fiftyone-embeddings-visualization
以 2D 視覺化呈現資料集嵌入向量
理解複雜的圖像資料集需要看到樣本在嵌入空間中的關聯性。此技能引導您計算嵌入向量,並使用 UMAP 或 t-SNE 創建 2D 視覺化,揭示您 FiftyOne 資料集中的聚類、異常值和類別分佈。
스킬 ZIP 다운로드
Claude에서 업로드
설정 → 기능 → 스킬 → 스킬 업로드로 이동
토글을 켜고 사용 시작
테스트해 보기
"fiftyone-embeddings-visualization" 사용 중입니다. Visualize my coco-subset dataset in 2D
예상 결과:
- Computing CLIP embeddings for 500 samples...
- Running UMAP dimensionality reduction...
- ✓ Visualization ready (brain_key: exploration)
- → Open http://localhost:5151/
- → Click the Embeddings panel icon
- → Select exploration from the dropdown
- → Use Color by to select ground_truth.label
"fiftyone-embeddings-visualization" 사용 중입니다. Find outliers in my image classification dataset
예상 결과:
- Computing uniqueness scores for all samples...
- Found 50 most unique samples (outliers)
- → Open http://localhost:5151/
- → Go to Embeddings panel
- → Select outliers brain_key
- → Use sort field to view by uniqueness score
보안 감사
안전This is a pure documentation skill containing only instructions in SKILL.md format. The static scanner produced false positives by misinterpreting: (1) SHA256 hashes as weak cryptographic algorithms and C2 keywords, (2) markdown code block backticks as shell command execution, and (3) documentation URLs as hardcoded network endpoints. No executable code, scripts, network operations, or external command execution is present. This skill merely guides users on using legitimate FiftyOne MCP server tools.
위험 요인
🌐 네트워크 접근 (11)
⚙️ 외부 명령어 (88)
품질 점수
만들 수 있는 것
探索訓練資料
在訓練前視覺化圖像資料集中的類別分佈並找出標籤錯誤的樣本。
分析模型預測
在嵌入空間中比較真實標籤與預測結果,以識別混淆模式。
尋找資料集異常值
檢測可能表明資料品質問題或罕見案例的異常或獨特樣本。
이 프롬프트를 사용해 보세요
Set context to my dataset and launch the app. Compute CLIP embeddings and create a UMAP visualization. Then tell me how to view it in the FiftyOne App embeddings panel.
My dataset has CLIP embeddings already. Compute a visualization and then find the 50 most unique (outlier) samples. Show me how to identify them in the embeddings panel.
Generate a visualization for my classification dataset and color points by ground truth label. Describe what well-separated vs overlapping clusters indicate about class similarity.
Create a t-SNE visualization using DINOv2 embeddings for publication-quality plots. What are the advantages of t-SNE over UMAP for this use case?
모범 사례
- 使用 CLIP 嵌入向量進行語義相似性分析;僅使用 DINOv2 進行視覺相似性分析
- 從 UMAP 開始進行快速探索;切換至 t-SNE 以獲得出版品質的局部結構
- 計算獨特性分數以比視覺檢查更可靠地找出異常值
피하기
- 跳過 launch_app() 呼叫 - brain 運算子需要 app executor
- 在執行 compute_visualization 之前忘記計算嵌入向量
- 在 compute_similarity 和 compute_visualization 之間使用錯誤的 brain_key