重新思考 AI 訓練的影片儲存:打破跨影格依賴,提升隨機存取效率
Rethinking Video Storage for AI Training: Breaking Cross-Frame Dependencies for Efficient Random Access
codec4AI 提出結合稀疏解碼與解耦跨影格依賴編碼的影片儲存方案,旨在提升 AI 訓練時隨機存取影格的效率。與論文評測的 LeRobot v3 配置相比,DDRA H.265 的媒體負載減少 58.7%,數據載入吞吐量提升 1.45×,重建品質相若。另有實驗顯示,在合適的壓縮設定下,下游 VLA、WAM 和 VLM 表現相若;論文、程式碼及文件已公開。
🛑 Stop using LeRobot’s default video storage.
🚀 Upgrade your training-data infra with codec4AI!
🏛️ From Academician Wen Gao’s team at Peking University.
Video training needs fast random frame access. codec4AI brings sparse decoding and dependency-decoupled encoding together to make that access more efficient.
Against the LeRobot v3 configuration evaluated in our paper, DDRA H.265 achieves:
📦 58.7% less media payload
⚡ 1.45× data-loading throughput
🔬 At comparable reconstruction quality
🧠 What about training quality? Separate experiments show comparable downstream VLA, WAM and VLM performance at suitable compression settings.
🔧 Less unnecessary frame reconstruction.
💾 More efficient video storage.
🤖 Built around the access patterns of AI training.
🔓 Paper, code and documentation are public. Try it, benchmark it, and build with us! 🤝
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co