{"id":241,"date":"2025-10-21T11:58:43","date_gmt":"2025-10-21T03:58:43","guid":{"rendered":"https:\/\/snaili.com\/?p=241"},"modified":"2025-11-05T01:39:35","modified_gmt":"2025-11-04T17:39:35","slug":"pytorch%e5%ad%a6%e4%b9%a0%e5%8f%b2-%e4%ba%8c-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb%e4%b8%8e%e6%95%b0%e6%8d%aedemo%e5%a4%84%e7%90%86","status":"publish","type":"post","link":"https:\/\/snaili.com\/index.php\/2025\/10\/21\/pytorch%e5%ad%a6%e4%b9%a0%e5%8f%b2-%e4%ba%8c-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb%e4%b8%8e%e6%95%b0%e6%8d%aedemo%e5%a4%84%e7%90%86\/","title":{"rendered":"\u6df1\u5ea6\u5b66\u4e60\u4e0e\u5927\u8bed\u8a00\u6a21\u578b \uff08\u4e8c\uff09 \u8bba\u6587bg\u4e0eAbstract"},"content":{"rendered":"\n

\u8bba\u6587\u9898\u76ee\u4e3a\uff1aBLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models<\/p>\n\n\n\n

\u9898\u76ee\u5185\u5bb9\u7ffb\u8bd1\uff1a\u901a\u7528\u9884\u6d4b\u6a21\u578b<\/strong> \u7684 \u5e73\u8861\u62bd\u6837<\/strong> \u65f6\u95f4\u5e8f\u5217<\/strong> \u8bed\u6599\u5e93<\/p>\n\n\n\n

\u9898\u76ee\u5185\u5bb9\u5206\u89e3\uff1aUniversal Forecasting Models Balanced Sampling Time Series Corpus<\/p>\n\n\n\n

keywords\uff1alarge-scale time series dataset, balanced sampling, universal time series forecasting<\/p>\n\n\n\n

Abstract\uff1aThe advent of universal time series forecasting models has revo-lutionized zero-shot forecasting across diverse domains, yet the critical role of data diversity in training these models remains un-derexplored. Existing large-scale time series datasets often suffer from inherent biases and imbalanced distributions, leading to sub-optimal model performance and generalization. To address this gap, we introduce BLAST, a novel pre-training corpus designed to enhance data diversity through a balanced sampling strategy. First,BLAST incorporates 321 billion observations from publicly available datasets and employs a comprehensive suite of statistical metrics to characterize time series patterns. Then, to facilitate pattern-oriented sampling, the data is implicitly clustered using grid-based parti-tioning. Furthermore, by integrating grid sampling and grid mixup techniques, BLAST ensures a balanced and representative coverage of diverse patterns. Experimental results demonstrate that models pre-trained on BLAST achieve state-of-the-art performance with a fraction of the computational resources and training tokens re-quired by existing methods. Our findings highlight the pivotal role of data diversity in improving both training efficiency and model performance for the universal forecasting task.<\/p>\n\n\n\n

\u5207\u5165\u70b9\uff1aUniversal Time Series Forecasting Models zero-shot forecasting BLAST grid-based partitioning{\u6570\u636e\u901a\u8fc7 \u7f51\u683c\u5212\u5206\uff08grid-based partitioning\uff09<\/strong> \u8fdb\u884c\u9690\u5f0f\u805a\u7c7b\uff0c\u4ee5\u4fbf\u6309\u6a21\u5f0f\u8fdb\u884c\u91c7\u6837\u3002 \u4f7f\u7528\u4e86 \u7f51\u683c\u91c7\u6837\uff08grid sampling\uff09<\/strong> \u548c \u7f51\u683c\u6df7\u5408\uff08grid mixup\uff09<\/strong> \u6280\u672f\uff0c\u786e\u4fdd\u91c7\u6837\u7684\u65f6\u95f4\u5e8f\u5217\u5177\u6709\u4ee3\u8868\u6027\u548c\u5747\u8861\u6027\u3002}<\/p>\n\n\n\n

\u4e00\u3001abstract<\/strong><\/h2>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u8fd9\u91cc\u5927\u767d\u8bdd\u53ef\u4ee5\u7406\u89e3\u4e3a\uff1a\u7528\u901a\u7528\u6a21\u578b\u8bad\u7ec3\u7684\u4e0d\u597d\u4e0d\u5747\u8861 \u7528\u6211\u641c\u96c6\u7684\u6570\u636eBLAST \u6211\u8fd9\u4e2a\u91c7\u6837\u8fc7\u7a0b\u597d\u6570\u636e\u5747\u8861 \u5f97\u51fa\u6765\u7684\u7ed3\u8bba\u4f1a\u6bd4\u666e\u901a\u7684\u597d<\/p>\n\n\n\n

ai\u7528\u79d1\u5b66\u7684\u8bed\u8a00\u544a\u8bc9\u6211\uff1a<\/p>\n\n\n\n

1.\u73b0\u72b6\u95ee\u9898<\/strong>\uff1a
\u901a\u7528\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u6a21\u578b\u867d\u7136\u80fd\u505a\u96f6\u6837\u672c\u9884\u6d4b\uff0c\u4f46\u5b83\u4eec\u4f9d\u8d56\u7684\u73b0\u6709\u5927\u89c4\u6a21\u6570\u636e\u96c6\u5b58\u5728\u504f\u5dee\u548c\u5206\u5e03\u4e0d\u5747\uff0c\u5bfc\u81f4\u8bad\u7ec3\u6548\u679c\u548c\u6cdb\u5316\u80fd\u529b\u4e0d\u8db3\u3002<\/p>\n\n\n\n

2.\u6539\u8fdb<\/strong>\uff1a
\u6587\u7ae0\u63d0\u51fa\u7684 BLAST \u6570\u636e\u96c6<\/strong>\u901a\u8fc7 \u5747\u8861\u91c7\u6837\u7b56\u7565<\/strong>\uff08\u7f51\u683c\u5212\u5206\u3001\u7f51\u683c\u91c7\u6837\u3001\u7f51\u683c\u6df7\u5408\uff09\u89e3\u51b3\u4e86\u6570\u636e\u4e0d\u5747\u8861\u95ee\u9898\uff0c\u786e\u4fdd\u8986\u76d6\u66f4\u591a\u6a21\u5f0f\uff0c\u63d0\u5347\u6570\u636e\u591a\u6837\u6027\u3002<\/p>\n\n\n\n

3.\u7ed3\u8bba\u4f18\u52bf<\/strong>\uff1a
\u4f7f\u7528 BLAST \u9884\u8bad\u7ec3\u7684\u6a21\u578b\u4e0d\u4ec5\u6027\u80fd\u8fbe\u5230 SOTA\uff0c\u8fd8\u663e\u8457\u51cf\u5c11\u4e86\u8ba1\u7b97\u8d44\u6e90\u548c\u8bad\u7ec3\u6807\u8bb0\uff0c\u76f8\u6bd4\u4f20\u7edf\u65b9\u6cd5\u66f4\u9ad8\u6548\u3001\u66f4\u6cdb\u5316\u3002<\/p>\n\n\n\n

\u4e00\u53e5\u8bdd\u6982\u62ec\u5c31\u662f\uff1a\u901a\u8fc7\u5747\u8861\u91c7\u6837\u6784\u5efa\u4e86\u4e00\u4e2a\u591a\u6837\u5316\u7684\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\uff0c\u8fbe\u5230\u4e86\u964d\u672c\u589e\u6548\u7684\u76ee\u6807\uff1a\u52a0\u901f\u6a21\u578b\u8bad\u7ec3\u3001\u63d0\u5347\u6a21\u578b\u6cdb\u5316\u80fd\u529b<\/p>\n\n\n\n

\u4e8c\u3001background<\/strong><\/h2>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n

\u80cc\u666f\u4e3b\u8981\u4ecb\u7ecd\u7684\u5c31\u662f\uff1a\u5148\u4ecb\u7ecd\u4e86\u901a\u7528\u65f6\u95f4\u5e8f\u5217\u6a21\u578b \u7136\u540e\u6307\u51fa\u5176\u7f3a\u70b9\uff1a \u521d\u59cb\u5206\u5e03\u5f80\u5f80\u5b58\u5728\u4e25\u91cd\u5931\u8861 \u539f\u59cb\u6570\u636e\u4e2d\u51fa\u73b0\u5927\u91cf\u91cd\u590d\u6a21\u5f0f \u7136\u540e\u4e3e\u4e86\u4e00\u4e2a\u4f8b\u5b50\u7b80\u5355\u62bd\u6837 \u6216\u5206\u5c42\u62bd\u6837\u7b49\u7b80\u5316\u7684\u91c7\u6837\u7b56\u7565 \u7b80\u5355\u62bd\u6837\u5b8c\u5168\u5ffd\u7565\u4e86\u8fd9\u4e9b\u504f\u5dee\uff0c\u5206\u5c42\u62bd\u6837\u5374\u5e38\u5047\u8bbe\u540c\u4e00\u6570\u636e\u96c6\u6216\u57df\u5185\u7684\u6570\u636e\u5177\u6709\u76f8\u4f3c\u6a21\u5f0f \u4ece\u800c\u53d1\u73b0\u6a21\u578b\u7ed3\u679c\u53ef\u80fd\u8fc7\u5ea6\u62df\u5408\u9ad8\u9891\u6a21\u5f0f\uff0c\u540c\u65f6\u5bf9\u4f4e\u9891\u6a21\u5f0f\u62df\u5408 \u4e0d\u8db3\uff0c\u4ece\u800c\u635f\u5bb3\u5176\u6cdb\u5316\u80fd\u529b \u8fd9\u90fd\u662f\u8bad\u7ec3\u6570\u636e\u7684\u591a\u6837\u6027\u7f3a\u5931\u5bfc\u81f4\u7684 \u6240\u4ee5\u63d0\u51fa\u4e86\u4e00\u4e2a\u540d\u4e3aBLAST\uff08\u5e73\u8861\u91c7\u6837\u65f6\u95f4\u5e8f\u5217\u8bed\u6599\u5e93\uff09<\/p>\n\n\n\n

\u89c4\u6574\u4e00\u4e0b\u8bed\u8a00\uff1a
\u901a\u7528\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u6a21\u578b\u5728\u96f6\u6837\u672c\u9884\u6d4b\u65b9\u9762\u6709\u5f88\u5927\u6f5c\u529b\uff0c\u4f46\u5b83\u4eec\u4f9d\u8d56\u4e8e\u5927\u89c4\u6a21\u3001\u591a\u6837\u5316\u7684\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n

1.\u6307\u51fa\u95ee\u9898<\/strong><\/p>\n\n\n\n

    \n
  • \u73b0\u6709\u5927\u89c4\u6a21\u6570\u636e\u96c6\u867d\u7136\u6570\u91cf\u5e9e\u5927\uff0c\u4f46\u521d\u59cb\u5206\u5e03\u4e25\u91cd\u5931\u8861<\/strong>\uff0c\u5bfc\u81f4\u5927\u91cf\u91cd\u590d\u6a21\u5f0f\u3002<\/li>\n\n\n\n
  • \u7b80\u5355\u62bd\u6837\uff08naive sampling\uff09\u5b8c\u5168\u5ffd\u7565\u504f\u5dee\uff0c\u5206\u5c42\u62bd\u6837\uff08stratified sampling\uff09\u5047\u8bbe\u540c\u4e00\u6570\u636e\u96c6\u6216\u57df\u5185\u6a21\u5f0f\u76f8\u4f3c\uff0c\u4f46\u5b9e\u9645\u5e76\u4e0d\u603b\u6210\u7acb\u3002<\/li>\n\n\n\n
  • \u7ed3\u679c\uff1a\u6a21\u578b\u5bb9\u6613\u8fc7\u62df\u5408\u9ad8\u9891\u6a21\u5f0f<\/strong>\uff0c\u4f4e\u9891\u6a21\u5f0f\u62df\u5408\u4e0d\u8db3<\/strong>\uff0c\u6cdb\u5316\u80fd\u529b\u4e0b\u964d\u3002<\/li>\n\n\n\n
  • \u6839\u672c\u539f\u56e0\uff1a\u8bad\u7ec3\u6570\u636e\u7f3a\u4e4f\u591a\u6837\u6027<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n

    2.\u63d0\u51fa\u89e3\u51b3\u65b9\u6848<\/strong>
    \u4e3a\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u4f5c\u8005\u63d0\u51fa BLAST\uff08Balanced Sampling Time series corpus\uff09<\/strong>\uff0c\u901a\u8fc7\u5747\u8861\u91c7\u6837\u7b56\u7565\u6784\u5efa\u4e00\u4e2a\u591a\u6837\u5316\u7684\u9884\u8bad\u7ec3\u8bed\u6599\u5e93\u3002<\/p>\n\n\n\n

    \"\"<\/div><\/figure>\n\n\n\n

    \u8fd9\u91cc\u662f\u8bf4blast\u7684\u5177\u4f53\u5b9e\u73b0\u65b9\u6cd5\uff1a\u6574\u5408\u5e7f\u6cdb\u7684\u516c\u5f00\u53ef\u7528\u6570\u636e\u96c6 \u7ed3\u5408\u5404\u79cd\u7edf\u8ba1\u5c5e\u6027\u8fd8\u5168\u9762\u8868\u5f81\u6bcf\u79cd\u6a21\u5f0f \uff0c\u5982\u5e73\u7a33\u6027\u3001\u5b63\u8282\u6027\u3001\u6ce2\u52a8\u6027\u7b49\u3002\u7136\u540e\u901a\u8fc7\u79bb\u6563\u5316\u5904\u7406\u5408\u5e76\u4e3a\u7edf\u4e00\u7684\u7279\u5f81\u5411\u91cf \u76f4\u89c2\u7684\u8868\u793a\u6570\u636e \u4f7f\u7528\u7f51\u683c\u91c7\u6837\u548c\u7f51\u683c\u6df7\u5408\uff0c\u4ee5\u786e\u4fdd\u5404\u79cd\u6a21\u5f0f \u7684\u5e73\u8861\u548c\u4ee3\u8868\u6027\u8986\u76d6 \u901a\u8fc7\u5bf9blast\u548cTimeMoE\u6a21\u578b\u7684\u5bf9\u6bd4\u8bad\u7ec3 \u53ef\u4ee5\u770b\u51fablast\u7684\u6a21\u578b\u7684\u4f18\u52bf \u7136\u540e\u603b\u7ed3\u4e86\u4e00\u4e0b\u8d21\u732e\uff1a\u63d0\u51fa\u9884\u8bad\u7ec3\u6570\u636e\u591a\u6837\u6027\u5bf9\u8bad\u7ec3\u6548\u7387\u548c\u6a21\u578b \u6027\u80fd\u7684\u5f71\u54cd\uff0c\u901a\u8fc7\u5e73\u8861\u91c7\u6837\u6280\u672f\uff0c\u7814\u53d1\u51fablast\u6a21\u578b\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u57fa\u4e8eblast\u7684\u9884\u8bad\u7ec3\u5728\u51cf\u5c11\u8d44\u6e90\u548c \u6570\u636e\u9700\u6c42\u7684\u540c\u65f6\u53d6\u5f97\u4e86\u8f83\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

    \u89c4\u6574\u8bed\u8a00\uff1a<\/p>\n\n\n\n

    1.BLAST\u7684\u5b9e\u73b0\u65b9\u6cd5<\/strong><\/p>\n\n\n\n

      \n
    • \u6574\u5408\u6570\u636e<\/strong>\uff1a\u6536\u96c6\u5927\u91cf\u516c\u5f00\u53ef\u7528\u7684\u6570\u636e\u96c6\uff0c\u603b\u89c4\u6a21\u8fbe\u5230 3210 \u4ebf\u89c2\u6d4b\u503c\u3002<\/li>\n\n\n\n
    • \u6a21\u5f0f\u8868\u5f81<\/strong>\uff1a\u4e0d\u4f9d\u8d56\u6570\u636e\u96c6\u6216\u9886\u57df\u6807\u7b7e\uff0c\u800c\u662f\u901a\u8fc7\u7edf\u8ba1\u5c5e\u6027\uff08\u5982\u5e73\u7a33\u6027\u3001\u5b63\u8282\u6027\u3001\u6ce2\u52a8\u6027\u7b49\uff09\u5168\u9762\u523b\u753b\u65f6\u95f4\u5e8f\u5217\u6a21\u5f0f\u3002<\/li>\n\n\n\n
    • \u7279\u5f81\u5904\u7406<\/strong>\uff1a\u5c06\u8fd9\u4e9b\u5c5e\u6027\u79bb\u6563\u5316\uff0c\u5408\u5e76\u4e3a\u7edf\u4e00\u7684\u7279\u5f81\u5411\u91cf\uff0c\u5e76\u6295\u5c04\u5230\u4f4e\u7ef4\u7a7a\u95f4\uff0c\u76f4\u89c2\u5c55\u793a\u6570\u636e\u5206\u5e03\u4e0d\u5747\u8861\u3002<\/li>\n\n\n\n
    • \u5747\u8861\u91c7\u6837<\/strong>\uff1a\u5728\u4f4e\u7ef4\u7a7a\u95f4\u4e2d\u91c7\u7528 \u7f51\u683c\u91c7\u6837\uff08grid sampling\uff09<\/strong> \u548c \u7f51\u683c\u6df7\u5408\uff08grid mixup\uff09<\/strong>\uff0c\u786e\u4fdd\u6a21\u5f0f\u7684\u5747\u8861\u548c\u4ee3\u8868\u6027\u8986\u76d6\u3002<\/li>\n<\/ul>\n\n\n\n

      2.\u5b9e\u9a8c\u9a8c\u8bc1<\/strong><\/p>\n\n\n\n

        \n
      • \u4f7f\u7528 BLAST \u8bad\u7ec3 TimeMoE \u6a21\u578b\uff0c\u4e0e\u539f\u59cb TimeMoE \u5bf9\u6bd4\uff1a\n
          \n
        • \u539f\u59cb\uff1a419B tokens + 128 A100 GPUs<\/li>\n\n\n\n
        • BLAST\uff1a78B tokens + 8 A100 GPUs<\/li>\n<\/ul>\n<\/li>\n\n\n\n
        • \u7ed3\u679c\uff1aBLAST \u5728\u66f4\u5c11\u8d44\u6e90\u4e0b\u5b9e\u73b0 SOTA \u6027\u80fd\uff0c\u8bc1\u660e\u6570\u636e\u591a\u6837\u6027\u663e\u8457\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u548c\u6a21\u578b\u6027\u80fd\u3002<\/li>\n<\/ul>\n\n\n\n

          3.\u8d21\u732e\u603b\u7ed3<\/strong><\/p>\n\n\n\n

          \u6784\u5efa BLAST \u8bed\u6599\u5e93\uff0c\u5b9e\u9a8c\u8868\u660e\u5176\u5728\u964d\u4f4e\u8d44\u6e90\u548c\u6570\u636e\u9700\u6c42\u7684\u540c\u65f6\u6027\u80fd\u66f4\u4f18\u3002<\/p>\n\n\n\n

          \u9996\u6b21\u7cfb\u7edf\u7814\u7a76\u9884\u8bad\u7ec3\u6570\u636e\u591a\u6837\u6027\u5bf9\u8bad\u7ec3\u6548\u7387\u548c\u6a21\u578b\u6027\u80fd\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n

          \u63d0\u51fa\u57fa\u4e8e\u6a21\u5f0f\u7684\u5e73\u8861\u91c7\u6837\u6280\u672f\uff08\u7edf\u8ba1\u5c5e\u6027 + \u7f51\u683c\u5212\u5206 + mixup\uff09\u3002<\/p>\n\n\n\n

          \u4e09\u3001\u9884\u5907\u77e5\u8bc6<\/strong><\/h2>\n\n\n\n
          \"\"<\/div><\/figure>\n\n\n\n

          1. Large-scale Time Series Forecasting Dataset<\/strong><\/h4>\n\n\n\n
            \n
          • \u4e00\u4e2a\u5927\u89c4\u6a21\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u6570\u636e\u96c6 D<\/strong> \u901a\u5e38\u7531\u591a\u4e2a\u5b50\u6570\u636e\u96c6\u7ec4\u6210\uff1a
            D={D1,D2,\u2026,DN}D = \\{ D_1, D_2, \\dots, D_N \\}D={D1\u200b,D2\u200b,\u2026,DN\u200b}<\/li>\n\n\n\n
          • \u6bcf\u4e2a\u5b50\u6570\u636e\u96c6 DnD_nDn\u200b \u5305\u542b KnK_nKn\u200b \u6761\u65f6\u95f4\u5e8f\u5217\uff1a
            Dn={X1n,X2n,\u2026,XKnn}D_n = \\{ X^n_1, X^n_2, \\dots, X^n_{K_n} \\}Dn\u200b={X1n\u200b,X2n\u200b,\u2026,XKn\u200bn\u200b}<\/li>\n\n\n\n
          • \u6bcf\u6761\u65f6\u95f4\u5e8f\u5217 XknX^n_kXkn\u200b \u6709 TknT^n_kTkn\u200b \u4e2a\u65f6\u95f4\u6b65\uff1a
            Xkn={xk,1n,xk,2n,\u2026,xk,Tknn}X^n_k = \\{ x^n_{k,1}, x^n_{k,2}, \\dots, x^n_{k,T^n_k} \\}Xkn\u200b={xk,1n\u200b,xk,2n\u200b,\u2026,xk,Tkn\u200bn\u200b}<\/li>\n\n\n\n
          • \u91cd\u70b9<\/strong>\uff1a\u4e0d\u540c\u5b50\u6570\u636e\u96c6\u7684\u5927\u5c0f\u548c\u5e8f\u5217\u957f\u5ea6\u5dee\u5f02\u5f88\u5927\u3002<\/li>\n<\/ul>\n\n\n\n
            \n\n\n\n

            2. Sampling Strategies<\/strong><\/h4>\n\n\n\n
              \n
            • \u539f\u59cb\u65f6\u95f4\u5e8f\u5217\u4e0d\u80fd\u76f4\u63a5\u7528\u4e8e\u8bad\u7ec3\uff0c\u9700\u8981\u901a\u8fc7 \u6ed1\u52a8\u7a97\u53e3\uff08sliding window\uff09<\/strong> \u751f\u6210\u5019\u9009\u6837\u672c\u96c6\u5408 W\u3002<\/li>\n\n\n\n
            • \u91c7\u6837\u7b56\u7565<\/strong>\u7684\u76ee\u6807\uff1a\u4ece\u5019\u9009\u6837\u672c\u4e2d\u9009\u62e9\u6700\u7ec8\u7528\u4e8e\u8bad\u7ec3\u7684\u6837\u672c\u3002<\/li>\n\n\n\n
            • \u5e38\u89c1\u7b56\u7565\uff1a\n
                \n
              • Naive Sampling<\/strong>\uff1a\u968f\u673a\u5747\u5300\u62bd\u6837\uff0c\u5ffd\u7565\u6570\u636e\u5206\u5e03\u504f\u5dee\u3002<\/li>\n\n\n\n
              • Stratified Sampling<\/strong>\uff1a\u6309\u6570\u636e\u96c6\u6216\u9886\u57df\u5206\u5c42\u62bd\u6837\uff0c\u4f46\u5047\u8bbe\u540c\u4e00\u57df\u6a21\u5f0f\u76f8\u4f3c\uff0c\u8fd9\u4e2a\u5047\u8bbe\u4e0d\u603b\u6210\u7acb\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
                \n\n\n\n

                3. Universal Forecasting Models<\/strong><\/h4>\n\n\n\n
                  \n
                • \u901a\u7528\u9884\u6d4b\u6a21\u578b<\/strong>\uff1a\u5728\u5927\u89c4\u6a21\u65f6\u95f4\u5e8f\u5217\u6570\u636e\u4e0a\u9884\u8bad\u7ec3\uff0c\u80fd\u591f\u5728\u4e0d\u540c\u9886\u57df\u5b9e\u73b0 \u96f6\u6837\u672c\u9884\u6d4b\uff08zero-shot forecasting\uff09<\/strong>\u3002<\/li>\n\n\n\n
                • \u4f8b\u5b50\uff1aTimeMoE<\/strong>\uff0c\u662f\u76ee\u524d SOTA \u6a21\u578b\u4e4b\u4e00\uff0c\u9884\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u8fbe 309B observations\u3002<\/li>\n<\/ul>\n\n\n\n
                  \"\"<\/div><\/figure>\n\n\n\n
                    \n
                  1. Universal Time Series Forecasting<\/strong>\n
                      \n
                    • \u4ecb\u7ecd\u901a\u7528\u65f6\u95f4\u5e8f\u5217\u9884\u6d4b\u6a21\u578b\u7684\u80cc\u666f\uff1a\n
                        \n
                      • \u7075\u611f\u6765\u81ea AI \u7684\u7a81\u7834\uff0c\u76ee\u6807\u662f\u901a\u8fc7\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u5b9e\u73b0\u8de8\u9886\u57df\u96f6\u6837\u672c\u9884\u6d4b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                      • \u5206\u7c7b\uff1a\n
                          \n
                        • Encoder-only<\/strong>\uff08\u63a9\u7801\u7f16\u7801\u7b56\u7565\uff09<\/li>\n\n\n\n
                        • Decoder-only<\/strong>\uff08\u81ea\u56de\u5f52\u9884\u8bad\u7ec3\uff09<\/li>\n\n\n\n
                        • Encoder-Decoder<\/strong>\uff08\u5b8c\u6574 Transformer \u6846\u67b6\uff09<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                        • \u6700\u65b0\u6280\u672f\uff1a\n
                            \n
                          • Mixture-of-Experts<\/li>\n\n\n\n
                          • Long-context modeling<\/li>\n\n\n\n
                          • Hierarchical modeling<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                          • \u8fd8\u63d0\u5230\u4e00\u4e9b\u63a2\u7d22 Transformer \u4e4b\u5916\u67b6\u6784\u7684\u7814\u7a76\u3002<\/li>\n\n\n\n
                          • \u603b\u7ed3<\/strong>\uff1a\u8fd9\u4e9b\u6a21\u578b\u901a\u8fc7\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u5c55\u73b0\u4e86\u5f3a\u5927\u7684\u96f6\u6837\u672c\u9884\u6d4b\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                          • Time Series Forecasting Pre-training Corpus<\/strong>\n
                              \n
                            • \u5f3a\u8c03\u65e0\u8bba\u6a21\u578b\u67b6\u6784\u5982\u4f55\uff0c\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6570\u636e\u662f\u57fa\u7840\u3002<\/li>\n\n\n\n
                            • \u5217\u4e3e\u5df2\u6709\u6570\u636e\u96c6\uff1a\n
                                \n
                              • ForecastPFN\uff08\u7eaf\u5408\u6210\u6570\u636e\uff09<\/li>\n\n\n\n
                              • Chronos\uff08Monash + M-competitions + \u5408\u6210\u6570\u636e\uff0c84B observations\uff09<\/li>\n\n\n\n
                              • MOIRAI\uff08LOTSA\uff0c231B\uff09<\/li>\n\n\n\n
                              • Timer\uff08UTSD\uff0c1B\uff09<\/li>\n\n\n\n
                              • TimeMoE\uff08Time-300B\uff0c309B\uff09<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                              • \u6307\u51fa\u95ee\u9898<\/strong>\uff1a\u8fd9\u4e9b\u5de5\u4f5c\u4e3b\u8981\u5173\u6ce8\u6570\u636e\u89c4\u6a21\uff0c\u6ca1\u6709\u7cfb\u7edf\u7814\u7a76\u6570\u636e\u591a\u6837\u6027<\/strong>\u3002<\/li>\n\n\n\n
                              • \u5f15\u51fa\u4f60\u7684\u8d21\u732e<\/strong>\uff1a\u63d0\u51fa BLAST\uff08321B observations\uff09\uff0c\u91c7\u7528\u5747\u8861\u91c7\u6837\u7b56\u7565\u4fdd\u8bc1\u591a\u6837\u6027\uff0c\u5e76\u9a8c\u8bc1\u5176\u5728\u8bad\u7ec3\u6548\u7387\u548c\u6027\u80fd\u4e0a\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n

                                <\/p>\n\n\n\n

                                demo\u6570\u636e\u5904\u7406<\/h2>\n\n\n\n

                                \u9e22\u5c3e\u82b1\u6570\u636e\u5904\u7406 \u6211\u8bb0\u5f97\u4e4b\u524d\u505a\u8fc7\u8fd9\u4e2a\u6765\u7740 \u4f46\u662f\u4e3a\u4ec0\u4e48\u6211\u6ca1\u6709\u600e\u4e48\u505a\u7684\u5370\u8c61<\/p>\n\n\n\n

                                \u4f7f\u7528 PyTorch \u8fdb\u884c\u6570\u636e\u5904\u7406\u3001\u6a21\u578b\u6784\u5efa\u3001\u8bad\u7ec3\u548c\u8bc4\u4f30\uff0c\u4ee5\u53ca\u5982\u4f55\u5bf9\u5206\u7c7b\u7ed3\u679c\u8fdb\u884c\u53ef\u89c6\u5316\u5206\u6790<\/p>\n\n\n\n

                                \"\"<\/div><\/figure>\n\n\n\n

                                \u4e0d\u592a\u61c2 \u4f46\u662f\u4ee3\u7801\u6ca1\u95ee\u9898<\/p>\n\n\n\n

                                \"\"<\/div><\/figure>\n\n\n\n
                                \"\"<\/div><\/figure>\n\n\n\n

                                \u54e6\u4e07\u80fd\u7684\u9e22\u5c3e\u82b1~<\/p>\n","protected":false},"excerpt":{"rendered":"

                                \u8bba\u6587\u9898\u76ee\u4e3a\uff1aBLAST: Balanced Sampling Time Series Corpus for U […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9,1],"tags":[],"class_list":["post-241","post","type-post","status-publish","format-standard","hentry","category-9","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/posts\/241","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/comments?post=241"}],"version-history":[{"count":9,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/posts\/241\/revisions"}],"predecessor-version":[{"id":361,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/posts\/241\/revisions\/361"}],"wp:attachment":[{"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/media?parent=241"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/categories?post=241"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/snaili.com\/index.php\/wp-json\/wp\/v2\/tags?post=241"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}