\u5176\u4e2d\u6700\u5173\u952e\u7684\u7ec6\u8282\u5728\u6700\u540e\u4e00\u884c\uff1a\u4f60\u5fc5\u987b\u4ece\u5df2\u4fdd\u5b58 epoch \u7684\u4e0b\u4e00\u4e2a epoch \u5f00\u59cb\uff0c\u800c\u4e0d\u662f\u4ece\u96f6\u91cd\u65b0\u5f00\u59cb\u3002\u8bad\u7ec3\u5faa\u73af\u4e5f\u9700\u8981\u76f8\u5e94\u8c03\u6574\uff0c\u4ee5\u4fbf\u9075\u5faa\u8fd9\u4e2a\u8d77\u70b9\u3002\u53ea\u9700\u4fee\u6539\u5faa\u73af\u8303\u56f4\uff0c\u4f7f\u5176\u4ece\u6062\u590d\u540e\u7684 epoch \u5f00\u59cb\uff0c\u5c31\u80fd\u907f\u514d\u91cd\u590d\u8ba1\u7b97\u5df2\u7ecf\u5b8c\u6210\u7684\u8f6e\u6b21\u3002\u4f60\u7684\u4efb\u52a1\u5c06\u4ece\u4e2d\u65ad\u5904\u7cbe\u786e\u7eed\u4e0a\uff0c\u4fdd\u7559\u6b64\u524d\u6240\u6709\u8fdb\u5ea6\u3002<\/p>\n

for epoch in range(start_epoch, total_epochs):\r\n    train_one_epoch(model, train_loader, optimizer, scheduler)\r\n    validate(model, val_loader)\r\n    save_checkpoint(model, optimizer, scheduler, epoch)<\/code><\/pre>\n
\u50cf fastai \u8fd9\u6837\u7684\u6846\u67b6\uff0c\u8fd8\u80fd\u628a\u8fd9\u4e2a\u8fc7\u7a0b\u8fdb\u4e00\u6b65\u7b80\u5316\u3002fit_one_cycle<\/code> \u51fd\u6570\u4f1a\u5e94\u7528\u5faa\u73af\u5f0f\u5b66\u4e60\u7387\u4e0e\u52a8\u91cf\u7b56\u7565\u3002\u5982\u679c\u4e2d\u65ad\u540e\u5904\u7406\u4e0d\u5f53\uff0c\u6574\u4e2a\u5faa\u73af\u4f1a\u4ece\u5934\u5f00\u59cb\uff0c\u5bfc\u81f4\u7ed3\u679c\u4e0e\u672a\u4e2d\u65ad\u8bad\u7ec3\u4e0d\u540c\u3002fastai \u901a\u8fc7 start_epoch<\/code> \u53c2\u6570\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\u3002\u4f60\u53ea\u9700\u91cd\u65b0\u5b9e\u4f8b\u5316 learner\uff0c\u5e76\u7528\u4e0b\u4e00\u4e2a epoch \u7f16\u53f7\u8c03\u7528 fit_one_cycle<\/code>\u3002\u6846\u67b6\u4f1a\u81ea\u52a8\u52a0\u8f7d\u6b64\u524d\u4e3a\u8be5 epoch \u4fdd\u5b58\u7684\u6587\u4ef6\uff0c\u65e0\u9700\u624b\u52a8\u91cd\u65b0\u52a0\u8f7d\u6743\u91cd\u3002\u8fd9\u6837\uff0c\u5b66\u4e60\u7387\u548c\u52a8\u91cf\u5faa\u73af\u5c31\u80fd\u4ece\u51c6\u786e\u4f4d\u7f6e\u7ee7\u7eed\u6267\u884c\uff0c\u8bad\u7ec3\u7b56\u7565\u5f97\u4ee5\u65e0\u7f1d\u5ef6\u7eed\u3002<\/p>\n
\u91cd\u65b0\u5efa\u7acb\u6570\u636e\u52a0\u8f7d\u5668\u4e0e\u968f\u673a\u79cd\u5b50<\/h3>\n
\u5982\u679c\u6570\u636e\u7ba1\u9053\u4e0d\u80fd\u4e00\u81f4\uff0c\u4ec5\u4ec5\u6062\u590d\u6a21\u578b\u4e5f\u6ca1\u6709\u610f\u4e49\u3002\u968f\u673a\u6570\u751f\u6210\u5668\u72b6\u6001\u51b3\u5b9a\u4e86\u6570\u636e\u52a0\u8f7d\u5668\u5982\u4f55\u6253\u4e71\u6837\u672c\u987a\u5e8f\u3002\u6bcf\u4e2a epoch \u90fd\u4f1a\u4ea7\u751f\u4e0d\u540c\u7684\u987a\u5e8f\uff0c\u800c\u56fe\u50cf\u589e\u5f3a\u4e2d\u7684\u5404\u79cd\u968f\u673a\u53d8\u6362\u4e5f\u90fd\u4f9d\u8d56 RNG \u72b6\u6001\u3002<\/p>\n
\u6062\u590d\u8fd9\u4e2a\u72b6\u6001\u5176\u5b9e\u5f88\u7b80\u5355\u3002\u53ea\u9700\u628a\u4e4b\u524d\u4fdd\u5b58\u7684 RNG \u503c\u91cd\u65b0\u4f20\u7ed9 PyTorch\uff0c\u6570\u636e\u52a0\u8f7d\u5668\u5c31\u4f1a\u751f\u6210\u4e0e\u672a\u5d29\u6e83\u65f6\u5b8c\u5168\u76f8\u540c\u7684\u6837\u672c\u5e8f\u5217\u3002\u8fd9\u79cd\u53ef\u590d\u73b0\u6027\u5bf9\u4e8e\u9a8c\u8bc1\u4e00\u81f4\u6027\u975e\u5e38\u91cd\u8981\u3002\u4f60\u7684\u635f\u5931\u66f2\u7ebf\u80fd\u591f\u5728\u4e2d\u65ad\u524d\u540e\u4fdd\u6301\u53ef\u6bd4\uff0c\u68af\u5ea6\u6a21\u5f0f\u4e5f\u4f1a\u66f4\u52a0\u7a33\u5b9a\u53ef\u9884\u6d4b\u3002<\/p>\n
torch.set_rng_state(checkpoint['rng_state'])<\/code><\/pre>\n
\u4f60\u8fd8\u5fc5\u987b\u6062\u590d\u5404\u4e2a worker \u4e13\u5c5e\u7684\u968f\u673a\u79cd\u5b50\u3002\u6570\u636e\u52a0\u8f7d\u5668\u901a\u5e38\u4f1a\u542f\u7528\u591a\u4e2a worker \u8fdb\u7a0b\uff0c\u800c\u6bcf\u4e2a worker \u90fd\u7ef4\u62a4\u7740\u81ea\u5df1\u7684\u968f\u673a\u72b6\u6001\u3002\u4f60\u9700\u8981\u5728\u91cd\u65b0\u521b\u5efa\u6570\u636e\u52a0\u8f7d\u5668\u5b9e\u4f8b\u4e4b\u524d\uff0c\u5148\u8bbe\u7f6e\u8fd9\u4e9b\u79cd\u5b50\u3002\u8fd9\u6837\u624d\u80fd\u786e\u4fdd\u6bcf\u4e2a worker \u90fd\u590d\u73b0\u51fa\u539f\u672c\u7684\u6570\u636e\u6253\u4e71\u6a21\u5f0f\u3002\u81f3\u6b64\uff0c\u4f60\u7684\u8bad\u7ec3\u73af\u5883\u624d\u7b97\u771f\u6b63\u56de\u5230\u4e86\u6545\u969c\u524d\u7684\u72b6\u6001\u3002<\/p>\n
\u6700\u540e\u7684\u9a8c\u8bc1\u6b65\u9aa4\uff0c\u7528\u4e8e\u786e\u8ba4\u6062\u590d\u662f\u5426\u771f\u7684\u6210\u529f\u3002\u5148\u8dd1\u51e0\u4e2a batch\uff0c\u5e76\u628a\u635f\u5931\u503c\u4e0e\u5d29\u6e83\u524d\u7684\u65e5\u5fd7\u8fdb\u884c\u5bf9\u6bd4\u3002\u4e8c\u8005\u5e94\u5f53\u5927\u81f4\u63a5\u8fd1\uff0c\u68af\u5ea6\u8303\u6570\u4e5f\u5e94\u843d\u5728\u5408\u7406\u8303\u56f4\u5185\u3002\u5982\u679c\u504f\u5dee\u660e\u663e\uff0c\u5c31\u8bf4\u660e\u6062\u590d\u8fc7\u7a0b\u5b58\u5728\u95ee\u9898\u3002\u5c3d\u65e9\u53d1\u73b0\u8fd9\u4e9b\u9519\u8bef\uff0c\u80fd\u907f\u514d\u4f60\u5728\u635f\u574f\u72b6\u6001\u4e0b\u7ee7\u7eed\u6d6a\u8d39\u6570\u5c0f\u65f6\u7b97\u529b\u3002<\/p>\n
\u4f60\u7684\u68c0\u67e5\u70b9\u52a0\u8f7d\u7b56\u7565\uff0c\u4e5f\u4f1a\u76f4\u63a5\u5f71\u54cd\u6062\u590d\u65f6\u95f4\u3002\u4e00\u4e2a\u7ed3\u6784\u6e05\u6670\u3001\u7ec4\u7ec7\u826f\u597d\u7684\u68c0\u67e5\u70b9\uff0c\u80fd\u8ba9\u4f60\u8fc5\u901f\u5b9a\u4f4d\u5230\u6b63\u786e\u6587\u4ef6\uff0c\u5e76\u987a\u5229\u5b8c\u6210\u52a0\u8f7d\uff0c\u65e0\u9700\u590d\u6742\u89e3\u6790\u3002\u8fd9\u6837\uff0c\u4f60\u7684\u4efb\u52a1\u901a\u5e38\u53ef\u4ee5\u5728\u51e0\u5206\u949f\u5185\u6062\u590d\u5230\u5b8c\u6574\u8fd0\u884c\u72b6\u6001\u3002\u9ad8\u6548\u7684\u6062\u590d\u673a\u5236\uff0c\u4f1a\u628a\u4e00\u573a\u672c\u53ef\u80fd\u5f88\u4e25\u91cd\u7684\u6545\u969c\uff0c\u53d8\u6210\u4e00\u6b21\u77ed\u6682\u7684\u6682\u505c\uff0c\u8ba9\u8bad\u7ec3\u4ee5\u6700\u5c0f\u505c\u673a\u3001\u6700\u5927\u628a\u63e1\u7ee7\u7eed\u63a8\u8fdb\u3002<\/p>\n
\u8c03\u6574\u5b66\u4e60\u7387\u4ee5\u7ee7\u7eed\u8bad\u7ec3<\/strong><\/h2>\n
\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u4fdd\u5b58\u7740\u8bad\u7ec3\u8fdb\u5c55\u4e2d\u7684\u5173\u952e\u4fe1\u606f\u3002\u5b83\u8bb0\u5f55\u4e86\u5f53\u524d\u5728\u8870\u51cf\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\u3002\u4ec5\u6062\u590d\u6a21\u578b\u548c\u4f18\u5316\u5668\uff0c\u5e76\u4e0d\u610f\u5473\u7740\u6062\u590d\u5de5\u4f5c\u5df2\u7ecf\u5b8c\u6210\uff0c\u8c03\u5ea6\u5668\u72b6\u6001\u540c\u6837\u5fc5\u987b\u56de\u5230\u5d29\u6e83\u524d\u7684\u4f4d\u7f6e\u3002\u5982\u679c\u4f60\u65b0\u5efa\u4e00\u4e2a\u8c03\u5ea6\u5668\uff0c\u6574\u4e2a\u5b66\u4e60\u7387\u8ba1\u5212\u5c31\u4f1a\u88ab\u6253\u4e71\uff0c\u56e0\u4e3a\u6a21\u578b\u6b64\u65f6\u672c\u5e94\u4f7f\u7528\u540e\u671f\u66f4\u4f4e\u7684\u5b66\u4e60\u7387\u3002<\/p>\n
\u8ba1\u7b97\u6b63\u786e\u7684 step \u548c epoch<\/h3>\n
\u8c03\u5ea6\u5668\u901a\u5e38\u6839\u636e step \u6216 epoch \u6765\u51b3\u5b9a\u5f53\u524d\u5e94\u4f7f\u7528\u7684\u5b66\u4e60\u7387\u3002\u4f60\u5fc5\u987b\u77e5\u9053\u5d29\u6e83\u53d1\u751f\u65f6\u6240\u5904\u7684\u7cbe\u786e\u4f4d\u7f6e\u3002\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u516c\u5f0f\uff1acurrent_step = epoch * len(train_loader) + batch_index<\/code>\u3002\u8fd9\u4e2a\u8ba1\u7b97\u7ed3\u679c\u7ed9\u51fa\u4e86\u4f60\u5728\u5b66\u4e60\u7387\u8ba1\u5212\u4e2d\u7684\u7cbe\u786e step \u7f16\u53f7\u3002\u7136\u540e\uff0c\u5c06 last_epoch<\/code> \u53c2\u6570\u8bbe\u7f6e\u4e3a\u8fd9\u4e2a step \u503c\uff0c\u8c03\u5ea6\u5668\u5c31\u80fd\u4ece\u90a3\u4e2a\u51c6\u786e\u4f4d\u7f6e\u7ee7\u7eed\u8fd0\u884c\u3002\u5b66\u4e60\u7387\u66f2\u7ebf\u5c06\u8de8\u8d8a\u4e2d\u65ad\u70b9\u4fdd\u6301\u5e73\u6ed1\uff0c\u6a21\u578b\u5728\u540e\u7eed\u6bcf\u4e00\u6b65\u90fd\u80fd\u83b7\u5f97\u6b63\u786e\u7684\u5b66\u4e60\u7387\u3002\u8fd9\u79cd\u505a\u6cd5\u53ef\u786e\u4fdd\u6062\u590d\u8fc7\u7a0b\u4e0d\u4f1a\u7834\u574f\u539f\u672c\u7684\u8bad\u7ec3\u8282\u594f\u3002<\/p>\n
\u5bf9\u4e8e\u591a\u5468\u671f\u8c03\u5ea6\u7b56\u7565\u6765\u8bf4\uff0cepoch \u7f16\u53f7\u540c\u6837\u91cd\u8981\u3002\u4f59\u5f26\u9000\u706b\uff08cosine annealing\uff09\u548c one-cycle \u7b56\u7565\u90fd\u4f9d\u8d56\u4e8e\u5b8c\u6574\u5468\u671f\u957f\u5ea6\u3002\u4f60\u7684\u68c0\u67e5\u70b9\u4e2d\u4fdd\u5b58\u4e86\u6700\u540e\u4e00\u4e2a\u5df2\u5b8c\u6210\u7684 epoch\uff0c\u53ea\u9700\u5728\u6b64\u57fa\u7840\u4e0a\u52a0\u4e00\uff0c\u518d\u4f20\u5165\u8c03\u5ea6\u5668\u6784\u9020\u51fd\u6570\uff0c\u8c03\u5ea6\u5668\u5c31\u80fd\u6b63\u786e\u8ba1\u7b97\u5269\u4f59\u8870\u51cf\u8fc7\u7a0b\u3002\u8fd9\u4e00\u6b65\u80fd\u8ba9\u6062\u590d\u8fc7\u7a0b\u907f\u514d\u62cd\u8111\u888b\u731c\u6d4b\u3002<\/p>\n
\u91cd\u65b0\u521d\u59cb\u5316\u8c03\u5ea6\u5668\u4ee5\u4fdd\u6301\u8fde\u7eed\u6027<\/h3>\n
\u8bad\u7ec3\u5d29\u6e83\u540e\uff0c\u4e0d\u8981\u65b0\u5efa\u4e00\u4e2a\u5168\u65b0\u7684\u8c03\u5ea6\u5668\u3002\u65b0\u7684\u8c03\u5ea6\u5668\u4f1a\u628a\u5b66\u4e60\u7387\u91cd\u7f6e\u4e3a\u521d\u59cb\u503c\uff0c\u800c\u6b64\u65f6\u6a21\u578b\u901a\u5e38\u5df2\u7ecf\u8fdb\u5165\u8bad\u7ec3\u540e\u671f\uff0c\u7406\u5e94\u4f7f\u7528\u66f4\u4f4e\u5b66\u4e60\u7387\u3002\u7a81\u7136\u56de\u5230\u8f83\u9ad8\u5b66\u4e60\u7387\uff0c\u53ef\u80fd\u5bfc\u81f4\u68af\u5ea6\u4e0d\u7a33\u5b9a\u3001\u635f\u5931\u66b4\u6da8\uff0c\u751a\u81f3\u6743\u91cd\u53d1\u6563\u3002\u8fd9\u6837\u7684\u9519\u8bef\uff0c\u4f1a\u8ba9\u4f60\u4e4b\u524d\u7684\u6062\u590d\u52aa\u529b\u524d\u529f\u5c3d\u5f03\uff0c\u751a\u81f3\u8feb\u4f7f\u4f60\u518d\u6b21\u91cd\u542f\u8bad\u7ec3\u3002<\/p>\n
\u5982\u679c\u5728\u5d29\u6e83\u540e\u91cd\u65b0\u521b\u5efa\u4e00\u4e2a\u65b0\u7684\u8c03\u5ea6\u5668\uff0c\u4f60\u7684\u5b66\u4e60\u7387\u8ba1\u5212\u5c31\u4f1a\u4ece\u5934\u5f00\u59cb\u3002\u8fd9\u4e2a\u7a81\u589e\u53ef\u80fd\u8ba9\u6a21\u578b\u4e0d\u7a33\u5b9a\uff0c\u5e76\u62b5\u6d88\u6570\u5c0f\u65f6\u7684\u8bad\u7ec3\u6210\u679c\u3002\u6b63\u786e\u505a\u6cd5\u59cb\u7ec8\u662f\u6062\u590d\u5df2\u4fdd\u5b58\u7684\u8c03\u5ea6\u5668\u72b6\u6001\u3002<\/p><\/blockquote>\n
\u6b63\u786e\u65b9\u5f0f\u662f\u5148\u521b\u5efa\u8c03\u5ea6\u5668\uff0c\u518d\u52a0\u8f7d\u5176\u4fdd\u5b58\u7684\u72b6\u6001\u3002scheduler.load_state_dict()<\/code> \u65b9\u6cd5\u4f1a\u6062\u590d\u8c03\u5ea6\u5668\u5185\u90e8\u6240\u6709\u53c2\u6570\uff0c\u8ba9\u5b83\u56de\u5230\u539f\u672c\u7684 step \u8ba1\u6570\u4f4d\u7f6e\u3002\u8fd9\u6837\uff0c\u5b66\u4e60\u7387\u5c31\u80fd\u6cbf\u7740\u65e2\u5b9a\u8f68\u8ff9\u7ee7\u7eed\u524d\u8fdb\uff0c\u4eff\u4f5b\u6545\u969c\u4ece\u672a\u53d1\u751f\u3002\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u5c06\u65e0\u7f1d\u8854\u63a5\uff0c\u4e0d\u9519\u8fc7\u4efb\u4f55\u4e00\u6b65\u3002<\/p>\n
\u59a5\u5584\u6062\u590d\u8c03\u5ea6\u5668\uff0c\u53ef\u4ee5\u5f7b\u5e95\u89c4\u907f\u4e0a\u8ff0\u98ce\u9669\u3002\u4f60\u7684\u8bad\u7ec3\u52a8\u6001\u5f97\u4ee5\u5ef6\u7eed\uff0c\u9a8c\u8bc1\u6307\u6807\u4e5f\u4fdd\u6301\u53ef\u6bd4\u3002\u5e94\u5f53\u628a\u8fd9\u79cd\u505a\u6cd5\u5e94\u7528\u5230\u6bcf\u4e00\u4e2a\u68c0\u67e5\u70b9\u4e2d\u3002\u4fdd\u5b58\u8c03\u5ea6\u5668\u72b6\u6001\u51e0\u4e4e\u4e0d\u9700\u8981\u989d\u5916\u6210\u672c\uff0c\u800c\u5ffd\u7565\u5b83\u7684\u4ee3\u4ef7\u5374\u53ef\u80fd\u975e\u5e38\u9ad8\u6602\u3002\u6b63\u662f\u8fd9\u79cd\u5bf9\u7ec6\u8282\u7684\u91cd\u89c6\uff0c\u624d\u80fd\u628a\u4e00\u6b21\u5d29\u6e83\u4ece\u91cd\u5927\u632b\u6298\uff0c\u53d8\u6210\u77ed\u6682\u6682\u505c\u3002<\/p>\n
\u9a8c\u8bc1\u6062\u590d\u662f\u5426\u6210\u529f<\/strong><\/h2>\n
\u5bf9\u635f\u5931\u548c\u68af\u5ea6\u8fdb\u884c\u5408\u7406\u6027\u68c0\u67e5<\/h3>\n
\u5728\u6062\u590d\u68c0\u67e5\u70b9\u4e4b\u540e\uff0c\u5148\u8fd0\u884c\u5c11\u91cf batch\u3002\u628a\u6b64\u65f6\u7684\u635f\u5931\u503c\u4e0e\u5d29\u6e83\u524d\u65e5\u5fd7\u4e2d\u7684\u6570\u503c\u8fdb\u884c\u6bd4\u8f83\u3002\u635f\u5931\u5e94\u4e0e\u8be5 epoch \u6700\u540e\u8bb0\u5f55\u7684\u503c\u5927\u81f4\u63a5\u8fd1\u3002\u5982\u679c\u5339\u914d\u826f\u597d\uff0c\u5c31\u8bf4\u660e\u8bad\u7ec3\u72b6\u6001\u6062\u590d\u6b63\u786e\uff1b\u5982\u679c\u51fa\u73b0\u660e\u663e\u8df3\u53d8\uff0c\u5219\u610f\u5473\u7740\u68c0\u67e5\u70b9\u52a0\u8f7d\u6d41\u7a0b\u5b58\u5728\u95ee\u9898\u3002<\/p>\n
\u4f60\u8fd8\u9700\u8981\u68c0\u67e5\u68af\u5ea6\u8303\u6570\u3002\u8fd9\u4e9b\u6570\u503c\u8861\u91cf\u7684\u662f\u6743\u91cd\u66f4\u65b0\u7684\u5e45\u5ea6\u3002\u628a\u5b83\u4eec\u4e0e\u5d29\u6e83\u524d\u7684\u65e5\u5fd7\u8fdb\u884c\u5bf9\u7167\uff0c\u82e5\u7ed3\u679c\u76f8\u8fd1\uff0c\u901a\u5e38\u8868\u793a\u6062\u590d\u6210\u529f\uff1b\u82e5\u5dee\u5f02\u660e\u663e\uff0c\u5219\u8bf4\u660e\u67d0\u5904\u51fa\u4e86\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u8c03\u5ea6\u5668\u72b6\u6001\u53ef\u80fd\u6ca1\u6709\u6b63\u786e\u6062\u590d\uff0c\u6216\u8005\u968f\u673a\u6570\u751f\u6210\u5668\u4e0e\u539f\u59cb\u73af\u5883\u4e0d\u4e00\u81f4\u3002\u8fd9\u4e9b\u7ec6\u5fae\u7684\u4e0d\u5339\u914d\uff0c\u90fd\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u7ed3\u679c\u4e0d\u7a33\u5b9a\u3002<\/p>\n
\u5982\u679c\u635f\u5931\u503c\u6216\u68af\u5ea6\u6570\u503c\u51fa\u73b0\u5927\u5e45\u8df3\u53d8\uff0c\u5c31\u8bf4\u660e\u6062\u590d\u8fc7\u7a0b\u51fa\u73b0\u4e86\u9519\u8bef\u3002\u68c0\u67e5\u70b9\u6587\u4ef6\u53ef\u80fd\u5df2\u635f\u574f\uff0c\u4e5f\u53ef\u80fd\u5185\u5bb9\u4e0d\u5b8c\u6574\u3002\u8fd9\u79cd\u5feb\u901f\u68c0\u67e5\u80fd\u5e2e\u52a9\u4f60\u5c3d\u65e9\u53d1\u73b0\u95ee\u9898\uff0c\u907f\u514d\u5728\u9519\u8bef\u72b6\u6001\u4e0b\u7ee7\u7eed\u8bad\u7ec3\u6570\u5c0f\u65f6\uff0c\u4ece\u800c\u8282\u7701\u5b9d\u8d35\u7b97\u529b\u3002<\/p>\n
\u5efa\u8bae\u91c7\u7528\u5bf9\u7167\u5f0f\u9a8c\u8bc1\u65b9\u6cd5\u3002\u63d0\u524d\u8bb0\u5f55\u5d29\u6e83\u524d\u67d0\u4e00\u6b65\u7684\u635f\u5931\u503c\u548c\u68af\u5ea6\u8303\u6570\uff0c\u6062\u590d\u540e\u4f7f\u7528\u76f8\u540c\u6570\u91cf\u7684 batch \u91cd\u65b0\u8dd1\u4e00\u904d\uff0c\u518d\u8bb0\u5f55\u65b0\u7684\u7ed3\u679c\uff0c\u5e76\u8fdb\u884c\u9010\u9879\u5bf9\u6bd4\u3002\u4e24\u7ec4\u6570\u503c\u5e94\u5f53\u9ad8\u5ea6\u63a5\u8fd1\u3002\u6574\u4e2a\u8fc7\u7a0b\u53ea\u9700\u8981\u51e0\u5206\u949f\uff0c\u5374\u80fd\u5728\u540e\u9762\u8282\u7701\u5927\u91cf\u65f6\u95f4\u3002<\/p>\n
\u76d1\u63a7\u6062\u590d\u521d\u671f\u662f\u5426\u51fa\u73b0\u53d1\u6563<\/h3>\n
\u6062\u590d\u540e\u7684\u6700\u521d\u51e0\u4e2a step\uff0c\u5f80\u5f80\u6700\u80fd\u8bf4\u660e\u95ee\u9898\u3002\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u8bad\u7ec3\u5e94\u5f53\u548c\u4e2d\u65ad\u524d\u5b8c\u5168\u4e00\u81f4\u5730\u7ee7\u7eed\u8fdb\u884c\uff0c\u635f\u5931\u66f2\u7ebf\u4fdd\u6301\u539f\u6709\u4e0b\u964d\u8d8b\u52bf\uff0c\u68af\u5ea6\u6a21\u5f0f\u4e5f\u5e94\u7ef4\u6301\u7a33\u5b9a\u3002<\/p>\n
\u4f60\u9700\u8981\u7279\u522b\u7559\u610f\u4e00\u4e9b\u8b66\u544a\u4fe1\u53f7\u3002\u6bd4\u5982\u68af\u5ea6\u7206\u70b8\uff0c\u4f1a\u8868\u73b0\u4e3a\u53c2\u6570\u66f4\u65b0\u7a81\u7136\u5f02\u5e38\u589e\u5927\uff1bNaN \u503c\u5219\u610f\u5473\u7740\u51fa\u73b0\u4e86\u6570\u503c\u4e0d\u7a33\u5b9a\u3002\u8fd9\u4e24\u79cd\u60c5\u51b5\u901a\u5e38\u90fd\u6e90\u4e8e\u6062\u590d\u9519\u8bef\u3002\u5982\u679c\u4e0d\u53ca\u65f6\u5904\u7406\uff0c\u8bad\u7ec3\u4e0d\u4f1a\u81ea\u884c\u56de\u5230\u6b63\u5e38\u72b6\u6001\uff0c\u53cd\u800c\u4f1a\u7ee7\u7eed\u6d6a\u8d39\u8d44\u6e90\u3002\u56e0\u6b64\uff0c\u5fc5\u987b\u5c3d\u65e9\u53d1\u73b0\u3002<\/p>\n
\u5982\u679c\u89c2\u5bdf\u5230\u53d1\u6563\u73b0\u8c61\uff0c\u5e94\u7acb\u5373\u505c\u6b62\u4efb\u52a1\uff0c\u4e0d\u8981\u62b1\u7740\u4fa5\u5e78\u5fc3\u7406\u7ee7\u7eed\u8fd0\u884c\u3002\u8bf7\u68c0\u67e5\u68c0\u67e5\u70b9\u6587\u4ef6\u662f\u5426\u5b8c\u6574\uff0c\u786e\u8ba4\u5404\u4e2a\u7ec4\u4ef6\u662f\u5426\u90fd\u5df2\u6b63\u786e\u52a0\u8f7d\uff0c\u5e76\u6838\u5bf9 RNG \u72b6\u6001\u662f\u5426\u4e0e\u539f\u59cb\u73af\u5883\u4e00\u81f4\u3002<\/p>\n
\u4e00\u6b21\u5e72\u51c0\u3001\u6b63\u786e\u7684\u6062\u590d\uff0c\u4e0d\u4f1a\u51fa\u73b0\u53d1\u6563\u3002\u635f\u5931\u66f2\u7ebf\u4f1a\u5728\u5404\u4e2a epoch \u4e4b\u95f4\u5e73\u6ed1\u5ef6\u7eed\uff0c\u68af\u5ea6\u8303\u6570\u4fdd\u6301\u7a33\u5b9a\u53ef\u9884\u6d4b\uff0c\u5b66\u4e60\u8fc7\u7a0b\u7ee7\u7eed\u6b63\u5e38\u63a8\u8fdb\u3002\u8fd9\u610f\u5473\u7740\u4f60\u5df2\u7ecf\u6210\u529f\u4ece\u6545\u969c\u4e2d\u6062\u590d\uff0c\u4efb\u52a1\u80fd\u591f\u6309\u8ba1\u5212\u7ee7\u7eed\u8fd0\u884c\uff0c\u505c\u673a\u9636\u6bb5\u6b63\u5f0f\u7ed3\u675f\uff0c\u8bad\u7ec3\u8fdb\u5ea6\u4e5f\u5f97\u4ee5\u5ef6\u7eed\u3002<\/p>\n
\u81f3\u6b64\uff0c\u6574\u4e2a\u9a8c\u8bc1\u6d41\u7a0b\u5b8c\u6210\u3002\u6b64\u540e\uff0c\u4f60\u5c31\u53ef\u4ee5\u5bf9\u6a21\u578b\u540e\u7eed\u7ed3\u679c\u4fdd\u6301\u8db3\u591f\u4fe1\u5fc3\u3002<\/p>\n
\n
<\/div>\n<\/div>\n
\u73b0\u5728\uff0c\u4f60\u5df2\u7ecf\u7406\u89e3\u4e86\u5b8c\u6574\u7684\u6062\u590d\u6d41\u7a0b\uff1a\u4fdd\u5b58\u5168\u90e8\u72b6\u6001\u2014\u2014\u6a21\u578b\u3001\u4f18\u5316\u5668\u3001\u8c03\u5ea6\u5668\u548c RNG\uff1b\u5b9e\u65bd\u7a33\u5065\u4e14\u95f4\u9694\u5408\u7406\u7684\u68c0\u67e5\u70b9\u673a\u5236\uff1b\u7cbe\u786e\u6062\u590d\u8bad\u7ec3\u73af\u5883\uff1b\u6b63\u786e\u5ef6\u7eed\u5b66\u4e60\u7387\u8ba1\u5212\u3002\u8fd9\u4e9b\u6b65\u9aa4\uff0c\u80fd\u591f\u628a\u670d\u52a1\u5668\u6545\u969c\u4ece\u707e\u96be\u6027\u4e8b\u4ef6\uff0c\u53d8\u6210\u53ef\u63a7\u7684\u77ed\u6682\u4e2d\u65ad\u3002<\/p>\n
\u53ea\u8981\u51c6\u5907\u5145\u5206\uff0c\u4f60\u7684\u8bad\u7ec3\u4efb\u52a1\u5c31\u80fd\u5728\u5d29\u6e83\u540e\u7ee7\u7eed\u5b58\u6d3b\u3002\u524d\u671f\u5728\u68c0\u67e5\u70b9\u673a\u5236\u4e0a\u7684\u6295\u5165\uff0c\u4f1a\u5728\u672a\u6765\u8282\u7701\u65e0\u6570\u5c0f\u65f6\u3002\u6062\u590d\u65f6\u95f4\u5c06\u4ece\u6570\u5929\u7f29\u77ed\u5230\u51e0\u5206\u949f\u3002\u4f60\u7684\u8bad\u7ec3\u4f1a\u4ece\u4e2d\u65ad\u5904\u51c6\u786e\u63a5\u7eed\uff0c\u4e0d\u6d6a\u8d39\u7b97\u529b\uff0c\u4e5f\u4e0d\u4e22\u5931\u8fdb\u5ea6\u3002<\/p>\n
\u4f60\u4fdd\u5b58\u4e0b\u7684\u6bcf\u4e00\u4e2a epoch\uff0c\u90fd\u662f\u5bf9\u505c\u673a\u98ce\u9669\u7684\u4fdd\u9669\uff1b\u4f60\u505a\u51fa\u7684\u6bcf\u4e00\u6b21\u5b66\u4e60\u7387\u4fee\u6b63\uff0c\u90fd\u662f\u5bf9\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u4fdd\u969c\uff1b\u4f60\u5199\u4e0b\u7684\u6bcf\u4e00\u4e2a\u68c0\u67e5\u70b9\uff0c\u90fd\u662f\u5bf9\u6a21\u578b\u8f68\u8ff9\u7684\u5b88\u62a4\u3002<\/p>\n
\u6709\u4e86\u8fd9\u4efd\u6307\u5357\uff0c\u4f60\u5c31\u80fd\u66f4\u4ece\u5bb9\u5730\u9762\u5bf9\u4e0b\u4e00\u6b21\u670d\u52a1\u5668\u6545\u969c\uff0c\u56e0\u4e3a\u4f60\u77e5\u9053\u8bad\u7ec3\u8fdb\u5ea6\u662f\u5b89\u5168\u7684\uff0c\u63a5\u4e0b\u6765\u7684\u9053\u8def\u4e5f\u662f\u6e05\u6670\u7684\u3002\u8bf7\u5728\u4e0b\u4e00\u6b21\u5d29\u6e83\u53d1\u751f\u4e4b\u524d\uff0c\u5c31\u628a\u8fd9\u4e9b\u5b9e\u8df5\u843d\u5b9e\u5230\u4f4d\uff0c\u800c\u4e0d\u662f\u4e8b\u540e\u8865\u6551\u3002<\/p>\n
\u5e38\u89c1\u95ee\u9898<\/strong><\/h2>\n
\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5e94\u591a\u4e45\u4fdd\u5b58\u4e00\u6b21\u68c0\u67e5\u70b9\uff1f<\/h3>\n
\u4f60\u5e94\u6839\u636e\u9884\u671f\u5d29\u6e83\u7387\u6765\u8bbe\u7f6e\u68c0\u67e5\u70b9\u95f4\u9694\u3002\u6bd4\u5982\u4e00\u4e2a\u4efb\u52a1\u5e73\u5747\u6bcf 150 \u5206\u949f\u4e2d\u65ad\u4e00\u6b21\uff0c\u90a3\u4e48\u6bcf 15 \u5206\u949f\u4fdd\u5b58\u4e00\u6b21\u4f1a\u6bd4\u8f83\u5408\u7406\u3002\u8fd9\u79cd\u505a\u6cd5\u65e2\u80fd\u51cf\u5c11\u91cd\u7b97\u635f\u5931\uff0c\u4e5f\u4e0d\u4f1a\u5e26\u6765\u8fc7\u591a\u989d\u5916\u5f00\u9500\u3002\u6700\u7ec8\u53ef\u884c\u7684\u95f4\u9694\uff0c\u8fd8\u53d6\u51b3\u4e8e\u4f60\u7684\u5b58\u50a8\u5199\u5165\u901f\u5ea6\u3002<\/p>\n
\u5982\u679c\u6211\u53ea\u4fdd\u5b58\u6a21\u578b\u6743\u91cd\uff0c\u4f1a\u53d1\u751f\u4ec0\u4e48\uff1f<\/h3>\n
\u4f60\u4f1a\u4e22\u5931\u5173\u952e\u7684\u4f18\u5316\u5668\u72b6\u6001\uff0c\u5305\u62ec Adam \u7684\u52a8\u91cf\u503c\uff1b\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u4f1a\u91cd\u7f6e\u5230\u521d\u59cb\u9636\u6bb5\uff1b\u968f\u673a\u6570\u751f\u6210\u5668\u4f1a\u751f\u6210\u4e0d\u540c\u7684\u6570\u636e\u6253\u4e71\u987a\u5e8f\u3002\u8fd9\u4e9b\u7f3a\u5931\u90fd\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u7ed3\u679c\u4e0d\u4e00\u81f4\u3002\u56e0\u6b64\uff0c\u82e5\u60f3\u5b9e\u73b0\u771f\u6b63\u7684\u6062\u590d\uff0c\u5fc5\u987b\u59cb\u7ec8\u4fdd\u5b58\u5b8c\u6574\u7684\u72b6\u6001\u5b57\u5178\u3002<\/p>\n
\u6211\u53ef\u4ee5\u5728\u4e0d\u540c\u6570\u91cf\u7684 GPU \u4e0a\u6062\u590d\u8bad\u7ec3\u5417\uff1f<\/h3>\n
\u53ef\u4ee5\uff0c\u4f60\u53ef\u4ee5\u5728\u4e0d\u540c GPU \u6570\u91cf\u7684\u914d\u7f6e\u4e0a\u6062\u590d\u8bad\u7ec3\u3002\u68c0\u67e5\u70b9\u4e2d\u4fdd\u5b58\u7684\u6a21\u578b\u53c2\u6570\u548c\u4f18\u5316\u5668\u72b6\u6001\u53ef\u4ee5\u8de8\u914d\u7f6e\u8fc1\u79fb\u3002\u4e0d\u8fc7\uff0c\u4f60\u9700\u8981\u76f8\u5e94\u8c03\u6574 batch size \u548c\u5b66\u4e60\u7387\u3002\u6570\u636e\u52a0\u8f7d\u5668\u7684\u5206\u53d1\u65b9\u5f0f\u53ef\u80fd\u4f1a\u6539\u53d8\uff0c\u4f46\u4fdd\u5b58\u7684 epoch \u548c\u6a21\u578b\u72b6\u6001\u4f9d\u7136\u6709\u6548\u3002<\/p>\n
\u6211\u600e\u4e48\u5224\u65ad\u6062\u590d\u662f\u5426\u6210\u529f\uff1f<\/h3>\n
\u5148\u8fd0\u884c\u51e0\u4e2a batch\uff0c\u5e76\u628a\u635f\u5931\u503c\u4e0e\u5d29\u6e83\u524d\u7684\u65e5\u5fd7\u8fdb\u884c\u6bd4\u8f83\uff1b\u68af\u5ea6\u8303\u6570\u4e5f\u5e94\u5f53\u5c3d\u91cf\u63a5\u8fd1\u3002\u540c\u65f6\u89c2\u5bdf\u6062\u590d\u521d\u671f\u662f\u5426\u51fa\u73b0 NaN \u503c\u6216\u68af\u5ea6\u7206\u70b8\u3002\u4efb\u4f55\u663e\u8457\u504f\u5dee\uff0c\u90fd\u8bf4\u660e\u6062\u590d\u6d41\u7a0b\u5b58\u5728\u9519\u8bef\uff0c\u5e94\u7acb\u523b\u505c\u6b62\u5e76\u6392\u67e5\uff0c\u907f\u514d\u7ee7\u7eed\u6d6a\u8d39\u7b97\u529b\u3002<\/p>\n
\u68c0\u67e5\u70b9\u4f1a\u62d6\u6162\u8bad\u7ec3\u5417\uff1f<\/h3>\n
\u68c0\u67e5\u70b9\u673a\u5236\u786e\u5b9e\u4f1a\u5e26\u6765\u4e00\u5b9a\u5f00\u9500\uff0c\u4f46\u73b0\u4ee3\u7cfb\u7edf\u5df2\u7ecf\u80fd\u628a\u8fd9\u90e8\u5206\u6210\u672c\u538b\u5f97\u5f88\u4f4e\u3002\u4f20\u7edf\u5b58\u50a8\u65b9\u5f0f\u901a\u5e38\u4f1a\u5360\u7528 5%\u201310% \u7684\u8bad\u7ec3\u65f6\u95f4\uff0c\u800c\u9ad8\u901f NVMe \u5206\u5c42\u5b58\u50a8\u53ef\u5c06\u5176\u964d\u5230 1% \u4ee5\u4e0b\u3002\u5f02\u6b65\u68c0\u67e5\u70b9\u8fd8\u80fd\u628a\u6570\u636e\u5148\u5378\u8f7d\u5230\u4e3b\u673a\u5185\u5b58\uff0c\u8ba9 GPU \u5728\u4fdd\u5b58\u8fc7\u7a0b\u4e2d\u7ee7\u7eed\u5de5\u4f5c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4f60\u7684 GPT-6 \u8bad\u7ec3\u4efb\u52a1\u5df2\u7ecf\u5728 AI \u8bad\u7ec3\u96c6\u7fa4\u4e0a\u8dd1\u4e86 10 \u5929\uff0c\u5374\u7a81\u7136\u5d29\u6e83\u4e86\u3002\u5c4f\u5e55\u7184\u706d\uff0c\u5fc3\u8df3\u52a0\u901f\u3002\u4e4b\u524d\u7684\u8fdb\u5ea6\u96be\u9053\u5168\u6ca1\u4e86\u5417\uff1f\u8981\u60f3\u6062\u590d\u8bad\u7ec3\u8fdb\u5ea6\uff0c\u5fc5\u987b\u63d0\u524d\u89c4\u5212\u3002\u6ca1\u4eba\u613f\u610f\u773c\u7741\u7741\u770b\u7740\u51e0\u5468\u7684\u7b97\u529b\u6295\u5165\u4ed8\u8bf8\u4e1c\u6d41\uff0c\u4e5f\u6ca1\u4eba\u613f\u610f\u9762\u5bf9\u6f2b\u957f\u505c\u673a\u6240\u5e26\u6765\u7684\u538b\u529b\u3002 \u8fd9\u4efd\u5b9e\u6218\u6307\u5357\u5c06\u63d0\u4f9b\u5207\u5b9e\u53ef\u884c\u7684\u6b65\u9aa4\u3002\u9996\u5148\uff0c\u5f04\u6e05\u695a\u54ea\u4e9b\u72b6\u6001\u4e22\u5931\u4e86\u3002\u7136\u540e\uff0c\u4fdd\u5b58\u5e76\u6062\u590d\u8fd9\u4e9b\u72b6\u6001\u4ee5\u7ee7\u7eed\u8bad\u7ec3\u3002\u63a5\u7740\uff0c\u8c03\u6574\u5b66\u4e60\u7387\u3002\u6700\u540e\uff0c\u9a8c\u8bc1\u6062\u590d\u662f\u5426\u6210\u529f\u3002 \u7406\u89e3\u88ab\u4e2d\u65ad\u7684\u8bad\u7ec3\u4f1a\u8bdd AI \u8bad\u7ec3\u96c6\u7fa4\u6545\u969c\u7684\u5e38\u89c1\u539f\u56e0 \u8bad\u7ec3\u4efb\u52a1\u4e2d\u65ad\u7684\u539f\u56e0\u6709\u5f88\u591a\uff0c\u5176\u4e2d\u786c\u4ef6\u6545\u969c\u6700\u4e3a\u5e38\u89c1\u3002GPU \u53ef\u80fd\u56e0\u8fc7\u70ed\u800c\u81ea\u52a8\u5173\u673a\uff0c\u5185\u5b58\u6a21\u5757\u53ef\u80fd\u51fa\u73b0\u9519\u8bef\uff0c\u7535\u6e90\u4e5f\u53ef\u80fd\u5728\u6beb\u65e0\u9884\u8b66\u7684\u60c5\u51b5\u4e0b\u5931\u6548\u3002\u8f6f\u4ef6\u7f3a\u9677\u540c\u6837\u4f1a\u5bfc\u81f4\u5d29\u6e83\u3002\u4e00\u4e2a\u7ec6\u5fae\u7684\u4ee3\u7801\u9519\u8bef\uff0c\u53ef\u80fd\u5728\u8fd0\u884c\u6570\u5929\u540e\u624d\u7834\u574f\u8bad\u7ec3\u5faa\u73af\u3002\u7f51\u7edc\u95ee\u9898\u4e5f\u53ef\u80fd\u8ba9\u8282\u70b9\u4e0e\u96c6\u7fa4\u5931\u53bb\u8fde\u63a5\u3002\u65e0\u8bba\u662f\u54ea\u79cd\u6545\u969c\uff0c\u7ed3\u679c\u90fd\u4e00\u6837\uff1a\u4f60\u7684\u8bad\u7ec3\u4efb\u52a1\u4f1a\u5728\u4e2d\u9014\u505c\u6b62\u3002 \u5176\u5f71\u54cd\u8fdc\u4e0d\u6b62\u4efb\u52a1\u505c\u6b62\u672c\u8eab\u3002\u4f60\u7684\u96c6\u7fa4\u4f1a\u5904\u4e8e\u7a7a\u95f2\u72b6\u6001\uff0c\u5176\u4ed6\u4efb\u52a1\u5728\u961f\u5217\u4e2d\u7b49\u5f85\uff0c\u56e2\u961f\u7684\u6709\u6548\u5de5\u4f5c\u65f6\u95f4\u88ab\u6d6a\u8d39\u3002\u8fd9\u6837\u7684\u505c\u673a\u635f\u5931\u4e0d\u53ea\u662f\u7b97\u529b\u6210\u672c\uff0c\u66f4\u4f1a\u6253\u4e71\u7814\u7a76\u8282\u594f\uff0c\u62d6\u6162\u9879\u76ee\u8fdb\u5ea6\u3002\u7406\u89e3\u8fd9\u4e9b\u539f\u56e0\uff0c\u6709\u52a9\u4e8e\u4f60\u4e3a\u4e0d\u53ef\u907f\u514d\u7684\u4e2d\u65ad\u505a\u597d\u51c6\u5907\u3002\u4f60\u65e0\u6cd5\u963b\u6b62\u6240\u6709\u6545\u969c\uff0c\u4f46\u4f60\u53ef\u4ee5\u638c\u63a7\u81ea\u5df1\u7684\u5e94\u5bf9\u65b9\u5f0f\u3002 \u8bad\u7ec3\u505c\u6b62\u65f6\u4f1a\u4e22\u5931\u54ea\u4e9b\u72b6\u6001\uff1f \u5f88\u591a\u5de5\u7a0b\u5e08\u8bef\u4ee5\u4e3a\u53ea\u9700\u8981\u4fdd\u5b58\u6a21\u578b\u6743\u91cd\uff0c\u8fd9\u79cd\u60f3\u6cd5\u5f80\u5f80\u4f1a\u5bfc\u81f4\u6062\u590d\u5931\u8d25\u3002\u771f\u6b63\u7684\u7eed\u8bad\uff0c\u8981\u6c42\u4f60\u6062\u590d\u5b8c\u6574\u7684\u8bad\u7ec3\u72b6\u6001\u3002\u6a21\u578b\u53c2\u6570\u53ea\u662f\u5176\u4e2d\u4e00\u90e8\u5206\uff0c\u4f18\u5316\u5668\u72b6\u6001\u540c\u6837\u81f3\u5173\u91cd\u8981\u3002\u4ee5 Adam \u4f18\u5316\u5668\u4e3a\u4f8b\uff0c\u5b83\u4fdd\u5b58\u4e86\u6bcf\u4e2a\u53c2\u6570\u5bf9\u5e94\u7684\u52a8\u91cf\u548c\u65b9\u5dee\u4f30\u8ba1\u503c\u3002\u6ca1\u6709\u8fd9\u4e9b\u503c\uff0c\u4f18\u5316\u5668\u7684\u884c\u4e3a\u5c31\u4f1a\u50cf\u521a\u5f00\u59cb\u8bad\u7ec3\u4e00\u6837\u3002\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u4e5f\u4fdd\u5b58\u4e86\u4f4d\u7f6e\u4fe1\u606f\uff0c\u5b83\u77e5\u9053\u5f53\u524d\u8bad\u7ec3\u8fdb\u884c\u5230\u8870\u51cf\u8ba1\u5212\u7684\u54ea\u4e00\u6b65\u3002 \u968f\u673a\u6570\u751f\u6210\u5668\uff08RNG\uff09\u72b6\u6001\u540c\u6837\u91cd\u8981\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6570\u636e\u6253\u4e71\u548c\u6570\u636e\u589e\u5f3a\u90fd\u4f9d\u8d56\u968f\u673a\u6027\u3002\u6bcf\u4e2a epoch \u7684\u6570\u636e\u987a\u5e8f\u90fd\u4f1a\u4e0d\u540c\uff0c\u56fe\u50cf\u53d8\u6362\u4f1a\u968f\u673a\u88c1\u526a\u548c\u7ffb\u8f6c\u3002\u8fd9\u4e9b\u64cd\u4f5c\u90fd\u7531 RNG \u72b6\u6001\u51b3\u5b9a\u3002\u5982\u679c\u4e0d\u6062\u590d\u5b83\uff0c\u6570\u636e\u7ba1\u9053\u5c31\u4f1a\u751f\u6210\u4e0d\u540c\u7684\u6837\u672c\u5e8f\u5217\u3002\u8fd9\u79cd\u4e0d\u4e00\u81f4\u4f1a\u5e26\u6765\u7ec6\u5fae\u4f46\u771f\u5b9e\u7684\u8bad\u7ec3\u5dee\u5f02\uff0c\u6a21\u578b\u7684\u6536\u655b\u8def\u5f84\u53ef\u80fd\u56e0\u6b64\u6539\u53d8\uff0c\u9a8c\u8bc1\u7ed3\u679c\u4e5f\u53ef\u80fd\u4e0e\u5d29\u6e83\u524d\u7684\u65e5\u5fd7\u4e0d\u518d\u4e00\u81f4\u3002 \u4e00\u4e2a\u5b8c\u6574\u7684\u68c0\u67e5\u70b9\u5e94\u5f53\u540c\u65f6\u4fdd\u5b58\u8fd9\u4e9b\u7ec4\u6210\u90e8\u5206\uff1a\u6a21\u578b\u3001\u4f18\u5316\u5668\u3001\u8c03\u5ea6\u5668\u4ee5\u53ca RNG \u72b6\u6001\u3002\u91c7\u7528\u8fd9\u79cd\u5b8c\u6574\u4fdd\u5b58\u65b9\u5f0f\uff0c\u6062\u590d\u8fc7\u7a0b\u5c31\u4f1a\u7b80\u5355\u5f97\u591a\u3002\u4f60\u53ef\u4ee5\u628a\u6240\u6709\u5185\u5bb9\u6062\u590d\u5230\u6545\u969c\u524d\u7684\u7cbe\u786e\u72b6\u6001\uff0c\u8ba9\u8bad\u7ec3\u4eff\u4f5b\u4ece\u672a\u4e2d\u65ad\u3002\u8fd9\u6837\u7684\u51c6\u5907\uff0c\u80fd\u591f\u628a\u539f\u672c\u53ef\u80fd\u662f\u4e00\u573a\u707e\u96be\u7684\u6545\u969c\uff0c\u53d8\u6210\u4e00\u6b21\u5c0f\u5c0f\u7684\u4e0d\u4fbf\u3002\u4f60\u4e3a\u5b8c\u5584\u68c0\u67e5\u70b9\u673a\u5236\u6295\u5165\u7684\u65f6\u95f4\uff0c\u4f1a\u5728\u4eca\u540e\u7684\u6bcf\u4e00\u6b21\u6545\u969c\u4e2d\u83b7\u5f97\u56de\u62a5\u3002 \u901a\u8fc7\u68c0\u67e5\u70b9\u6062\u590d\u8bad\u7ec3\u8fdb\u5ea6 \u4fdd\u5b58\u5b8c\u6574\u7684\u6a21\u578b\u4e0e\u4f18\u5316\u5668\u72b6\u6001 \u68c0\u67e5\u70b9\u5fc5\u987b\u6355\u83b7\u8bad\u7ec3\u72b6\u6001\u4e2d\u7684\u6bcf\u4e00\u4e2a\u5173\u952e\u90e8\u5206\u3002\u4f60\u4e0d\u80fd\u53ea\u4fdd\u5b58\u6a21\u578b\u6743\u91cd\u3002\u4f18\u5316\u5668\u4fdd\u5b58\u4e86\u6bcf\u4e2a\u53c2\u6570\u7684\u52a8\u91cf\u503c\uff0c\u8c03\u5ea6\u5668\u8bb0\u5f55\u4e86\u81ea\u5df1\u5728\u5b66\u4e60\u7387\u8870\u51cf\u66f2\u7ebf\u4e2d\u7684\u4f4d\u7f6e\uff0c\u968f\u673a\u6570\u751f\u6210\u5668\u51b3\u5b9a\u4e86\u6570\u636e\u6253\u4e71\u7684\u6a21\u5f0f\u3002\u6bcf\u4e00\u4e2a\u7ec4\u4ef6\uff0c\u90fd\u662f\u5e73\u6ed1\u6062\u590d\u8bad\u7ec3\u4e0d\u53ef\u6216\u7f3a\u7684\u4e00\u73af\u3002 \u4f60\u53ef\u4ee5\u901a\u8fc7\u4e00\u6761\u547d\u4ee4\u521b\u5efa\u4e00\u4e2a\u5b8c\u6574\u7684\u68c0\u67e5\u70b9\u3002\u5b57\u5178\u7ed3\u6784\u80fd\u8ba9\u6240\u6709\u5185\u5bb9\u4fdd\u6301\u6e05\u6670\u6709\u5e8f\u3002\u4fdd\u5b58\u5f53\u524d epoch \u7f16\u53f7\uff0c\u4ee5\u4fbf\u77e5\u9053\u4ece\u54ea\u91cc\u91cd\u65b0\u5f00\u59cb\uff1b\u4fdd\u5b58\u6a21\u578b\u7684\u72b6\u6001\u5b57\u5178\uff0c\u4ee5\u8fd8\u539f\u7f51\u7edc\u7ed3\u6784\u4e2d\u7684\u53c2\u6570\uff1b\u4fdd\u5b58\u4f18\u5316\u5668\u7684\u72b6\u6001\u5b57\u5178\uff0c\u4ee5\u4fdd\u7559\u5df2\u7ecf\u7d2f\u79ef\u7684\u68af\u5ea6\u4fe1\u606f\uff1b\u4fdd\u5b58\u8c03\u5ea6\u5668\u72b6\u6001\uff0c\u4ee5\u5ef6\u7eed\u65e2\u5b9a\u7684\u5b66\u4e60\u7387\u8ba1\u5212\uff1b\u8bb0\u5f55 RNG \u72b6\u6001\uff0c\u4ee5\u4fdd\u8bc1\u6570\u636e\u7ba1\u9053\u7684\u53ef\u590d\u73b0\u6027\u3002\u8fd9\u4e2a\u5b8c\u6574\u5feb\u7167\uff0c\u80fd\u5728\u4efb\u4f55\u6545\u969c\u540e\u771f\u6b63\u5b9e\u73b0\u8bad\u7ec3\u8fdb\u5ea6\u7684\u6062\u590d\u3002 torch.save({ ‘epoch’: epoch, ‘model_state_dict’: model.state_dict(), ‘optimizer_state_dict’: optimizer.state_dict(), ‘scheduler_state_dict’: scheduler.state_dict(), ‘rng_state’: torch.get_rng_state() }, ‘checkpoint.pth’) \u968f\u7740\u6a21\u578b\u89c4\u6a21\u589e\u5927\uff0c\u8fd9\u4e2a\u6587\u4ef6\u4f1a\u8fc5\u901f\u81a8\u80c0\u3002\u4e00\u4e2a\u7b80\u5355\u7ecf\u9a8c\u6cd5\u5219\u662f\uff1a\u82e5\u5305\u542b\u4f18\u5316\u5668\u72b6\u6001\uff0c\u6bcf\u4e2a\u53c2\u6570\u5927\u7ea6\u9700\u8981 12 \u5b57\u8282\u3002\u6743\u91cd\u672c\u8eab\u5360 4 \u5b57\u8282\uff0c\u4f18\u5316\u5668\u72b6\u6001\u5360 8 \u5b57\u8282\u3002\u8fd9\u4e2a\u4f30\u7b97\u89c4\u5f8b\u53ef\u4ee5\u5e2e\u52a9\u4f60\u5728\u8bad\u7ec3\u5f00\u59cb\u524d\u5c31\u89c4\u5212\u597d\u5b58\u50a8\u5bb9\u91cf\u3002 \u5b58\u50a8\u68c0\u67e5\u70b9\u4ee5\u5b9e\u73b0\u5feb\u901f\u6062\u590d […]<\/p>\n
Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":34576,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[62],"tags":[13232,13233,13234,13235,13236],"class_list":["post-34579","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-sc","tag-gpt-6-training-recovery","tag-server-crash-recovery","tag-ai-training-checkpoint","tag-learning-rate-recovery","tag-distributed-training-failure-recovery"],"acf":[],"yoast_head":"\n