gap by providing a step-by-step, interactive breakdown that aligns model architecture with actual data flow and code implementation.

    What it does

    • Architectural mapping provides clear flowcharts of data moving from raw input through internal transformations to the final prediction head.
    • Code-to-paper alignment links specific class methods and functions in official repositories to corresponding equations and sections in the text.
    • Pseudo-code generation constructs educational Python logic when official source code is unavailable or incomplete.
    • Tensor shape tracking details dimension changes [B, L, H, C] at critical module boundaries to explain how temporal or variable features interact.
    • Component attribution identifies which parts of the model are direct reuses, adaptations of prior work, or new mechanisms.

    How it works

    1. Invoke the skill using $lingerskill or a specific request to explain a paper.
    2. The agent reads the provided text, figures, and code to generate an initial research problem and architecture overview.
    3. The process moves through interactive stages: data pipeline, core modules, training/prediction, and final source synthesis.
    4. You provide a "continue" or "next" prompt to advance through stages, or request a deep dive into a specific formula or code block.

    Frameworks & tools

    This skill is optimized for deep learning papers using PyTorch or TensorFlow, specifically focusing on time-series domains. It uses Mermaid for diagramming and standard Python for pseudo-code representations.

    Why this beats prompting it yourself

    Generic prompts often hallucinate architecture details or fail to track tensor dimensions accurately across modules. This skill enforces a strict grounding rule, citing specific page numbers, section headers, and line numbers to ensure every explanation is verifiable.

    Use cases

    • Deconstruct complex time-series architectures like Transformers or Informers into digestible computation steps.
    • Debug the gap between a paper's mathematical description and its official GitHub implementation.
    • Understand the training vs. inference data flow, including inverse scaling and loss computation.
    • Evaluate the actual "novelty" of a model by comparing its modules against cited prior work.

    Known limitations

    Does not support running or executing code. Scope is limited to static analysis and pedagogical explanation of provided or accessible paper and code materials.

    Cannot execute or environment-test code. Requires access to paper text/code to avoid hallucination. Defaults to interactive stages unless full analysis is requested.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 6 days ago

    • Free with an Agensi account

    Requires

    Python
    Git

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Listed6 days ago

    What's inside

    Frequently Asked Questions

    Lingerskill specializes in the complex alignment of research paper math with actual GitHub code, specifically tracking tensor dimension changes [B, L, H, C] to show exactly how data flows through a model.

    The skill is optimized for any AI agent that supports Markdown and Mermaid.js diagrams, allowing it to render the architectural flowcharts and code-to-paper mappings natively in your chat interface.

    While it excels at time-series forecasting models (like Informers or Autoformers), it can effectively deconstruct any PyTorch or TensorFlow-based architecture that provides a clear paper-to-code relationship.

    No, Lingerskill performs static analysis and pedagogical breakdown; it cannot execute code, train models, or perform live debugging on your local machine.

    Simply invoke '$lingerskill' followed by a PDF link or paper title; the skill will then guide you through a step-by-step interactive analysis of the architecture, data pipeline, and source code.

    You receive the full skill logic optimized for deep learning analysis, including the specialized grounding rules for tensor shape tracking and the Mermaid diagramming templates for architectural mapping.

    New in Research & Analysis