{"id":2567,"date":"2024-03-25T02:16:38","date_gmt":"2024-03-24T18:16:38","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2567"},"modified":"2024-04-28T20:01:42","modified_gmt":"2024-04-28T12:01:42","slug":"%e4%bb%8e%e5%a4%b4%e5%bc%80%e5%a7%8b%e7%bc%96%e5%86%99-lora-%e4%bb%a3%e7%a0%81","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2567","title":{"rendered":"\u4ece\u5934\u5f00\u59cb\u7f16\u5199 LoRA \u4ee3\u7801"},"content":{"rendered":"\n

\u5728 PyTorch \u4e2d\u5b9e\u73b0 LLMs \u7684\u4f4e\u9636\u9002\u5e94<\/a><\/p>\n\n\n\n

LoRA \u4ee3\u8868\u4f4e\u9636\u9002\u5e94\uff0c\u662f\u4e00\u79cd\u66f4\u6709\u6548\u5730\u5fae\u8c03 LLMs \u7684\u6d41\u884c\u6280\u672f\u3002 LoRA \u4e0d\u8c03\u6574\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u7684\u6240\u6709\u53c2\u6570\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\u4ec5\u66f4\u65b0\u4e00\u5c0f\u90e8\u5206\u4f4e\u79e9\u77e9\u9635.<\/p>\n\n\n\n

\u4ece\u5934\u5f00\u59cb\u7f16\u7801\u6765\u89e3\u91ca LoRA \u7684\u5de5\u4f5c\u539f\u7406\uff0c\u8fd9\u662f\u4e86\u89e3\u7b97\u6cd5\u5e95\u5c42\u7684\u7edd\u4f73\u7ec3\u4e60.<\/p>\n\n\n\n

1. \u4e86\u89e3 LoRA<\/strong><\/h2>\n\n\n\n

\u9884\u8bad\u7ec3\u7684 LLMs \u901a\u5e38\u88ab\u79f0\u4e3a\u57fa\u7840\u6a21\u578b\uff0c\u56e0\u4e3a\u5b83\u4eec\u5728\u5404\u79cd\u4efb\u52a1\u4e2d\u5177\u6709\u591a\u529f\u80fd\u6027\u3002\u7136\u800c\uff0c\u9488\u5bf9\u7279\u5b9a\u6570\u636e\u96c6\u6216\u4efb\u52a1\u8c03\u6574\u9884\u8bad\u7ec3\u7684 LLM \u901a\u5e38\u5f88\u6709\u7528\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5fae\u8c03\u6765\u5b8c\u6210\u3002<\/p>\n\n\n\n

\u5fae\u8c03\u5141\u8bb8\u6a21\u578b\u9002\u5e94\u7279\u5b9a\u9886\u57df\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u6602\u8d35\u7684\u9884\u8bad\u7ec3\uff0c\u4f46\u66f4\u65b0\u6240\u6709\u5c42\u7684\u8ba1\u7b97\u6210\u672c\u4ecd\u7136\u5f88\u9ad8\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u8f83\u5927\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

LoRA \u63d0\u4f9b\u4e86\u6bd4\u5e38\u89c4\u5fae\u8c03\u66f4\u6709\u6548\u7684\u66ff\u4ee3\u65b9\u6848\u3002\u6b63\u5982\u300aLoRA\uff1a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u4f4e\u79e9\u9002\u5e94\u300b\u4e00\u6587\u4e2d\u66f4\u8be6\u7ec6\u8ba8\u8bba\u7684\u90a3\u6837\uff0cLoRA \u4ee5\u4f4e\u79e9\u683c\u5f0f\u8fd1\u4f3c\u8bad\u7ec3\u671f\u95f4\u5c42\u7684\u6743\u91cd\u53d8\u5316 \u0394W<\/em>\u3002<\/p>\n\n\n\n

\u4f8b\u5982\uff0c\u5728\u5e38\u89c4\u5fae\u8c03\u4e2d\uff0c\u6211\u4eec\u5c06\u6743\u91cd\u77e9\u9635 W \u7684\u6743\u91cd\u66f4\u65b0\u8ba1\u7b97\u4e3a \u0394W<\/em>\uff0c\u800c\u5728 LoRA \u4e2d\uff0c\u6211\u4eec\u901a\u8fc7\u4e24\u4e2a\u8f83\u5c0f\u77e9\u9635 AB \u7684\u77e9\u9635\u4e58\u6cd5\u6765\u8fd1\u4f3c \u0394W<\/em>\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002 \uff08\u5982\u679c\u60a8\u719f\u6089 PCA \u6216 SVD\uff0c\u8bf7\u5c06\u5176\u89c6\u4e3a\u5c06 \u0394W<\/em> \u5206\u89e3\u4e3a A \u548c B\u3002\uff09<\/p>\n\n\n\n

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

\u5e38\u89c4\u5fae\u8c03\uff08\u5de6\uff09\u548c LoRA\uff08\u53f3\uff09\u4e2d\u524d\u5411\u4f20\u9012\u8fc7\u7a0b\u4e2d\u6743\u91cd\u66f4\u65b0\u4e4b\u95f4\u7684\u6bd4\u8f83\u3002<\/em><\/p>\n\n\n\n

\u8bf7\u6ce8\u610f\uff0c\u4e0a\u56fe\u4e2d\u7684 r \u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u6211\u4eec\u53ef\u4ee5\u7528\u5b83\u6765\u6307\u5b9a\u7528\u4e8e\u9002\u5e94\u7684\u4f4e\u79e9\u77e9\u9635\u7684\u79e9\u3002\u8f83\u5c0f\u7684 r \u4f1a\u5bfc\u81f4\u66f4\u7b80\u5355\u7684\u4f4e\u79e9\u77e9\u9635\uff0c\u4ece\u800c\u5bfc\u81f4\u5728\u9002\u5e94\u8fc7\u7a0b\u4e2d\u9700\u8981\u5b66\u4e60\u7684\u53c2\u6570\u66f4\u5c11\u3002\u8fd9\u53ef\u4ee5\u5e26\u6765\u66f4\u5feb\u7684\u8bad\u7ec3\u5e76\u53ef\u80fd\u51cf\u5c11\u8ba1\u7b97\u9700\u6c42\u3002\u7136\u800c\uff0c\u968f\u7740 r \u7684\u51cf\u5c0f\uff0c\u4f4e\u79e9\u77e9\u9635\u6355\u83b7\u7279\u5b9a\u4efb\u52a1\u4fe1\u606f\u7684\u80fd\u529b\u4f1a\u964d\u4f4e\u3002<\/p>\n\n\n\n

\u4e3e\u4e2a\u5177\u4f53\u7684\u4f8b\u5b50\uff0c\u5047\u8bbe\u7ed9\u5b9a\u5c42\u7684\u6743\u91cd\u77e9\u9635W\u7684\u5927\u5c0f\u4e3a5,000×10,000\uff08\u603b\u517150M\u4e2a\u53c2\u6570\uff09\u3002\u5982\u679c\u6211\u4eec\u9009\u62e9\u79e9r=8\uff0c\u6211\u4eec\u521d\u59cb\u5316\u4e24\u4e2a\u77e9\u9635\uff1a5,000×8\u7ef4\u77e9\u9635B\u548c8×10,000\u7ef4\u77e9\u9635A\u3002\u52a0\u5728\u4e00\u8d77\uff0cA\u548cB\u53ea\u670980,000 + 40,000 = 120,000\u4e2a\u53c2\u6570\uff0c\u5373400\u500d\u5c0f\u4e8e\u901a\u8fc7\u0394W\u8fdb\u884c\u5e38\u89c4\u5fae\u8c03\u768450M\u53c2\u6570\u3002<\/p>\n\n\n\n

\u5728\u5b9e\u8df5\u4e2d\uff0c\u5c1d\u8bd5\u4e0d\u540c\u7684 r \u503c\u4ee5\u627e\u5230\u9002\u5f53\u7684\u5e73\u8861\u4ee5\u5728\u65b0\u4efb\u52a1\u4e2d\u5b9e\u73b0\u6240\u9700\u7684\u6027\u80fd\u975e\u5e38\u91cd\u8981\u3002<\/p>\n\n\n\n

2. \u4ece\u5934\u5f00\u59cb\u7f16\u5199 LoRA \u4ee3\u7801<\/strong><\/h2>\n\n\n\n

\u7531\u4e8e\u6982\u5ff5\u89e3\u91ca\u6709\u65f6\u53ef\u80fd\u5f88\u62bd\u8c61\uff0c\u56e0\u6b64\u73b0\u5728\u8ba9\u6211\u4eec\u81ea\u5df1\u5b9e\u73b0 LoRA\uff0c\u4ee5\u66f4\u597d\u5730\u4e86\u89e3\u5b83\u7684\u5de5\u4f5c\u539f\u7406\u3002\u5728\u4ee3\u7801\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u6309\u5982\u4e0b\u65b9\u5f0f\u5b9e\u73b0 LoRA \u5c42\uff1a<\/p>\n\n\n\n

class LoRALayer(torch.nn.Module):\n    def __init__(self, in_dim, out_dim, rank, alpha):\n        super().__init__()\n        std_dev = 1 \/ torch.sqrt(torch.tensor(rank).float())\n        self.A = torch.nn.Parameter(torch.randn(in_dim, rank) * std_dev)\n        self.B = torch.nn.Parameter(torch.zeros(rank, out_dim))\n        self.alpha = alpha\n\n    def forward(self, x):\n        x = self.alpha * (x @ self.A @ self.B)\n        return x<\/pre><\/div>\n\n\n\n
\u5728\u4e0a\u9762\u7684\u4ee3\u7801\u4e2d\uff0c in_dim<\/strong><\/code><\/code> \u662f\u6211\u4eec\u8981\u4f7f\u7528LoRA\u4fee\u6539\u7684\u5c42\u7684\u8f93\u5165\u7ef4\u5ea6\uff0c out_dim<\/strong><\/code><\/code> \u662f\u8be5\u5c42\u5404\u81ea\u7684\u8f93\u51fa\u7ef4\u5ea6\u3002<\/p>\n\n\n\n
\u6211\u4eec\u4e4b\u524d\u8ba8\u8bba\u8fc7\uff0c\u77e9\u9635 A<\/em><\/strong><\/code> \u548c B<\/em><\/strong><\/code> \u7684 rank<\/code> \u662f\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u7528\u4e8e\u63a7\u5236LoRA\u5f15\u5165\u7684\u9644\u52a0\u53c2\u6570\u7684\u590d\u6742\u6027\u548c\u6570\u91cf\u3002<\/p>\n\n\n\n
\u7136\u800c\uff0c\u770b\u770b\u4e0a\u9762\u7684\u4ee3\u7801\uff0c\u6211\u4eec\u6dfb\u52a0\u4e86\u53e6\u4e00\u4e2a\u8d85\u53c2\u6570\uff0c\u5373\u7f29\u653e\u56e0\u5b50 alpha<\/strong><\/code><\/code> \u3002\u8be5\u56e0\u7d20\u51b3\u5b9a\u4e86 LoRA \u5c42\u5bf9\u6a21\u578b\u73b0\u6709\u6743\u91cd\u5f15\u5165\u7684\u53d8\u5316\u5e45\u5ea6\uff1a alpha * (x @ A @ B)<\/code><\/code> \u3002 alpha<\/code> \u503c\u8d8a\u9ad8\uff0c\u610f\u5473\u7740\u5bf9\u6a21\u578b\u884c\u4e3a\u7684\u8c03\u6574\u8d8a\u5927\uff0c\u800c\u503c\u8d8a\u4f4e\uff0c\u53d8\u5316\u8d8a\u7ec6\u5fae\u3002<\/p>\n\n\n\n
\u53e6\u4e00\u4ef6\u9700\u8981\u6ce8\u610f\u7684\u4e8b\u60c5\u662f\uff0c\u6211\u4eec\u4f7f\u7528\u968f\u673a\u5206\u5e03\u4e2d\u7684\u5c0f\u503c\u521d\u59cb\u5316\u4e86 A<\/code><\/code> \u3002\u8fd9\u91cc\uff0c\u8fd9\u4e2a\u5206\u5e03\u7684\u6807\u51c6\u5dee\u662f\u7531\u79e9\u7684\u5e73\u65b9\u6839\u51b3\u5b9a\u7684\uff08\u8fd9\u4e2a\u9009\u62e9\u4fdd\u8bc1\u4e86 A<\/code><\/code> \u4e2d\u7684\u521d\u59cb\u503c\u4e0d\u4f1a\u592a\u5927\u3002\uff09\u4f46\u662f\uff0c\u6211\u4eec\u521d\u59cb\u5316\u4e86 B<\/code><\/code> \u5e26\u96f6\u3002\u8fd9\u91cc\u7684\u57fa\u672c\u539f\u7406\u662f\uff0c\u5728\u8bad\u7ec3\u5f00\u59cb\u65f6\uff0c\u5728\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u66f4\u65b0 A<\/code><\/code> \u548c B<\/code><\/code> \u4e4b\u524d\uff0c LoRALayer<\/code> \u4e0d\u4f1a\u5f71\u54cd\u539f\u59cb\u6743\u91cd\uff0c\u56e0\u4e3a AB=0<\/strong><\/em><\/code> if B=0<\/em><\/strong><\/code> \u3002<\/p>\n\n\n\n
\u8bf7\u6ce8\u610f\uff0cLoRA \u901a\u5e38\u5e94\u7528\u4e8e\u795e\u7ecf\u7f51\u7edc\u7684\u7ebf\u6027\uff08\u524d\u9988\uff09\u5c42\u3002\u4f8b\u5982\uff0c\u5047\u8bbe\u6211\u4eec\u6709\u4e00\u4e2a\u7b80\u5355\u7684 PyTorch \u6a21\u578b\u6216\u5177\u6709\u4e24\u4e2a\u7ebf\u6027\u5c42\u7684\u6a21\u5757\uff08\u4f8b\u5982\uff0c\u8fd9\u53ef\u80fd\u662f\u53d8\u538b\u5668\u5757\u7684\u524d\u9988\u6a21\u5757\uff09\u3002\u5047\u8bbe\u8be5\u6a21\u5757\u7684forward\u65b9\u6cd5\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n
def forward(self, x):\n    x = self.linear_1(x)\n    x = F.relu(x)\n    x = self.linear_2(x)\n    return x<\/pre><\/div>\n\n\n\n
\u5982\u679c\u6211\u4eec\u4f7f\u7528 LoRA\uff0c\u6211\u4eec\u4f1a\u5c06 LoRA \u66f4\u65b0\u6dfb\u52a0\u5230\u8fd9\u4e9b\u7ebf\u6027\u5c42\u8f93\u51fa\u4e2d\uff0c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n
def forward(self, x):\n    x = self.linear_1(x) + self.lora_1(x)\n    x = F.relu(x)\n    x = self.linear_2(x) + self.lora_2(x)\n    return logits<\/pre><\/div>\n\n\n\n
\u5728\u4ee3\u7801\u4e2d\uff0c\u5f53\u901a\u8fc7\u4fee\u6539\u73b0\u6709 PyTorch \u6a21\u578b\u6765\u5b9e\u73b0 LoRA \u65f6\uff0c\u5b9e\u73b0\u7ebf\u6027\u5c42\u7684\u8fd9\u79cd LoRA \u4fee\u6539\u7684\u4e00\u79cd\u7b80\u5355\u65b9\u6cd5\u662f\u5c06\u6bcf\u4e2a\u7ebf\u6027\u5c42\u66ff\u6362\u4e3a LinearWithLoRA<\/code><\/code> \u5c42\uff0c\u8be5\u5c42\u5c06 Linear<\/code><\/code> \u5c42\u4e0e\u6211\u4eec\u4e4b\u524d\u7684 LoRALayer<\/code><\/code> \u5b9e\u73b0\u76f8\u7ed3\u5408\uff1a<\/p>\n\n\n\n
class LinearWithLoRA(torch.nn.Module):\n    def __init__(self, linear, rank, alpha):\n        super().__init__()\n        self.linear = linear\n        self.lora = LoRALayer(\n            linear.in_features, linear.out_features, rank, alpha\n        )\n\n    def forward(self, x):\n        return self.linear(x) + self.lora(x)<\/pre><\/div>\n\n\n\n
\u4e0a\u8ff0\u8fd9\u4e9b\u6982\u5ff5\u603b\u7ed3\u5982\u4e0b\u56fe<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u5728\u5b9e\u8df5\u4e2d\uff0c\u8981\u4f7f\u7528 LoRA \u88c5\u5907\u548c\u5fae\u8c03\u6a21\u578b\uff0c\u6211\u4eec\u6240\u8981\u505a\u7684\u5c31\u662f\u7528\u65b0\u7684 LinearWithLoRA<\/code> \u5c42\u66ff\u6362\u5176\u9884\u8bad\u7ec3\u7684 Linear<\/code> \u5c42\u3002\u6211\u4eec\u5c06\u5728\u4e0b\u9762\u7684\u5b9e\u8df5\u90e8\u5206\u4e2d\u901a\u8fc7\u5c06 LinearWithLoRA<\/code> \u5c42\u5e94\u7528\u4e8e\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u6765\u4e86\u89e3\u5176\u5de5\u4f5c\u539f\u7406\u3002<\/p>\n\n\n\n
3. \u4f7f\u7528 LoRA \u8fdb\u884c\u5fae\u8c03\u2014\u2014\u4e00\u4e2a\u5b9e\u8df5\u793a\u4f8b<\/strong><\/h2>\n\n\n\n
LoRA \u662f\u4e00\u79cd\u53ef\u4ee5\u5e94\u7528\u4e8e\u5404\u79cd\u7c7b\u578b\u795e\u7ecf\u7f51\u7edc\u7684\u65b9\u6cd5\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f GPT \u6216\u56fe\u50cf\u751f\u6210\u6a21\u578b\u7b49\u751f\u6210\u6a21\u578b\u3002\u5bf9\u4e8e\u8fd9\u4e2a\u5b9e\u8df5\u793a\u4f8b\uff0c\u6211\u4eec\u5c06\u8bad\u7ec3\u4e00\u4e2a\u7528\u4e8e\u6587\u672c\u5206\u7c7b\u7684\u5c0f\u578b BERT \u6a21\u578b\uff0c\u56e0\u4e3a\u5206\u7c7b\u51c6\u786e\u6027\u6bd4\u751f\u6210\u7684\u6587\u672c\u66f4\u5bb9\u6613\u8bc4\u4f30\u3002 <\/p>\n\n\n\n
\u7279\u522b\u662f\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 Transformer \u5e93\u4e2d\u7684\u9884\u8bad\u7ec3 DistilBERT\uff08BERT \u7684\u8f83\u5c0f\u7248\u672c\uff09\u6a21\u578b<\/p>\n\n\n\n
from transformers import AutoModelForSequenceClassification\n\nmodel = AutoModelForSequenceClassification.from_pretrained(\n    \"distilbert-base-uncased\", num_labels=2)<\/pre><\/div>\n\n\n\n
\u7531\u4e8e\u6211\u4eec\u53ea\u60f3\u8bad\u7ec3\u65b0\u7684 LoRA \u6743\u91cd\uff0c\u56e0\u6b64\u6211\u4eec\u901a\u8fc7\u5c06\u6240\u6709\u53ef\u8bad\u7ec3\u7684 requires_grad<\/code><\/code> \u8bbe\u7f6e\u4e3a False<\/code><\/code> \u6765\u51bb\u7ed3\u6240\u6709\u6a21\u578b\u53c2\u6570\u53c2\u6570\uff1a<\/p>\n\n\n\n
for param in model.parameters():\n    param.requires_grad = False\nprint(model)<\/pre><\/div>\n\n\n\n
\u63a5\u4e0b\u6765\uff0c\u8ba9\u6211\u4eec\u4f7f\u7528 print(model)<\/code> \u7b80\u8981\u68c0\u67e5\u6a21\u578b\u7684\u7ed3\u6784\uff1a<\/p>\n\n\n\n
DistilBertForSequenceClassification(\n  (distilbert): DistilBertModel(\n    (embeddings): Embeddings(\n      (word_embeddings): Embedding(30522, 768, padding_idx=0)\n      (position_embeddings): Embedding(512, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (transformer): Transformer(\n      (layer): ModuleList(\n        (0-5): 6 x TransformerBlock(\n          (attention): MultiHeadSelfAttention(\n            (dropout): Dropout(p=0.1, inplace=False)\n            (q_lin): Linear(in_features=768, out_features=768, bias=True)\n            (k_lin): Linear(in_features=768, out_features=768, bias=True)\n            (v_lin): Linear(in_features=768, out_features=768, bias=True)\n            (out_lin): Linear(in_features=768, out_features=768, bias=True)\n          )\n          (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n          (ffn): FFN(\n            (dropout): Dropout(p=0.1, inplace=False)\n            (lin1): Linear(in_features=768, out_features=3072, bias=True)\n            (lin2): Linear(in_features=3072, out_features=768, bias=True)\n            (activation): GELUActivation()\n          )\n          (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n        )\n      )\n    )\n  )\n  (pre_classifier): Linear(in_features=768, out_features=768, bias=True)\n  (classifier): Linear(in_features=768, out_features=2, bias=True)\n  (dropout): Dropout(p=0.2, inplace=False)\n)<\/pre><\/div>\n\n\n\n
\u6839\u636e\u4e0b\u9762\u7684\u8f93\u51fa\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u8be5\u6a21\u578b\u7531 6 \u4e2a\u5305\u542b\u7ebf\u6027\u5c42\u7684 Transformer \u5c42\u7ec4\u6210\uff1a<\/p>\n\n\n\n
(0-5): 6 x TransformerBlock(\n          (attention): MultiHeadSelfAttention(\n            ...\n            (q_lin): Linear(in_features=768, out_features=768, bias=True)\n\t\t\t\t\t\t...\n          )\n          ...<\/pre><\/div>\n\n\n\n
\u6b64\u5916\uff0c\u8be5\u6a21\u578b\u8fd8\u6709\u4e24\u4e2a Linear<\/code> \u8f93\u51fa\u5c42\uff1a<\/p>\n\n\n\n
...\n(pre_classifier): Linear(in_features=768, out_features=768, bias=True)\n(classifier): Linear(in_features=768, out_features=2, bias=True)\n<\/pre><\/div>\n\n\n\n
\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u5b9a\u4e49\u4ee5\u4e0b\u5206\u914d\u51fd\u6570\u548c\u5faa\u73af\u6765\u9009\u62e9\u6027\u5730\u4e3a\u8fd9\u4e9b Linear<\/code> \u5c42\u542f\u7528 LoRA\uff1a<\/p>\n\n\n\n
from transformers import AutoModelForSequenceClassification\nimport torch\n\nclass LoRALayer(torch.nn.Module):\n    def __init__(self, in_dim, out_dim, rank, alpha):\n        super().__init__()\n        std_dev = 1 \/ torch.sqrt(torch.tensor(rank).float())\n        self.A = torch.nn.Parameter(torch.randn(in_dim, rank) * std_dev)\n        self.B = torch.nn.Parameter(torch.zeros(rank, out_dim))\n        self.alpha = alpha\n\n    def forward(self, x):\n        x = self.alpha * (x @ self.A @ self.B)\n        return x\n\nclass LinearWithLoRA(torch.nn.Module):\n    def __init__(self, linear, rank, alpha):\n        super().__init__()\n        self.linear = linear\n        self.lora = LoRALayer(\n            linear.in_features, linear.out_features, rank, alpha\n        )\n\n    def forward(self, x):\n        return self.linear(x) + self.lora(x)\n        \nmodel = AutoModelForSequenceClassification.from_pretrained(\n    \"distilbert-base-uncased\", num_labels=2)\n\nfor param in model.parameters():\n    param.requires_grad = False\n\nfrom functools import partial\n\n# default hyperparameter choices\nlora_r = 8\nlora_alpha = 16\nlora_dropout = 0.05\nlora_query = True\nlora_key = False\nlora_value = True\nlora_projection = False\nlora_mlp = False\nlora_head = False\n\nlayers = []\n\nassign_lora = partial(LinearWithLoRA, rank=lora_r, alpha=lora_alpha)\n\nfor layer in model.distilbert.transformer.layer:\n    if lora_query:\n        layer.attention.q_lin = assign_lora(layer.attention.q_lin)\n    if lora_key:\n        layer.attention.k_lin = assign_lora(layer.attention.k_lin)\n    if lora_value:\n        layer.attention.v_lin = assign_lora(layer.attention.v_lin)\n    if lora_projection:\n        layer.attention.out_lin = assign_lora(layer.attention.out_lin)\n    if lora_mlp:\n        layer.ffn.lin1 = assign_lora(layer.ffn.lin1)\n        layer.ffn.lin2 = assign_lora(layer.ffn.lin2)\nif lora_head:\n    model.pre_classifier = assign_lora(model.pre_classifier)\n    model.classifier = assign_lora(model.classifier)\n\nprint(model)\n    <\/pre><\/div>\n\n\n\n
\u6211\u4eec\u73b0\u5728\u53ef\u4ee5\u4f7f\u7528 print(model)<\/code> \u518d\u6b21\u68c0\u67e5\u6a21\u578b\u4ee5\u68c0\u67e5\u5176\u66f4\u65b0\u540e\u7684\u7ed3\u6784\uff1a<\/p>\n\n\n\n
DistilBertForSequenceClassification(\n  (distilbert): DistilBertModel(\n    (embeddings): Embeddings(\n      (word_embeddings): Embedding(30522, 768, padding_idx=0)\n      (position_embeddings): Embedding(512, 768)\n      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n      (dropout): Dropout(p=0.1, inplace=False)\n    )\n    (transformer): Transformer(\n      (layer): ModuleList(\n        (0-5): 6 x TransformerBlock(\n          (attention): MultiHeadSelfAttention(\n            (dropout): Dropout(p=0.1, inplace=False)\n            (q_lin): LinearWithLoRA(\n              (linear): Linear(in_features=768, out_features=768, bias=True)\n              (lora): LoRALayer()\n            )\n            (k_lin): Linear(in_features=768, out_features=768, bias=True)\n            (v_lin): LinearWithLoRA(\n              (linear): Linear(in_features=768, out_features=768, bias=True)\n              (lora): LoRALayer()\n            )\n            (out_lin): Linear(in_features=768, out_features=768, bias=True)\n          )\n          (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n          (ffn): FFN(\n            (dropout): Dropout(p=0.1, inplace=False)\n            (lin1): Linear(in_features=768, out_features=3072, bias=True)\n            (lin2): Linear(in_features=3072, out_features=768, bias=True)\n            (activation): GELUActivation()\n          )\n          (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\n        )\n      )\n    )\n  )\n  (pre_classifier): Linear(in_features=768, out_features=768, bias=True)\n  (classifier): Linear(in_features=768, out_features=2, bias=True)\n  (dropout): Dropout(p=0.2, inplace=False)\n)<\/pre><\/div>\n\n\n\n
\u5982\u4e0a\u6240\u793a\uff0c Linear<\/code> \u5c42\u5df2\u6210\u529f\u66ff\u6362\u4e3a LinearWithLoRA<\/code> \u5c42<\/p>\n\n\n\n
\u5982\u679c\u6211\u4eec\u4f7f\u7528\u4e0a\u9762\u663e\u793a\u7684\u9ed8\u8ba4\u8d85\u53c2\u6570\u9009\u62e9\u6765\u8bad\u7ec3\u6a21\u578b\uff0c\u5219\u4f1a\u5728 IMDb \u7535\u5f71\u8bc4\u8bba\u5206\u7c7b\u6570\u636e\u96c6\u4e0a\u4ea7\u751f\u4ee5\u4e0b\u6027\u80fd\uff1a<\/p>\n\n\n\n