File size: 38,561 Bytes
729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 a5e200c 729bc74 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 |
{
"cells": [
{
"cell_type": "markdown",
"id": "963e9ae0-ac68-44be-8c7d-fb9842784362",
"metadata": {},
"source": [
"# 4.5 peft简介"
]
},
{
"cell_type": "markdown",
"id": "f4288594-c676-4369-aca1-730446f293d7",
"metadata": {},
"source": [
"## peft"
]
},
{
"cell_type": "markdown",
"id": "182b82c4-d484-4c15-a600-03c3b51367ec",
"metadata": {},
"source": [
"**PEFT**（Parameter-Efficient Fine-Tuning，参数高效微调）是一种优化技术，旨在以最小的参数更新实现对大规模预训练模型（如 GPT、BERT 等）的微调。PEFT 技术通过减少微调所需的参数量，显著降低了存储和计算开销，同时保留模型的性能，特别适合资源受限的场景和领域特定任务的定制化。\n",
"\n",
"---\n",
"\n",
"### **1. 核心思想**\n",
"传统的微调方式需要更新整个预训练模型的所有参数，PEFT 技术通过只调整少量的参数（如特定层或额外添加的小型模块）实现微调目标，大幅减少了训练开销和存储需求。\n",
"\n",
"---\n",
"\n",
"### **2. 常见的 PEFT 方法**\n",
"\n",
"#### **（1）Adapter 模型**\n",
"- 在每一层 Transformer 的输出中插入小型适配器模块，仅训练适配器模块的参数。\n",
"- 原始模型参数保持冻结不变。\n",
"- 优点：适配器模块参数量小，能适应不同任务。\n",
"\n",
"示例方法：\n",
"- **AdapterFusion**\n",
"- **MAD-X**\n",
"\n",
"---\n",
"\n",
"#### **（2）Prefix Tuning**\n",
"- 在 Transformer 的输入前添加一组可学习的前缀向量，这些前缀与模型的注意力机制交互。\n",
"- 只调整前缀向量的参数，而不更新原始模型。\n",
"- 优点：对生成任务效果显著，参数量进一步减少。\n",
"\n",
"---\n",
"\n",
"#### **（3）LoRA（Low-Rank Adaptation）**\n",
"- 将预训练模型中的部分权重分解为两个低秩矩阵，仅调整这些低秩矩阵的参数。\n",
"- 原始权重保持冻结状态。\n",
"- 优点：参数量极小，计算高效。\n",
" \n",
"---\n",
"\n",
"#### **（4）Prompt Tuning**\n",
"- 在输入文本中添加可学习的提示（Prompt）。\n",
"- 适合 NLP 任务中的文本生成、分类等。\n",
"- 优点：实现简单，易于集成到现有框架。\n",
"\n",
"---\n",
"\n",
"### **3. PEFT 的优势**\n",
"\n",
"1. **显著减少参数更新量**：\n",
" - 微调传统的大模型（如 GPT-3）需要更新数百亿参数，而 PEFT 仅需更新百万级别甚至更少的参数。\n",
"\n",
"2. **高效存储**：\n",
" - 每个任务的微调结果只需存储少量额外参数，而不是整个模型。\n",
"\n",
"3. **适用多任务**：\n",
" - 同一预训练模型可以通过不同的 PEFT 模块适配多个任务，无需重新训练。\n",
"\n",
"4. **降低计算开销**：\n",
" - 训练所需的内存和计算显著减少，适合资源有限的环境。\n",
"\n",
"---\n",
"\n",
"### **4. 应用场景**\n",
"\n",
"1. **领域特定任务**：\n",
" - 医疗、法律、金融等领域微调预训练模型。\n",
"\n",
"2. **多任务学习**：\n",
" - 适配多个任务，复用同一模型的预训练权重。\n",
"\n",
"3. **资源受限场景**：\n",
" - 移动设备、边缘设备上的模型部署。\n",
"\n",
"---\n",
"\n",
"### **5. Hugging Face PEFT 库**\n",
"\n",
"Hugging Face 提供了专门的 PEFT 库，支持多种参数高效微调技术：\n",
"- **安装**：\n",
" ```bash\n",
" pip install peft\n",
" ```\n",
"- **使用 LoRA 微调示例**：\n",
" ```python\n",
" from transformers import AutoModelForCausalLM, AutoTokenizer\n",
" from peft import LoraConfig, get_peft_model, TaskType\n",
"\n",
" # 加载模型和分词器\n",
" model_name = \"gpt2\"\n",
" model = AutoModelForCausalLM.from_pretrained(model_name)\n",
" tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
"\n",
" # 配置 LoRA\n",
" lora_config = LoraConfig(\n",
" task_type=TaskType.CAUSAL_LM,\n",
" r=8,\n",
" lora_alpha=32,\n",
" target_modules=[\"q_proj\", \"v_proj\"],\n",
" lora_dropout=0.1,\n",
" bias=\"none\"\n",
" )\n",
"\n",
" # 使用 LoRA 微调模型\n",
" model = get_peft_model(model, lora_config)\n",
" model.print_trainable_parameters()\n",
"\n",
" # 微调代码...\n",
" ```\n",
"\n",
"---\n",
"\n",
"### **6. PEFT 的局限性**\n",
"1. **特定任务限制**：\n",
" - 在一些复杂任务中，PEFT 方法可能不如全量微调效果好。\n",
"\n",
"2. **需要设计合适的模块**：\n",
" - 不同任务需要选择和设计合适的 PEFT 技术。\n",
"\n",
"3. **与模型架构相关**：\n",
" - PEFT 技术可能需要对模型架构进行一定程度的修改。\n",
"\n",
"---\n",
"\n",
"### **7. 小结**\n",
"PEFT 是一个极具潜力的技术，特别适合在有限资源下对大模型进行微调。它在许多领域和任务中已显示出良好的效果，例如 LoRA 和 Adapter 模型已经成为高效微调的主流方法。\n",
"\n",
"如果您需要实现高效微调，可以结合 Hugging Face 的 PEFT 库快速上手。"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "a70b2631-c9b9-49da-96c6-6760c63040ac",
"metadata": {},
"outputs": [],
"source": []
},
{
"cell_type": "markdown",
"id": "7b47ddf3-85c9-4dd8-bbbb-34fc3bd6aa1b",
"metadata": {},
"source": [
"## GPT2使用peft样例"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "5aa3d240-44e1-4811-8f61-d6ff2500a798",
"metadata": {},
"outputs": [],
"source": [
"import subprocess\n",
"import os\n",
"# 设置环境变量, autodl一般区域\n",
"result = subprocess.run('bash -c \"source /etc/network_turbo && env | grep proxy\"', shell=True, capture_output=True, text=True)\n",
"output = result.stdout\n",
"for line in output.splitlines():\n",
" if '=' in line:\n",
" var, value = line.split('=', 1)\n",
" os.environ[var] = value"
]
},
{
"cell_type": "markdown",
"id": "17bdb69d-3f0f-465e-bd60-2047a088e264",
"metadata": {},
"source": [
"如果您不确定模型中有哪些模块可以微调，可以打印模型结构："
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "41a0c049-9134-4d89-aad0-1aa2241a9fca",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "4becc479adbc472bb7672d49da16aafd",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"generation_config.json: 0%| | 0.00/124 [00:00<?, ?B/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"transformer\n",
"transformer.wte\n",
"transformer.wpe\n",
"transformer.drop\n",
"transformer.h\n",
"transformer.h.0\n",
"transformer.h.0.ln_1\n",
"transformer.h.0.attn\n",
"transformer.h.0.attn.c_attn\n",
"transformer.h.0.attn.c_proj\n",
"transformer.h.0.attn.attn_dropout\n",
"transformer.h.0.attn.resid_dropout\n",
"transformer.h.0.ln_2\n",
"transformer.h.0.mlp\n",
"transformer.h.0.mlp.c_fc\n",
"transformer.h.0.mlp.c_proj\n",
"transformer.h.0.mlp.act\n",
"transformer.h.0.mlp.dropout\n",
"transformer.h.1\n",
"transformer.h.1.ln_1\n",
"transformer.h.1.attn\n",
"transformer.h.1.attn.c_attn\n",
"transformer.h.1.attn.c_proj\n",
"transformer.h.1.attn.attn_dropout\n",
"transformer.h.1.attn.resid_dropout\n",
"transformer.h.1.ln_2\n",
"transformer.h.1.mlp\n",
"transformer.h.1.mlp.c_fc\n",
"transformer.h.1.mlp.c_proj\n",
"transformer.h.1.mlp.act\n",
"transformer.h.1.mlp.dropout\n",
"transformer.h.2\n",
"transformer.h.2.ln_1\n",
"transformer.h.2.attn\n",
"transformer.h.2.attn.c_attn\n",
"transformer.h.2.attn.c_proj\n",
"transformer.h.2.attn.attn_dropout\n",
"transformer.h.2.attn.resid_dropout\n",
"transformer.h.2.ln_2\n",
"transformer.h.2.mlp\n",
"transformer.h.2.mlp.c_fc\n",
"transformer.h.2.mlp.c_proj\n",
"transformer.h.2.mlp.act\n",
"transformer.h.2.mlp.dropout\n",
"transformer.h.3\n",
"transformer.h.3.ln_1\n",
"transformer.h.3.attn\n",
"transformer.h.3.attn.c_attn\n",
"transformer.h.3.attn.c_proj\n",
"transformer.h.3.attn.attn_dropout\n",
"transformer.h.3.attn.resid_dropout\n",
"transformer.h.3.ln_2\n",
"transformer.h.3.mlp\n",
"transformer.h.3.mlp.c_fc\n",
"transformer.h.3.mlp.c_proj\n",
"transformer.h.3.mlp.act\n",
"transformer.h.3.mlp.dropout\n",
"transformer.h.4\n",
"transformer.h.4.ln_1\n",
"transformer.h.4.attn\n",
"transformer.h.4.attn.c_attn\n",
"transformer.h.4.attn.c_proj\n",
"transformer.h.4.attn.attn_dropout\n",
"transformer.h.4.attn.resid_dropout\n",
"transformer.h.4.ln_2\n",
"transformer.h.4.mlp\n",
"transformer.h.4.mlp.c_fc\n",
"transformer.h.4.mlp.c_proj\n",
"transformer.h.4.mlp.act\n",
"transformer.h.4.mlp.dropout\n",
"transformer.h.5\n",
"transformer.h.5.ln_1\n",
"transformer.h.5.attn\n",
"transformer.h.5.attn.c_attn\n",
"transformer.h.5.attn.c_proj\n",
"transformer.h.5.attn.attn_dropout\n",
"transformer.h.5.attn.resid_dropout\n",
"transformer.h.5.ln_2\n",
"transformer.h.5.mlp\n",
"transformer.h.5.mlp.c_fc\n",
"transformer.h.5.mlp.c_proj\n",
"transformer.h.5.mlp.act\n",
"transformer.h.5.mlp.dropout\n",
"transformer.h.6\n",
"transformer.h.6.ln_1\n",
"transformer.h.6.attn\n",
"transformer.h.6.attn.c_attn\n",
"transformer.h.6.attn.c_proj\n",
"transformer.h.6.attn.attn_dropout\n",
"transformer.h.6.attn.resid_dropout\n",
"transformer.h.6.ln_2\n",
"transformer.h.6.mlp\n",
"transformer.h.6.mlp.c_fc\n",
"transformer.h.6.mlp.c_proj\n",
"transformer.h.6.mlp.act\n",
"transformer.h.6.mlp.dropout\n",
"transformer.h.7\n",
"transformer.h.7.ln_1\n",
"transformer.h.7.attn\n",
"transformer.h.7.attn.c_attn\n",
"transformer.h.7.attn.c_proj\n",
"transformer.h.7.attn.attn_dropout\n",
"transformer.h.7.attn.resid_dropout\n",
"transformer.h.7.ln_2\n",
"transformer.h.7.mlp\n",
"transformer.h.7.mlp.c_fc\n",
"transformer.h.7.mlp.c_proj\n",
"transformer.h.7.mlp.act\n",
"transformer.h.7.mlp.dropout\n",
"transformer.h.8\n",
"transformer.h.8.ln_1\n",
"transformer.h.8.attn\n",
"transformer.h.8.attn.c_attn\n",
"transformer.h.8.attn.c_proj\n",
"transformer.h.8.attn.attn_dropout\n",
"transformer.h.8.attn.resid_dropout\n",
"transformer.h.8.ln_2\n",
"transformer.h.8.mlp\n",
"transformer.h.8.mlp.c_fc\n",
"transformer.h.8.mlp.c_proj\n",
"transformer.h.8.mlp.act\n",
"transformer.h.8.mlp.dropout\n",
"transformer.h.9\n",
"transformer.h.9.ln_1\n",
"transformer.h.9.attn\n",
"transformer.h.9.attn.c_attn\n",
"transformer.h.9.attn.c_proj\n",
"transformer.h.9.attn.attn_dropout\n",
"transformer.h.9.attn.resid_dropout\n",
"transformer.h.9.ln_2\n",
"transformer.h.9.mlp\n",
"transformer.h.9.mlp.c_fc\n",
"transformer.h.9.mlp.c_proj\n",
"transformer.h.9.mlp.act\n",
"transformer.h.9.mlp.dropout\n",
"transformer.h.10\n",
"transformer.h.10.ln_1\n",
"transformer.h.10.attn\n",
"transformer.h.10.attn.c_attn\n",
"transformer.h.10.attn.c_proj\n",
"transformer.h.10.attn.attn_dropout\n",
"transformer.h.10.attn.resid_dropout\n",
"transformer.h.10.ln_2\n",
"transformer.h.10.mlp\n",
"transformer.h.10.mlp.c_fc\n",
"transformer.h.10.mlp.c_proj\n",
"transformer.h.10.mlp.act\n",
"transformer.h.10.mlp.dropout\n",
"transformer.h.11\n",
"transformer.h.11.ln_1\n",
"transformer.h.11.attn\n",
"transformer.h.11.attn.c_attn\n",
"transformer.h.11.attn.c_proj\n",
"transformer.h.11.attn.attn_dropout\n",
"transformer.h.11.attn.resid_dropout\n",
"transformer.h.11.ln_2\n",
"transformer.h.11.mlp\n",
"transformer.h.11.mlp.c_fc\n",
"transformer.h.11.mlp.c_proj\n",
"transformer.h.11.mlp.act\n",
"transformer.h.11.mlp.dropout\n",
"transformer.ln_f\n",
"lm_head\n"
]
}
],
"source": [
"from transformers import AutoModelForCausalLM\n",
"\n",
"model = AutoModelForCausalLM.from_pretrained(\"gpt2\")\n",
"\n",
"# 打印所有模块名称\n",
"for name, module in model.named_modules():\n",
" print(name)"
]
},
{
"cell_type": "markdown",
"id": "0add2f79-f35c-4638-80bb-0d8a87a9b6a7",
"metadata": {},
"source": [
"在选择 `target_modules` 时，通常会根据模块的名称选择模型的特定部分，通常使用列表中最后一个点 `.` 后的字段名或整个路径名（如果需要更精确）。以下是对这些模块的详细分析和选择建议：\n",
"\n",
"---\n",
"\n",
"### **1. 分析模块结构**\n",
"\n",
"从列表中可以看出，GPT-2 的模块层次分为以下几类：\n",
"\n",
"1. **Embedding 层**：\n",
" - `transformer.wte`：词嵌入层（Word Token Embeddings）。\n",
" - `transformer.wpe`：位置嵌入层（Position Embeddings）。\n",
"\n",
"2. **Transformer 编码器层**：\n",
" - 每层编号为 `transformer.h.<层号>`（共 12 层）。\n",
" - 每层中包含：\n",
" - **层归一化**：\n",
" - `transformer.h.<层号>.ln_1`：第一层归一化。\n",
" - `transformer.h.<层号>.ln_2`：第二层归一化。\n",
" - **自注意力模块**：\n",
" - `transformer.h.<层号>.attn.c_attn`：注意力模块的 Query、Key 和 Value 投影。\n",
" - `transformer.h.<层号>.attn.c_proj`：注意力的输出投影。\n",
" - `transformer.h.<层号>.attn.attn_dropout`：注意力的 Dropout。\n",
" - `transformer.h.<层号>.attn.resid_dropout`：残差的 Dropout。\n",
" - **前馈网络模块（MLP）**：\n",
" - `transformer.h.<层号>.mlp.c_fc`：MLP 的第一层全连接。\n",
" - `transformer.h.<层号>.mlp.c_proj`：MLP 的第二层全连接（输出投影）。\n",
" - `transformer.h.<层号>.mlp.act`：激活函数（如 GELU）。\n",
" - `transformer.h.<层号>.mlp.dropout`：MLP 的 Dropout。\n",
"\n",
"3. **最终层**：\n",
" - `transformer.ln_f`：最终层归一化（LayerNorm）。\n",
" - `lm_head`：语言建模头，用于生成预测的 token 分布。\n",
"\n",
"---\n",
"\n",
"### **2. 如何选择 `target_modules`**\n",
"\n",
"#### **（1）常见目标模块**\n",
"- `transformer.h.<层号>.attn.c_attn`：对自注意力模块的 Query、Key 和 Value 投影层微调。\n",
"- `transformer.h.<层号>.attn.c_proj`：对注意力输出的投影层微调。\n",
"- `transformer.h.<层号>.mlp.c_fc`：对前馈网络的输入全连接层微调。\n",
"- `transformer.h.<层号>.mlp.c_proj`：对前馈网络的输出投影层微调。\n",
"\n",
"#### **（2）推荐设置**\n",
"- **文本生成任务**：\n",
" ```python\n",
" target_modules = [\"transformer.h.*.attn.c_attn\", \"transformer.h.*.attn.c_proj\"]\n",
" ```\n",
" 解释：\n",
" - `*.attn.c_attn`：调整 Query、Key、Value 的生成。\n",
" - `*.attn.c_proj`：调整注意力输出。\n",
"\n",
"- **文本分类任务**：\n",
" ```python\n",
" target_modules = [\"transformer.h.*.attn.c_attn\"]\n",
" ```\n",
" 解释：\n",
" - 微调自注意力模块最重要的部分即可。\n",
"\n",
"- **特定任务需要更细粒度控制**：\n",
" - 仅微调某几层：\n",
" ```python\n",
" target_modules = [\"transformer.h.0.attn.c_attn\", \"transformer.h.0.mlp.c_fc\"]\n",
" ```\n",
"\n",
"#### **（3）通配符选择**\n",
"使用 `*` 通配符可以指定所有层的某些模块：\n",
"- `transformer.h.*.attn.c_attn`：所有层的 Query、Key 和 Value 投影。\n",
"- `transformer.h.*.mlp.*`：所有层的 MLP 模块。\n",
"\n",
"---\n",
"\n",
"### **3. 示例：指定多个模块**\n",
"\n",
"```python\n",
"lora_config = LoraConfig(\n",
" task_type=TaskType.CAUSAL_LM,\n",
" r=8,\n",
" lora_alpha=32,\n",
" target_modules=[\n",
" \"transformer.h.*.attn.c_attn\",\n",
" \"transformer.h.*.mlp.c_fc\"\n",
" ],\n",
" lora_dropout=0.1,\n",
" bias=\"none\"\n",
")\n",
"```\n",
"\n",
"- 这表示对所有层的 `attn.c_attn` 和 `mlp.c_fc` 模块进行 LoRA 微调。\n",
"\n",
"---\n",
"\n",
"### **4. 小提示：如何确定适合的模块**\n",
"\n",
"1. **任务相关性**：\n",
" - 文本生成：优先选择自注意力模块（如 `c_attn`）。\n",
" - 文本分类：通常需要全局语义表示，选择 `attn.c_attn` 或 `mlp.c_fc`。\n",
"\n",
"2. **性能与资源平衡**：\n",
" - 如果显存有限，可以只微调部分层。例如，仅选择浅层和深层的模块：\n",
" ```python\n",
" target_modules = [\"transformer.h.0.attn.c_attn\", \"transformer.h.11.attn.c_attn\"]\n",
" ```\n",
"\n",
"3. **打印模块名称以调试**：\n",
" - 确保选择的 `target_modules` 在模型中实际存在：\n",
" ```python\n",
" for name, _ in model.named_modules():\n",
" if \"c_attn\" in name:\n",
" print(name)\n",
" ```\n",
"\n",
"---\n",
"\n",
"### **建议**\n",
"- 一般情况下，`c_attn` 和 `c_proj` 是首选模块。\n",
"- 使用 `transformer.h.*` 通配符可以轻松指定多层。\n",
"- 根据任务需求和资源限制灵活调整目标模块，以实现最佳性能和效率。"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "14f20171-0719-4dfa-b888-147b657ebff4",
"metadata": {},
"outputs": [],
"source": []
},
{
"cell_type": "markdown",
"id": "b4e7bff2-2a4f-4a1d-9cb1-dd02aead2f85",
"metadata": {},
"source": [
"## LoraConfig具体配置"
]
},
{
"cell_type": "markdown",
"id": "10c99eb9-8007-4297-972e-7be71768c9c3",
"metadata": {},
"source": [
"以下是对 `LoraConfig` 配置的更详细解释，特别是如何设置微调哪些参数、冻结哪些参数，以及一般如何选择这些设置：\n",
"\n",
"---\n",
"\n",
"### **1. `LoraConfig` 参数解析**\n",
"\n",
"```python\n",
"lora_config = LoraConfig(\n",
" task_type=TaskType.SEQ_CLS, # 序列分类任务\n",
" r=8, # 降低矩阵秩\n",
" lora_alpha=32, # LoRA 的 alpha 超参数\n",
" target_modules=[\"c_attn\"], # GPT-2 中的自注意力模块\n",
" lora_dropout=0.1, # dropout 概率\n",
" bias=\"none\", # 是否微调偏置参数\n",
")\n",
"```\n",
"\n",
"#### **（1）`task_type`**\n",
"- 定义任务类型，用于指导 PEFT 的具体行为。\n",
"- **常见选项**：\n",
" - `TaskType.CAUSAL_LM`：自回归语言建模（如 GPT 系列模型）。\n",
" - `TaskType.SEQ_CLS`：序列分类（如情感分析）。\n",
" - `TaskType.TOKEN_CLS`：标注任务（如命名实体识别）。\n",
" - `TaskType.SEQ_2_SEQ_LM`：序列到序列任务（如翻译、摘要）。\n",
"\n",
"**当前设置**：\n",
"- `TaskType.SEQ_CLS` 表示目标是文本分类任务。\n",
"\n",
"---\n",
"\n",
"#### **（2）`r`**\n",
"- 表示 LoRA 的 **秩**（rank），是降低矩阵秩的核心参数。\n",
"- LoRA 通过将模型的权重分解为两个低秩矩阵（`A` 和 `B`），只更新这两个矩阵。\n",
"- `r` 的值越大，微调能力越强，但需要的额外参数也越多。\n",
"- **典型范围**：`4` 至 `64`，大多数任务中 `8` 或 `16` 是常用值。\n",
"\n",
"**当前设置**：\n",
"- `r=8` 表示使用低秩分解，并微调 8 维的参数矩阵。\n",
"\n",
"---\n",
"\n",
"#### **（3）`lora_alpha`**\n",
"- 是 LoRA 的一个缩放因子，用于调节两个低秩矩阵的更新速率。\n",
"- **公式**：实际更新 = LoRA 输出 × `lora_alpha / r`\n",
"- **典型范围**：`16` 至 `128`，较大任务中可以选择更高的值。\n",
"\n",
"**当前设置**：\n",
"- `lora_alpha=32`，表示适中幅度的更新速率。\n",
"\n",
"---\n",
"\n",
"#### **（4）`target_modules`**\n",
"- 指定要应用 LoRA 微调的模块。\n",
"- **常见选择**：\n",
" - 对 Transformer 模型中的 **注意力模块**（如 `query`、`key`、`value`）进行微调，因为这些模块对任务性能影响较大。\n",
" - 对 GPT-2，通常选择 `c_attn`（GPT-2 中负责自注意力机制的组合模块）。\n",
"\n",
"**当前设置**：\n",
"- `target_modules=[\"c_attn\"]` 表示只对 GPT-2 的自注意力模块 `c_attn` 应用 LoRA。\n",
"\n",
"---\n",
"\n",
"#### **（5）`lora_dropout`**\n",
"- 表示 LoRA 层的 dropout 概率，用于防止过拟合。\n",
"- **典型范围**：`0.0` 至 `0.1`，视任务复杂性而定。\n",
"\n",
"**当前设置**：\n",
"- `lora_dropout=0.1`，表示有 10% 的概率随机丢弃 LoRA 层的输出。\n",
"\n",
"---\n",
"\n",
"#### **（6）`bias`**\n",
"- 决定是否微调偏置参数。\n",
"- **选项**：\n",
" - `\"none\"`：不微调任何偏置。\n",
" - `\"all\"`：微调所有偏置。\n",
" - `\"lora_only\"`：只微调 LoRA 层的偏置。\n",
"\n",
"**当前设置**：\n",
"- `bias=\"none\"`，表示所有偏置参数保持冻结。\n",
"\n",
"---\n",
"\n",
"### **5. 总结建议**\n",
"- **微调的参数**：优先选择模型中注意力相关模块。\n",
"- **冻结的参数**：大部分参数默认冻结以节省显存。\n",
"- **配置选择**：根据任务复杂性调整 `r` 和 `target_modules`。\n",
"- **推荐起点**：\n",
" - 文本分类：`target_modules=[\"c_attn\"]`, `r=8`, `lora_dropout=0.1`。\n",
" - 文本生成：`target_modules=[\"q_proj\", \"v_proj\"]`, `r=16`, `lora_dropout=0.1`。\n",
"\n",
"通过这些设置，LoRA 可以在参数量极小的情况下实现高效微调，适合各种任务场景。"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "bbc080ba-3ee8-4bc6-afd9-2a3241f1bcda",
"metadata": {},
"outputs": [],
"source": [
"import subprocess\n",
"import os\n",
"# 设置环境变量, autodl一般区域\n",
"result = subprocess.run('bash -c \"source /etc/network_turbo && env | grep proxy\"', shell=True, capture_output=True, text=True)\n",
"output = result.stdout\n",
"for line in output.splitlines():\n",
" if '=' in line:\n",
" var, value = line.split('=', 1)\n",
" os.environ[var] = value"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "26d9f362-18cc-471f-b208-f29a6933c06a",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"Some weights of GPT2ForSequenceClassification were not initialized from the model checkpoint at dnagpt/dna_gpt2_v0 and are newly initialized: ['score.weight']\n",
"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
]
},
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "f7e72521368341d38a2b11028715a871",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
"Map: 0%| | 0/5920 [00:00<?, ? examples/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trainable params: 296,448 || all params: 109,180,416 || trainable%: 0.2715\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"/root/miniconda3/lib/python3.12/site-packages/peft/tuners/lora/layer.py:1264: UserWarning: fan_in_fan_out is set to False but the target module is `Conv1D`. Setting fan_in_fan_out to True.\n",
" warnings.warn(\n"
]
}
],
"source": [
"from transformers import AutoModelForSequenceClassification, AutoTokenizer, TrainingArguments, Trainer\n",
"from peft import LoraConfig, get_peft_model, TaskType\n",
"from datasets import load_dataset\n",
"from sklearn.metrics import accuracy_score, precision_recall_fscore_support\n",
"from transformers import DataCollatorWithPadding\n",
"\n",
"# **1. 加载模型和分词器**\n",
"model_name = \"dnagpt/dna_gpt2_v0\" # 基础模型\n",
"num_labels = 2 # 二分类任务\n",
"model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=num_labels)\n",
"tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
"\n",
"tokenizer.pad_token = tokenizer.eos_token\n",
"model.config.pad_token_id = tokenizer.pad_token_id\n",
"\n",
"\n",
"# **2. 定义数据集**\n",
"# 示例数据集：dna_promoter_300\n",
"dataset = load_dataset(\"dnagpt/dna_promoter_300\")['train'].train_test_split(test_size=0.1)\n",
"\n",
"# **3. 数据预处理**\n",
"def preprocess_function(examples):\n",
" examples['label'] = [int(item) for item in examples['label']]\n",
" return tokenizer(\n",
" examples[\"sequence\"], truncation=True, padding=\"max_length\", max_length=128\n",
" )\n",
"\n",
"tokenized_datasets = dataset.map(preprocess_function, batched=True)\n",
"#tokenized_datasets = tokenized_datasets.rename_column(\"label\", \"labels\") # Hugging Face Trainer 要求标签列名为 'labels'\n",
"\n",
"# 4. 创建一个数据收集器，用于动态填充和遮蔽\n",
"data_collator = DataCollatorWithPadding(tokenizer=tokenizer)\n",
"\n",
"# **4. 划分数据集**\n",
"train_dataset = tokenized_datasets[\"train\"]\n",
"test_dataset = tokenized_datasets[\"test\"]\n",
"\n",
"# **5. 配置 LoRA**\n",
"lora_config = LoraConfig(\n",
" task_type=TaskType.SEQ_CLS, # 序列分类任务\n",
" r=8, # 降低矩阵秩\n",
" lora_alpha=32, # LoRA 的 alpha 超参数\n",
" target_modules=[\"c_attn\"], # GPT-2 中的自注意力模块\n",
" lora_dropout=0.1, # dropout 概率\n",
" bias=\"none\", # 是否微调偏置参数\n",
")\n",
"\n",
"# 使用 LoRA 包装模型\n",
"model = get_peft_model(model, lora_config)\n",
"model.print_trainable_parameters() # 打印可训练的参数信息"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "7da39e7f-db92-483c-888d-19707ab35c5f",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/root/miniconda3/lib/python3.12/site-packages/transformers/training_args.py:1575: FutureWarning: `evaluation_strategy` is deprecated and will be removed in version 4.46 of 🤗 Transformers. Use `eval_strategy` instead\n",
" warnings.warn(\n",
"/tmp/ipykernel_2399/3695291394.py:28: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n",
" trainer = Trainer(\n"
]
},
{
"data": {
"text/html": [
"\n",
" <div>\n",
" \n",
" <progress value='66600' max='66600' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
" [66600/66600 34:07, Epoch 10/10]\n",
" </div>\n",
" <table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: left;\">\n",
" <th>Epoch</th>\n",
" <th>Training Loss</th>\n",
" <th>Validation Loss</th>\n",
" <th>Accuracy</th>\n",
" <th>Precision</th>\n",
" <th>Recall</th>\n",
" <th>F1</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <td>1</td>\n",
" <td>0.268300</td>\n",
" <td>0.307843</td>\n",
" <td>0.909797</td>\n",
" <td>0.916809</td>\n",
" <td>0.901987</td>\n",
" <td>0.909338</td>\n",
" </tr>\n",
" <tr>\n",
" <td>2</td>\n",
" <td>0.287400</td>\n",
" <td>0.278804</td>\n",
" <td>0.913514</td>\n",
" <td>0.901339</td>\n",
" <td>0.929269</td>\n",
" <td>0.915091</td>\n",
" </tr>\n",
" <tr>\n",
" <td>3</td>\n",
" <td>0.282800</td>\n",
" <td>0.291222</td>\n",
" <td>0.914527</td>\n",
" <td>0.913116</td>\n",
" <td>0.916807</td>\n",
" <td>0.914958</td>\n",
" </tr>\n",
" <tr>\n",
" <td>4</td>\n",
" <td>0.255200</td>\n",
" <td>0.281572</td>\n",
" <td>0.916385</td>\n",
" <td>0.896474</td>\n",
" <td>0.942068</td>\n",
" <td>0.918706</td>\n",
" </tr>\n",
" <tr>\n",
" <td>5</td>\n",
" <td>0.252000</td>\n",
" <td>0.271950</td>\n",
" <td>0.914527</td>\n",
" <td>0.913116</td>\n",
" <td>0.916807</td>\n",
" <td>0.914958</td>\n",
" </tr>\n",
" <tr>\n",
" <td>6</td>\n",
" <td>0.242300</td>\n",
" <td>0.288199</td>\n",
" <td>0.916385</td>\n",
" <td>0.916498</td>\n",
" <td>0.916807</td>\n",
" <td>0.916653</td>\n",
" </tr>\n",
" <tr>\n",
" <td>7</td>\n",
" <td>0.253500</td>\n",
" <td>0.268673</td>\n",
" <td>0.918750</td>\n",
" <td>0.909480</td>\n",
" <td>0.930616</td>\n",
" <td>0.919927</td>\n",
" </tr>\n",
" <tr>\n",
" <td>8</td>\n",
" <td>0.235900</td>\n",
" <td>0.277893</td>\n",
" <td>0.917568</td>\n",
" <td>0.906855</td>\n",
" <td>0.931290</td>\n",
" <td>0.918910</td>\n",
" </tr>\n",
" <tr>\n",
" <td>9</td>\n",
" <td>0.238600</td>\n",
" <td>0.280647</td>\n",
" <td>0.917568</td>\n",
" <td>0.913362</td>\n",
" <td>0.923206</td>\n",
" <td>0.918258</td>\n",
" </tr>\n",
" <tr>\n",
" <td>10</td>\n",
" <td>0.237900</td>\n",
" <td>0.284149</td>\n",
" <td>0.917736</td>\n",
" <td>0.913391</td>\n",
" <td>0.923543</td>\n",
" <td>0.918439</td>\n",
" </tr>\n",
" </tbody>\n",
"</table><p>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"训练完成，模型已保存至 ./gpt2_lora_text_classification\n"
]
}
],
"source": [
"# **6. 计算指标**\n",
"def compute_metrics(eval_pred):\n",
" predictions, labels = eval_pred\n",
" preds = predictions.argmax(axis=-1)\n",
" precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average=\"binary\")\n",
" acc = accuracy_score(labels, preds)\n",
" return {\"accuracy\": acc, \"precision\": precision, \"recall\": recall, \"f1\": f1}\n",
"\n",
"# **7. 定义训练参数**\n",
"training_args = TrainingArguments(\n",
" output_dir=\"./gpt2_lora_text_classification\", # 模型保存路径\n",
" evaluation_strategy=\"epoch\", # 每个 epoch 评估一次\n",
" save_strategy=\"epoch\", # 每个 epoch 保存一次\n",
" learning_rate=2e-5, # 学习率\n",
" per_device_train_batch_size=8, # 每设备的批量大小\n",
" per_device_eval_batch_size=8, # 每设备评估的批量大小\n",
" num_train_epochs=10, # 训练轮数\n",
" weight_decay=0.01, # 权重衰减\n",
" logging_dir=\"./logs\", # 日志路径\n",
" fp16=True, # 启用混合精度训练\n",
" save_total_limit=2, # 保留最多两个检查点\n",
" load_best_model_at_end=True, # 加载最佳模型\n",
" metric_for_best_model=\"accuracy\", # 根据准确率选择最佳模型\n",
" greater_is_better=True,\n",
")\n",
"\n",
"# **8. 定义 Trainer**\n",
"trainer = Trainer(\n",
" model=model,\n",
" args=training_args,\n",
" train_dataset=train_dataset,\n",
" eval_dataset=test_dataset,\n",
" tokenizer=tokenizer,\n",
" data_collator=data_collator,\n",
" compute_metrics=compute_metrics,\n",
")\n",
"\n",
"# **9. 开始训练**\n",
"trainer.train()\n",
"\n",
"# **10. 保存模型**\n",
"model.save_pretrained(\"./gpt2_lora_text_classification\")\n",
"tokenizer.save_pretrained(\"./gpt2_lora_text_classification\")\n",
"\n",
"print(\"训练完成，模型已保存至 ./gpt2_lora_text_classification\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "49a60fed-3a7d-4608-98b1-b4e313b94dbb",
"metadata": {},
"outputs": [],
"source": [
"from transformers import AutoModelForSequenceClassification, AutoTokenizer\n",
"from peft import PeftModel\n",
"\n",
"# 加载分词器\n",
"model_path = \"./gpt2_lora_text_classification\"\n",
"tokenizer = AutoTokenizer.from_pretrained(model_path)\n",
"\n",
"# 加载微调后的 PEFT 模型\n",
"base_model = AutoModelForSequenceClassification.from_pretrained(\"gpt2\", num_labels=2)\n",
"model = PeftModel.from_pretrained(base_model, model_path)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3c0d8f02-c3dc-4961-8b3a-50eefc5f9448",
"metadata": {},
"outputs": [],
"source": [
"import torch\n",
"\n",
"def predict(texts, model, tokenizer):\n",
" \"\"\"\n",
" 使用微调后的 PEFT 模型进行推理。\n",
" \n",
" Args:\n",
" texts (list of str): 待分类的文本列表。\n",
" model (PeftModel): 微调后的模型。\n",
" tokenizer (AutoTokenizer): 分词器。\n",
" \n",
" Returns:\n",
" list of dict: 每个文本的预测结果，包括 logits 和预测的类别标签。\n",
" \"\"\"\n",
" # 对输入文本进行分词和编码\n",
" inputs = tokenizer(\n",
" texts,\n",
" padding=True,\n",
" truncation=True,\n",
" max_length=512,\n",
" return_tensors=\"pt\"\n",
" )\n",
" \n",
" # 将输入数据移动到模型的设备上（CPU/GPU）\n",
" inputs = {key: value.to(model.device) for key, value in inputs.items()}\n",
" \n",
" # 模型推理\n",
" model.eval()\n",
" with torch.no_grad():\n",
" outputs = model(**inputs)\n",
" \n",
" # 获取 logits 并计算预测类别\n",
" logits = outputs.logits\n",
" probs = torch.nn.functional.softmax(logits, dim=-1)\n",
" predictions = torch.argmax(probs, dim=-1)\n",
" \n",
" # 返回每个文本的预测结果\n",
" results = [\n",
" {\"text\": text, \"logits\": logit.tolist(), \"predicted_class\": int(pred)}\n",
" for text, logit, pred in zip(texts, logits, predictions)\n",
" ]\n",
" return results\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "9c0cfe65-f4f3-4274-a4f4-1ac13725b15a",
"metadata": {},
"outputs": [],
"source": [
"Text: This movie was fantastic! I loved every part of it.\n",
"Predicted Class: 1\n",
"Logits: [-2.345, 3.567]\n",
"\n",
"Text: The plot was terrible and the acting was worse.\n",
"Predicted Class: 0\n",
"Logits: [4.123, -1.234]\n"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.3"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
|