\u672c\u6307\u5357\u5c55\u793a\u4e86\u5982\u4f55\u5fae\u8c03 Segformer\uff0c\u8fd9\u662f\u4e00\u79cd\u6700\u5148\u8fdb\u7684\u8bed\u4e49\u5206\u5272\u6a21\u578b\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u4e3a\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u5efa\u7acb\u4e00\u4e2a\u6a21\u578b\uff0c\u8fd9\u6837\u5b83\u5c31\u53ef\u4ee5\u770b\u5230\u8981\u884c\u9a76\u7684\u65b9\u5411\u5e76\u8bc6\u522b\u969c\u788d\u7269 \ud83e\udd16\u3002<\/p>\n
\u6211\u4eec\u5c06\u9996\u5148\u5728 Segments.ai \u4e0a\u6807\u8bb0\u4e00\u7ec4\u4eba\u884c\u9053\u56fe\u50cf\u3002\u7136\u540e\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 \ud83e\udd17 transformers \u5fae\u8c03\u9884\u5148\u8bad\u7ec3\u7684 SegFormer \u6a21\u578b\uff0ctransformers \u662f\u4e00\u4e2a\u5f00\u6e90\u5e93\uff0c\u63d0\u4f9b\u6700\u5148\u8fdb\u6a21\u578b\u7684\u6613\u4e8e\u4f7f\u7528\u7684\u5b9e\u73b0\u3002\u5728\u6b64\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u4f7f\u7528 Hugging Face Hub\uff0c\u8fd9\u662f\u6700\u5927\u7684\u5f00\u6e90\u6a21\u578b\u548c\u6570\u636e\u96c6\u76ee\u5f55\u3002<\/p>\n
\u8bed\u4e49\u5206\u5272\u662f\u5bf9\u56fe\u50cf\u4e2d\u7684\u6bcf\u4e2a\u50cf\u7d20\u8fdb\u884c\u5206\u7c7b\u7684\u4efb\u52a1\u3002\u4f60\u53ef\u4ee5\u5c06\u5176\u89c6\u4e3a\u5bf9\u56fe\u50cf\u8fdb\u884c\u66f4\u7cbe\u786e\u5206\u7c7b\u7684\u65b9\u6cd5\u3002\u5b83\u5728\u533b\u5b66\u6210\u50cf\u548c\u81ea\u52a8\u9a7e\u9a76\u7b49\u9886\u57df\u6709\u5e7f\u6cdb\u7684\u7528\u4f8b\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u6211\u4eec\u7684\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u6765\u8bf4\uff0c\u91cd\u8981\u7684\u662f\u8981\u786e\u5207\u5730\u77e5\u9053\u4eba\u884c\u9053\u5728\u56fe\u50cf\u4e2d\u7684\u4f4d\u7f6e\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u662f\u5426\u6709\u4eba\u884c\u9053\u3002<\/p>\n
\u56e0\u4e3a\u8bed\u4e49\u5206\u5272\u662f\u4e00\u79cd\u5206\u7c7b\uff0c\u6240\u4ee5\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\u548c\u8bed\u4e49\u5206\u5272\u7684\u7f51\u7edc\u67b6\u6784\u975e\u5e38\u76f8\u4f3c\u3002 2014 \u5e74\uff0cLong \u7b49\u4eba\u53d1\u8868\u4e86\u4e00\u7bc7\uff0c\u4f7f\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u8bed\u4e49\u5206\u5272\u3002 \u6700\u8fd1\uff0cTransformers \u5df2\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\uff08\u4f8b\u5982 ViT\uff09\uff0c\u73b0\u5728\u5b83\u4eec\u4e5f\u7528\u4e8e\u8bed\u4e49\u5206\u5272\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86\u6700\u5148\u8fdb\u7684\u6280\u672f\u3002<\/p>\n
\u662f Xie \u7b49\u4eba\u4e8e 2021 \u5e74\u5f15\u5165\u7684\u8bed\u4e49\u5206\u5272\u6a21\u578b\u3002 \u5b83\u6709\u4e00\u4e2a\u4e0d\u4f7f\u7528\u4f4d\u7f6e\u7f16\u7801\u7684\u5206\u5c42 Transformer \u7f16\u7801\u5668\uff08\u4e0e ViT \u76f8\u53cd\uff09\u548c\u4e00\u4e2a\u7b80\u5355\u7684\u591a\u5c42\u611f\u77e5\u5668\u89e3\u7801\u5668\u3002 SegFormer \u5728\u591a\u4e2a\u5e38\u89c1\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002 \u8ba9\u6211\u4eec\u770b\u770b\u6211\u4eec\u7684\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u5728\u4eba\u884c\u9053\u56fe\u50cf\u4e0a\u7684\u8868\u73b0\u5982\u4f55\u3002<\/p>\n
\u62ab\u8428\u5916\u9001\u673a\u5668\u4eba\u5206\u5272\u573a\u666f <\/p>\n
\u8ba9\u6211\u4eec\u4ece\u5b89\u88c5\u5fc5\u8981\u7684\u4f9d\u8d56\u9879\u5f00\u59cb\u3002\u56e0\u4e3a\u6211\u4eec\u8981\u5c06\u6570\u636e\u96c6\u548c\u6a21\u578b\u63a8\u9001\u5230 Hugging Face Hub\uff0c\u6240\u4ee5\u6211\u4eec\u9700\u8981\u5b89\u88c5 Git LFS \u5e76\u767b\u5f55 Hugging Face\u3002<\/p>\n
git-lfs \u7684\u5b89\u88c5\u5728\u60a8\u7684\u7cfb\u7edf\u4e0a\u53ef\u80fd\u6709\u6240\u4e0d\u540c\u3002\u8bf7\u6ce8\u610f\uff0cGoogle Colab \u5df2\u9884\u88c5 Git LFS\u3002<\/p>\n
pip install -q transformers datasets evaluate segments-ai\napt-get install git-lfs\ngit lfs install\nhuggingface-cli login\n<\/code><\/pre>\n1\u3001\u521b\u5efa\/\u9009\u62e9\u6570\u636e\u96c6<\/h2>\n
\u4efb\u4f55 ML \u9879\u76ee\u7684\u7b2c\u4e00\u6b65\u90fd\u662f\u7ec4\u88c5\u4e00\u4e2a\u597d\u7684\u6570\u636e\u96c6\u3002\u4e3a\u4e86\u8bad\u7ec3\u8bed\u4e49\u5206\u5272\u6a21\u578b\uff0c\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u5e26\u6709\u8bed\u4e49\u5206\u5272\u6807\u7b7e\u7684\u6570\u636e\u96c6\u3002\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528 Hugging Face Hub \u4e2d\u7684\u73b0\u6709\u6570\u636e\u96c6\uff08\u4f8b\u5982 \uff09\uff0c\u4e5f\u53ef\u4ee5\u521b\u5efa\u81ea\u5df1\u7684\u6570\u636e\u96c6\u3002<\/p>\n
\u5bf9\u4e8e\u6211\u4eec\u7684\u62ab\u8428\u5916\u9001\u673a\u5668\u4eba\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u73b0\u6709\u7684\u81ea\u52a8\u9a7e\u9a76\u6570\u636e\u96c6\uff08\u4f8b\u5982 \u6216 \uff09\u3002\u4f46\u662f\uff0c\u8fd9\u4e9b\u6570\u636e\u96c6\u662f\u7531\u884c\u9a76\u5728\u9053\u8def\u4e0a\u7684\u6c7d\u8f66\u6355\u83b7\u7684\u3002\u7531\u4e8e\u6211\u4eec\u7684\u9001\u8d27\u673a\u5668\u4eba\u5c06\u5728\u4eba\u884c\u9053\u4e0a\u884c\u9a76\uff0c\u56e0\u6b64\u8fd9\u4e9b\u6570\u636e\u96c6\u4e2d\u7684\u56fe\u50cf\u4e0e\u673a\u5668\u4eba\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u770b\u5230\u7684\u6570\u636e\u4f1a\u4e0d\u5339\u914d\u3002<\/p>\n
\u6211\u4eec\u4e0d\u5e0c\u671b\u9001\u8d27\u673a\u5668\u4eba\u611f\u5230\u56f0\u60d1\uff0c\u56e0\u6b64\u6211\u4eec\u5c06\u4f7f\u7528\u5728\u4eba\u884c\u9053\u4e0a\u62cd\u6444\u7684\u56fe\u50cf\u521b\u5efa\u81ea\u5df1\u7684\u8bed\u4e49\u5206\u5272\u6570\u636e\u96c6\u3002\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u6b65\u9aa4\u4e2d\u5c55\u793a\u5982\u4f55\u6807\u8bb0\u6211\u4eec\u62cd\u6444\u7684\u56fe\u50cf\u3002\u5982\u679c\u4f60\u53ea\u60f3\u4f7f\u7528\u6211\u4eec\u5b8c\u6210\u7684\u6807\u8bb0\u6570\u636e\u96c6\uff0c\u5219\u53ef\u4ee5\u8df3\u8fc7\u201c\u521b\u5efa\u81ea\u5df1\u7684\u6570\u636e\u96c6\u201d\u90e8\u5206\uff0c\u7136\u540e\u7ee7\u7eed\u201c\u4f7f\u7528 Hub \u4e2d\u7684\u6570\u636e\u96c6\u201d\u3002<\/p>\n
1.1 \u521b\u5efa\u81ea\u5df1\u7684\u6570\u636e\u96c6<\/h3>\n
\u8981\u521b\u5efa\u8bed\u4e49\u5206\u5272\u6570\u636e\u96c6\uff0c\u9700\u8981\u4e24\u6837\u4e1c\u897f\uff1a<\/p>\n
\n- \u6db5\u76d6\u6a21\u578b\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u5c06\u9047\u5230\u7684\u60c5\u51b5\u7684\u56fe\u50cf<\/li>\n
- \u5206\u5272\u6807\u7b7e\uff0c\u5373\u6bcf\u4e2a\u50cf\u7d20\u4ee3\u8868\u4e00\u4e2a\u7c7b\/\u7c7b\u522b\u7684\u56fe\u50cf\u3002<\/li>\n<\/ul>\n
\u6211\u4eec\u7ee7\u7eed\u62cd\u6444\u4e86\u4e00\u5343\u5f20\u6bd4\u5229\u65f6\u4eba\u884c\u9053\u7684\u56fe\u50cf\u3002\u6536\u96c6\u548c\u6807\u8bb0\u8fd9\u6837\u7684\u6570\u636e\u96c6\u53ef\u80fd\u9700\u8981\u5f88\u957f\u65f6\u95f4\uff0c\u56e0\u6b64\u4f60\u53ef\u4ee5\u4ece\u8f83\u5c0f\u7684\u6570\u636e\u96c6\u5f00\u59cb\uff0c\u5982\u679c\u6a21\u578b\u8868\u73b0\u4e0d\u591f\u597d\uff0c\u5219\u53ef\u4ee5\u6269\u5c55\u5b83\u3002<\/p>\n
\u4eba\u884c\u9053\u6570\u636e\u96c6\u4e2d\u539f\u59cb\u56fe\u50cf\u7684\u4e00\u4e9b\u793a\u4f8b <\/p>\n
\u8981\u83b7\u5f97\u5206\u5272\u6807\u7b7e\uff0c\u6211\u4eec\u9700\u8981\u6307\u51fa\u8fd9\u4e9b\u56fe\u50cf\u4e2d\u6240\u6709\u533a\u57df\/\u5bf9\u8c61\u7684\u7c7b\u522b\u3002\u8fd9\u53ef\u80fd\u662f\u4e00\u9879\u8017\u65f6\u7684\u5de5\u4f5c\uff0c\u4f46\u4f7f\u7528\u6b63\u786e\u7684\u5de5\u5177\u53ef\u4ee5\u5927\u5927\u52a0\u5feb\u4efb\u52a1\u901f\u5ea6\u3002\u5bf9\u4e8e\u6807\u6ce8\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 \uff0c\u56e0\u4e3a\u5b83\u5177\u6709\u7528\u4e8e\u56fe\u50cf\u5206\u5272\u7684\u667a\u80fd\u6807\u8bb0\u5de5\u5177\u548c\u6613\u4e8e\u4f7f\u7528\u7684 Python SDK\u3002<\/p>\n
1.2 \u5728 Segments.ai \u4e0a\u8bbe\u7f6e\u6807\u6ce8\u4efb\u52a1<\/h3>\n
\u9996\u5148\uff0c\u524d\u5f80 \u521b\u5efa\u4e00\u4e2a\u5e10\u6237\u3002\u63a5\u4e0b\u6765\uff0c\u521b\u5efa\u4e00\u4e2a\u65b0\u7684\u6570\u636e\u96c6\u5e76\u4e0a\u4f20\u4f60\u7684\u56fe\u50cf\u3002\u4f60\u53ef\u4ee5\u4ece Web \u754c\u9762\u6216\u901a\u8fc7 Python SDK \u6267\u884c\u6b64\u64cd\u4f5c\uff08\u53c2\u89c1\uff09\u3002<\/p>\n
1.3 \u6807\u6ce8\u56fe\u50cf<\/h3>\n
\u73b0\u5728\u539f\u59cb\u6570\u636e\u5df2\u52a0\u8f7d\uff0c\u8bf7\u8f6c\u5230\u5e76\u6253\u5f00\u65b0\u521b\u5efa\u7684\u6570\u636e\u96c6\u3002\u5355\u51fb\u201c\u5f00\u59cb\u6807\u8bb0\u201d\u5e76\u521b\u5efa\u5206\u5272\u8499\u7248\u3002\u4f60\u53ef\u4ee5\u4f7f\u7528\u7531 ML \u9a71\u52a8\u7684\u8d85\u50cf\u7d20\u548c\u81ea\u52a8\u5206\u5272\u5de5\u5177\u66f4\u5feb\u5730\u8fdb\u884c\u6807\u8bb0\u3002<\/p>\n
\u63d0\u793a\uff1a\u4f7f\u7528\u8d85\u50cf\u7d20\u5de5\u5177\u65f6\uff0c\u6eda\u52a8\u4ee5\u66f4\u6539\u8d85\u50cf\u7d20\u5927\u5c0f\uff0c\u7136\u540e\u5355\u51fb\u5e76\u62d6\u52a8\u4ee5\u9009\u62e9\u7247\u6bb5\u3002 <\/p>\n
1.4 \u5c06\u7ed3\u679c\u63a8\u9001\u5230 Hugging Face Hub<\/h3>\n
\u5b8c\u6210\u6807\u6ce8\u540e\uff0c\u521b\u5efa\u4e00\u4e2a\u5305\u542b\u6807\u8bb0\u6570\u636e\u7684\u65b0\u6570\u636e\u96c6\u7248\u672c\u3002\u4f60\u53ef\u4ee5\u5728 Segments.ai \u4e0a\u7684\u53d1\u5e03\u9009\u9879\u5361\u4e0a\u6267\u884c\u6b64\u64cd\u4f5c\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7 SDK \u4ee5\u7f16\u7a0b\u65b9\u5f0f\u6267\u884c\u6b64\u64cd\u4f5c\uff0c\u5982\u7b14\u8bb0\u672c\u4e2d\u6240\u793a\u3002<\/p>\n
\u8bf7\u6ce8\u610f\u521b\u5efa\u53d1\u5e03\u53ef\u80fd\u9700\u8981\u51e0\u79d2\u949f\u3002\u4f60\u53ef\u4ee5\u68c0\u67e5 Segments.ai \u4e0a\u7684\u53d1\u5e03\u9009\u9879\u5361\uff0c\u4ee5\u68c0\u67e5\u4f60\u7684\u53d1\u5e03\u662f\u5426\u4ecd\u5728\u521b\u5efa\u4e2d\u3002<\/p>\n
\u73b0\u5728\uff0c\u6211\u4eec\u5c06\u901a\u8fc7 Segments.ai Python SDK \u5c06\u53d1\u5e03\u8f6c\u6362\u4e3a\u3002\u5982\u679c\u4f60\u5c1a\u672a\u8bbe\u7f6e Segments Python \u5ba2\u6237\u7aef\uff0c\u8bf7\u6309\u7167\u7684\u201c\u5728 Segments.ai \u4e0a\u8bbe\u7f6e\u6807\u8bb0\u4efb\u52a1\u201d\u90e8\u5206\u4e2d\u7684\u8bf4\u660e\u8fdb\u884c\u64cd\u4f5c\u3002<\/p>\n
\u8bf7\u6ce8\u610f\uff0c\u8f6c\u6362\u53ef\u80fd\u9700\u8981\u4e00\u6bb5\u65f6\u95f4\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u6570\u636e\u96c6\u7684\u5927\u5c0f\u3002<\/p>\n
from segments.huggingface import release2dataset\n\nrelease = segments_client.get_release(dataset_identifier, release_name)\nhf_dataset = release2dataset(release)\n<\/code><\/pre>\n\u5982\u679c\u6211\u4eec\u68c0\u67e5\u65b0\u6570\u636e\u96c6\u7684\u7279\u5f81\uff0c\u53ef\u4ee5\u770b\u5230\u56fe\u50cf\u5217\u548c\u76f8\u5e94\u7684\u6807\u7b7e\u3002\u6807\u7b7e\u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff1a\u6807\u6ce8\u5217\u8868\u548c\u5206\u5272\u4f4d\u56fe\u3002\u6807\u6ce8\u5bf9\u5e94\u4e8e\u56fe\u50cf\u4e2d\u7684\u4e0d\u540c\u5bf9\u8c61\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u5bf9\u8c61\uff0c\u6ce8\u91ca\u5305\u542b\u4e00\u4e2a id<\/code> \u548c\u4e00\u4e2a category_id<\/code>\u3002\u5206\u5272\u4f4d\u56fe\u662f\u4e00\u5e45\u56fe\u50cf\uff0c\u5176\u4e2d\u6bcf\u4e2a\u50cf\u7d20\u90fd\u5305\u542b\u8be5\u50cf\u7d20\u5904\u5bf9\u8c61\u7684 ID\u3002\u66f4\u591a\u4fe1\u606f\u53ef\u5728\u4e2d\u627e\u5230\u3002<\/p>\n\u5bf9\u4e8e\u8bed\u4e49\u5206\u5272\uff0c\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u8bed\u4e49\u4f4d\u56fe\uff0c\u5176\u4e2d\u5305\u542b\u6bcf\u4e2a\u50cf\u7d20\u7684 category_id<\/code>\u3002\u6211\u4eec\u5c06\u4f7f\u7528 Segments.ai SDK \u4e2d\u7684 get_semantic_bitmap<\/code> \u51fd\u6570\u5c06\u4f4d\u56fe\u8f6c\u6362\u4e3a\u8bed\u4e49\u4f4d\u56fe\u3002\u8981\u5c06\u6b64\u51fd\u6570\u5e94\u7528\u4e8e\u6570\u636e\u96c6\u4e2d\u7684\u6240\u6709\u884c\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 dataset.map<\/code>\u3002<\/p>\nfrom segments.utils import get_semantic_bitmap\n\ndef convert_segmentation_bitmap(example):\n return {\n \"label.segmentation_bitmap\":\n get_semantic_bitmap(\n example[\"label.segmentation_bitmap\"],\n example[\"label.annotations\"],\n id_increment=0,\n )\n }\n\n\nsemantic_dataset = hf_dataset.map(\n convert_segmentation_bitmap,\n)\n<\/code><\/pre>\n\u4f60\u8fd8\u53ef\u4ee5\u91cd\u5199 convert_segmentation_bitmap<\/code> \u51fd\u6570\u4ee5\u4f7f\u7528\u6279\u6b21\u5e76\u5c06 batched=True<\/code> \u4f20\u9012\u7ed9 dataset.map<\/code>\u3002\u8fd9\u5c06\u663e\u8457\u52a0\u5feb\u6620\u5c04\u901f\u5ea6\uff0c\u4f46\u4f60\u53ef\u80fd\u9700\u8981\u8c03\u6574 batch_size<\/code> \u4ee5\u786e\u4fdd\u8be5\u8fc7\u7a0b\u4e0d\u4f1a\u8017\u5c3d\u5185\u5b58\u3002<\/p>\n\u6211\u4eec\u7a0d\u540e\u8981\u5fae\u8c03\u7684 SegFormer \u6a21\u578b\u9700\u8981\u4e3a\u7279\u5f81\u6307\u5b9a\u7279\u5b9a\u540d\u79f0\u3002\u4e3a\u65b9\u4fbf\u8d77\u89c1\uff0c\u6211\u4eec\u73b0\u5728\u5c06\u5339\u914d\u6b64\u683c\u5f0f\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u56fe\u50cf\u7279\u5f81\u91cd\u547d\u540d\u4e3a pixel_values<\/code>\uff0c\u5c06 label.segmentation_bitmap<\/code> \u91cd\u547d\u540d\u4e3a label<\/code>\uff0c\u5e76\u4e22\u5f03\u5176\u4ed6\u7279\u5f81\u3002<\/p>\nsemantic_dataset = semantic_dataset.rename_column('image', 'pixel_values')\nsemantic_dataset = semantic_dataset.rename_column('label.segmentation_bitmap', 'label')\nsemantic_dataset = semantic_dataset.remove_columns(['name', 'uuid', 'status', 'label.annotations'])\n<\/code><\/pre>\n\u6211\u4eec\u73b0\u5728\u53ef\u4ee5\u5c06\u8f6c\u6362\u540e\u7684\u6570\u636e\u96c6\u63a8\u9001\u5230 Hugging Face Hub\u3002\u8fd9\u6837\uff0c\u4f60\u7684\u56e2\u961f\u548c Hugging Face \u793e\u533a\u5c31\u53ef\u4ee5\u4f7f\u7528\u5b83\u4e86\u3002\u5728\u4e0b\u4e00\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u4e86\u89e3\u5982\u4f55\u4ece Hub \u52a0\u8f7d\u6570\u636e\u96c6\u3002<\/p>\n
hf_dataset_identifier = f\"{hf_username}\/{dataset_name}\"\n\nsemantic_dataset.push_to_hub(hf_dataset_identifier)\n<\/code><\/pre>\n1.5 \u4f7f\u7528 Hub \u4e2d\u7684\u6570\u636e\u96c6<\/h3>\n
\u5982\u679c\u4f60\u4e0d\u60f3\u521b\u5efa\u81ea\u5df1\u7684\u6570\u636e\u96c6\uff0c\u4f46\u5728 Hugging Face Hub \u4e0a\u627e\u5230\u4e86\u9002\u5408\u4f60\u7528\u4f8b\u7684\u6570\u636e\u96c6\uff0c\u53ef\u4ee5\u5728\u6b64\u5904\u5b9a\u4e49\u6807\u8bc6\u7b26\u3002<\/p>\n
\u4f8b\u5982\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u5b8c\u6574\u6807\u6ce8\u7684\u4eba\u884c\u9053\u6570\u636e\u96c6\u3002\u8bf7\u6ce8\u610f\uff0c\u60a8\u53ef\u4ee5\u76f4\u63a5\u67e5\u770b\u793a\u4f8b\u3002<\/p>\n
hf_dataset_identifier = \"segments\/sidewalk-semantic\"\n<\/code><\/pre>\n2\u3001\u52a0\u8f7d\u5e76\u51c6\u5907 Hugging Face \u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3<\/h2>\n
\u73b0\u5728\u6211\u4eec\u5df2\u7ecf\u521b\u5efa\u4e86\u4e00\u4e2a\u65b0\u6570\u636e\u96c6\u5e76\u5c06\u5176\u63a8\u9001\u5230 Hugging Face Hub\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u4e00\u884c\u4e2d\u52a0\u8f7d\u6570\u636e\u96c6\u3002<\/p>\n
from datasets import load_dataset\n\nds = load_dataset(hf_dataset_identifier)\n<\/code><\/pre>\n\u8ba9\u6211\u4eec\u6253\u4e71\u6570\u636e\u96c6\uff0c\u5e76\u5c06\u6570\u636e\u96c6\u62c6\u5206\u4e3a\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\u3002<\/p>\n
ds = ds.shuffle(seed=1)\nds = ds[\"train\"].train_test_split(test_size=0.2)\ntrain_ds = ds[\"train\"]\ntest_ds = ds[\"test\"]\n<\/code><\/pre>\n\u6211\u4eec\u5c06\u63d0\u53d6\u6807\u7b7e\u6570\u91cf\u548c\u4eba\u7c7b\u53ef\u8bfb\u7684 ID\uff0c\u4ee5\u4fbf\u6211\u4eec\u7a0d\u540e\u53ef\u4ee5\u6b63\u786e\u914d\u7f6e\u5206\u5272\u6a21\u578b\u3002<\/p>\n
import json\nfrom huggingface_hub import hf_hub_download\n\nrepo_id = f\"datasets\/{hf_dataset_identifier}\"\nfilename = \"id2label.json\"\nid2label = json.load(open(hf_hub_download(repo_id=hf_dataset_identifier, filename=filename, repo_type=\"dataset\"), \"r\"))\nid2label = {int(k): v for k, v in id2label.items()}\nlabel2id = {v: k for k, v in id2label.items()}\n\nnum_labels = len(id2label)\n<\/code><\/pre>\n2.1 \u56fe\u50cf\u5904\u7406\u5668\u548c\u6570\u636e\u589e\u5f3a<\/h3>\n
SegFormer \u6a21\u578b\u8981\u6c42\u8f93\u5165\u5177\u6709\u7279\u5b9a\u5f62\u72b6\u3002\u4e3a\u4e86\u8f6c\u6362\u6211\u4eec\u7684\u8bad\u7ec3\u6570\u636e\u4ee5\u5339\u914d\u9884\u671f\u5f62\u72b6\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528 SegFormerImageProcessor<\/code>\u3002\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528 ds.map<\/code> \u51fd\u6570\u63d0\u524d\u5c06\u56fe\u50cf\u5904\u7406\u5668\u5e94\u7528\u4e8e\u6574\u4e2a\u8bad\u7ec3\u6570\u636e\u96c6\uff0c\u4f46\u8fd9\u4f1a\u5360\u7528\u5927\u91cf\u78c1\u76d8\u7a7a\u95f4\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u8f6c\u6362\uff0c\u5b83\u53ea\u4f1a\u5728\u5b9e\u9645\u4f7f\u7528\u6570\u636e\u65f6\uff08\u5373\u65f6\uff09\u51c6\u5907\u4e00\u6279\u6570\u636e\u3002\u8fd9\u6837\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5f00\u59cb\u8bad\u7ec3\uff0c\u800c\u65e0\u9700\u7b49\u5f85\u8fdb\u4e00\u6b65\u7684\u6570\u636e\u9884\u5904\u7406\u3002<\/p>\n\u5728\u6211\u4eec\u7684\u8f6c\u6362\u4e2d\uff0c\u6211\u4eec\u8fd8\u5c06\u5b9a\u4e49\u4e00\u4e9b\u6570\u636e\u589e\u5f3a\uff0c\u4ee5\u4f7f\u6211\u4eec\u7684\u6a21\u578b\u66f4\u80fd\u9002\u5e94\u4e0d\u540c\u7684\u5149\u7167\u6761\u4ef6\u3002\u6211\u4eec\u5c06\u4f7f\u7528 torchvision \u7684 ColorJitter<\/code> \u51fd\u6570\u968f\u673a\u66f4\u6539\u6279\u6b21\u4e2d\u56fe\u50cf\u7684\u4eae\u5ea6\u3001\u5bf9\u6bd4\u5ea6\u3001\u9971\u548c\u5ea6\u548c\u8272\u8c03\u3002<\/p>\nfrom torchvision.transforms import ColorJitter\nfrom transformers import SegformerImageProcessor\n\nprocessor = SegformerImageProcessor()\njitter = ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.1) \n\ndef train_transforms(example_batch):\n images = [jitter(x) for x in example_batch['pixel_values']]\n labels = [x for x in example_batch['label']]\n inputs = processor(images, labels)\n return inputs\n\n\ndef val_transforms(example_batch):\n images = [x for x in example_batch['pixel_values']]\n labels = [x for x in example_batch['label']]\n inputs = processor(images, labels)\n return inputs\n\n\n# Set transforms\ntrain_ds.set_transform(train_transforms)\ntest_ds.set_transform(val_transforms)\n<\/code><\/pre>\n3\u3001\u5fae\u8c03 SegFormer \u6a21\u578b<\/h2>\n
SegFormer \u4f5c\u8005\u5b9a\u4e49\u4e86 5 \u4e2a\u5927\u5c0f\u9010\u6e10\u589e\u5927\u7684\u6a21\u578b\uff1aB0 \u5230 B5\u3002\u4e0b\u56fe\uff08\u53d6\u81ea\u539f\u59cb\u8bba\u6587\uff09\u663e\u793a\u4e86\u8fd9\u4e9b\u4e0d\u540c\u6a21\u578b\u5728 ADE20K \u6570\u636e\u96c6\u4e0a\u4e0e\u5176\u4ed6\u6a21\u578b\u76f8\u6bd4\u7684\u6027\u80fd\u3002<\/p>\n
SegFormer \u6a21\u578b\u53d8\u4f53\u4e0e\u5176\u4ed6\u5206\u5272\u6a21\u578b\u7684\u6bd4\u8f83 – <\/p>\n
\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u5c06\u52a0\u8f7d\u6700\u5c0f\u7684 SegFormer \u6a21\u578b (B0)\uff0c\u8be5\u6a21\u578b\u5728 ImageNet-1k \u4e0a\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\u3002\u5b83\u53ea\u6709\u5927\u7ea6 14MB \u7684\u5927\u5c0f\uff01\u4f7f\u7528\u5c0f\u6a21\u578b\u5c06\u786e\u4fdd\u6211\u4eec\u7684\u6a21\u578b\u53ef\u4ee5\u5728\u6211\u4eec\u7684\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u4e0a\u987a\u5229\u8fd0\u884c\u3002<\/p>\n
from transformers import SegformerForSemanticSegmentation\n\npretrained_model_name = \"nvidia\/mit-b0\" \nmodel = SegformerForSemanticSegmentation.from_pretrained(\n pretrained_model_name,\n id2label=id2label,\n label2id=label2id\n)\n<\/code><\/pre>\n3.1 \u8bbe\u7f6e Trainer<\/h2>\n
\u4e3a\u4e86\u5728\u6211\u4eec\u7684\u6570\u636e\u4e0a\u5fae\u8c03\u6a21\u578b\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 Hugging Face \u7684 \u3002\u6211\u4eec\u9700\u8981\u8bbe\u7f6e\u8bad\u7ec3\u914d\u7f6e\u548c\u8bc4\u4f30\u6307\u6807\u4ee5\u4f7f\u7528 Trainer\u3002<\/p>\n
\u9996\u5148\uff0c\u6211\u4eec\u5c06\u8bbe\u7f6e \u3002\u8fd9\u5b9a\u4e49\u4e86\u6240\u6709\u8bad\u7ec3\u8d85\u53c2\u6570\uff0c\u4f8b\u5982\u5b66\u4e60\u7387\u548c epoch \u6570\u3001\u4fdd\u5b58\u6a21\u578b\u7684\u9891\u7387\u7b49\u3002\u6211\u4eec\u8fd8\u6307\u5b9a\u5728\u8bad\u7ec3\u540e\u5c06\u6a21\u578b\u63a8\u9001\u5230 hub\uff08 push_to_hub=True<\/code>\uff09\u5e76\u6307\u5b9a\u6a21\u578b\u540d\u79f0\uff08 hub_model_id<\/code>\uff09\u3002<\/p>\nfrom transformers import TrainingArguments\n\nepochs = 50\nlr = 0.00006\nbatch_size = 2\n\nhub_model_id = \"segformer-b0-finetuned-segments-sidewalk-2\"\n\ntraining_args = TrainingArguments(\n \"segformer-b0-finetuned-segments-sidewalk-outputs\",\n learning_rate=lr,\n num_train_epochs=epochs,\n per_device_train_batch_size=batch_size,\n per_device_eval_batch_size=batch_size,\n save_total_limit=3,\n evaluation_strategy=\"steps\",\n save_strategy=\"steps\",\n save_steps=20,\n eval_steps=20,\n logging_steps=1,\n eval_accumulation_steps=5,\n load_best_model_at_end=True,\n push_to_hub=True,\n hub_model_id=hub_model_id,\n hub_strategy=\"end\",\n)\n<\/code><\/pre>\n\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u5b9a\u4e49\u4e00\u4e2a\u51fd\u6570\u6765\u8ba1\u7b97\u6211\u4eec\u60f3\u8981\u4f7f\u7528\u7684\u8bc4\u4f30\u6307\u6807\u3002\u56e0\u4e3a\u6211\u4eec\u6b63\u5728\u8fdb\u884c\u8bed\u4e49\u5206\u5272\uff0c\u6240\u4ee5\u6211\u4eec\u5c06\u4f7f\u7528\uff0c\u53ef\u76f4\u63a5\u5728\u4e2d\u8bbf\u95ee\u3002IoU \u8868\u793a\u5206\u5272\u63a9\u7801\u7684\u91cd\u53e0\u3002\u5e73\u5747 IoU \u662f\u6240\u6709\u8bed\u4e49\u7c7b\u7684 IoU \u7684\u5e73\u5747\u503c\u3002\u8bf7\u67e5\u770b\u4ee5\u4e86\u89e3\u56fe\u50cf\u5206\u5272\u8bc4\u4f30\u6307\u6807\u7684\u6982\u8ff0\u3002<\/p>\n
\u7531\u4e8e\u6211\u4eec\u7684\u6a21\u578b\u8f93\u51fa\u5c3a\u5bf8\u4e3a\u9ad8\u5ea6\/4 \u548c\u5bbd\u5ea6\/4 \u7684 logits\uff0c\u56e0\u6b64\u6211\u4eec\u5fc5\u987b\u5bf9\u5176\u8fdb\u884c\u5347\u7ea7\uff0c\u7136\u540e\u624d\u80fd\u8ba1\u7b97 mIoU\u3002<\/p>\n
import torch\nfrom torch import nn\nimport evaluate\n\nmetric = evaluate.load(\"mean_iou\")\n\ndef compute_metrics(eval_pred):\n with torch.no_grad():\n logits, labels = eval_pred\n logits_tensor = torch.from_numpy(logits)\n # scale the logits to the size of the label\n logits_tensor = nn.functional.interpolate(\n logits_tensor,\n size=labels.shape[-2:],\n mode=\"bilinear\",\n align_corners=False,\n ).argmax(dim=1)\n\n pred_labels = logits_tensor.detach().cpu().numpy()\n metrics = metric.compute(\n predictions=pred_labels,\n references=labels,\n num_labels=len(id2label),\n ignore_index=0,\n reduce_labels=processor.do_reduce_labels,\n )\n \n # add per category metrics as individual key-value pairs\n per_category_accuracy = metrics.pop(\"per_category_accuracy\").tolist()\n per_category_iou = metrics.pop(\"per_category_iou\").tolist()\n\n metrics.update({f\"accuracy_{id2label[i]}\": v for i, v in enumerate(per_category_accuracy)})\n metrics.update({f\"iou_{id2label[i]}\": v for i, v in enumerate(per_category_iou)})\n \n return metrics\n<\/code><\/pre>\n\u6700\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u5b9e\u4f8b\u5316\u4e00\u4e2a Trainer<\/code> \u5bf9\u8c61\u3002<\/p>\nfrom transformers import Trainer\n\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=train_ds,\n eval_dataset=test_ds,\n compute_metrics=compute_metrics,\n)\n<\/code><\/pre>\n\u73b0\u5728\u6211\u4eec\u7684\u8bad\u7ec3\u5668\u5df2\u7ecf\u8bbe\u7f6e\u597d\u4e86\uff0c\u8bad\u7ec3\u5c31\u50cf\u8c03\u7528 train \u51fd\u6570\u4e00\u6837\u7b80\u5355\u3002\u6211\u4eec\u4e0d\u9700\u8981\u62c5\u5fc3\u7ba1\u7406\u6211\u4eec\u7684 GPU\uff0c\u8bad\u7ec3\u5668\u4f1a\u5904\u7406\u597d\u8fd9\u4e9b\u3002<\/p>\n
trainer.train()\n<\/code><\/pre>\n\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06\u7ecf\u8fc7\u5fae\u8c03\u7684\u6a21\u578b\u548c\u56fe\u50cf\u5904\u7406\u5668\u63a8\u9001\u5230 Hub\u3002<\/p>\n
\u8fd9\u8fd8\u5c06\u81ea\u52a8\u521b\u5efa\u4e00\u4e2a\u5305\u542b\u7ed3\u679c\u7684\u6a21\u578b\u5361\u3002\u6211\u4eec\u5c06\u5728 kwargs<\/code> \u4e2d\u63d0\u4f9b\u4e00\u4e9b\u989d\u5916\u4fe1\u606f\u6765\u8ba9\u6a21\u578b\u5361\u66f4\u5b8c\u6210\u3002<\/p>\nkwargs = {\n \"tags\": [\"vision\", \"image-segmentation\"],\n \"finetuned_from\": pretrained_model_name,\n \"dataset\": hf_dataset_identifier,\n}\n\nprocessor.push_to_hub(hub_model_id)\ntrainer.push_to_hub(**kwargs)\n<\/code><\/pre>\n4\u3001\u63a8\u7406<\/h2>\n
\u73b0\u5728\u662f\u6fc0\u52a8\u4eba\u5fc3\u7684\u90e8\u5206\uff0c\u4f7f\u7528\u6211\u4eec\u7ecf\u8fc7\u5fae\u8c03\u7684\u6a21\u578b\uff01\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u5c55\u793a\u5982\u4f55\u4ece\u4e2d\u5fc3\u52a0\u8f7d\u6a21\u578b\u5e76\u5c06\u5176\u7528\u4e8e\u63a8\u7406\u3002<\/p>\n
\u4f46\u662f\uff0c\u4f60\u4e5f\u53ef\u4ee5\u76f4\u63a5\u5728 Hugging Face Hub \u4e0a\u8bd5\u7528\u4f60\u7684\u6a21\u578b\uff0c\u8fd9\u8981\u5f52\u529f\u4e8e\u652f\u6301\u7684\u9177\u70ab\u5c0f\u90e8\u4ef6\u3002\u5982\u679c\u4f60\u5728\u4e0a\u4e00\u6b65\u4e2d\u5c06\u6a21\u578b\u63a8\u9001\u5230 Hub\uff0c\u5e94\u8be5\u4f1a\u5728\u6a21\u578b\u9875\u9762\u4e0a\u770b\u5230\u4e00\u4e2a\u63a8\u7406\u5c0f\u90e8\u4ef6\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7\u5728\u6a21\u578b\u5361\u4e2d\u5b9a\u4e49\u793a\u4f8b\u56fe\u50cf URL \u5411\u5c0f\u90e8\u4ef6\u6dfb\u52a0\u9ed8\u8ba4\u793a\u4f8b\u3002\u8bf7\u53c2\u9605\u4f5c\u4e3a\u793a\u4f8b\u3002<\/p>\n
4.1 \u4f7f\u7528\u6765\u81ea Hub \u7684\u6a21\u578b<\/h3>\n
\u6211\u4eec\u9996\u5148\u4f7f\u7528 SegformerForSemanticSegmentation.from_pretrained()<\/code> \u4ece Hub \u52a0\u8f7d\u6a21\u578b\u3002<\/p>\nfrom transformers import SegformerImageProcessor, SegformerForSemanticSegmentation\n\nprocessor = SegformerImageProcessor.from_pretrained(\"nvidia\/segformer-b0-finetuned-ade-512-512\")\nmodel = SegformerForSemanticSegmentation.from_pretrained(f\"{hf_username}\/{hub_model_id}\")\n<\/code><\/pre>\n\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u4ece\u6d4b\u8bd5\u6570\u636e\u96c6\u52a0\u8f7d\u56fe\u50cf\u3002<\/p>\n
image = test_ds[0]['pixel_values']\ngt_seg = test_ds[0]['label']\nimage\n<\/code><\/pre>\n\u8981\u5206\u5272\u6b64\u6d4b\u8bd5\u56fe\u50cf\uff0c\u6211\u4eec\u9996\u5148\u9700\u8981\u4f7f\u7528\u56fe\u50cf\u5904\u7406\u5668\u51c6\u5907\u56fe\u50cf\u3002\u7136\u540e\u6211\u4eec\u5c06\u5176\u8f6c\u53d1\u7ed9\u6a21\u578b\u3002<\/p>\n
\u6211\u4eec\u8fd8\u9700\u8981\u8bb0\u4f4f\u5c06\u8f93\u51fa logits \u653e\u5927\u5230\u539f\u59cb\u56fe\u50cf\u5927\u5c0f\u3002\u4e3a\u4e86\u83b7\u5f97\u5b9e\u9645\u7684\u7c7b\u522b\u9884\u6d4b\uff0c\u6211\u4eec\u53ea\u9700\u5728 logits \u4e0a\u5e94\u7528 argmax\u3002<\/p>\n
from torch import nn\n\ninputs = processor(images=image, return_tensors=\"pt\")\noutputs = model(**inputs)\nlogits = outputs.logits # shape (batch_size, num_labels, height\/4, width\/4)\n\n# First, rescale logits to original image size\nupsampled_logits = nn.functional.interpolate(\n logits,\n size=image.size[::-1], # (height, width)\n mode='bilinear',\n align_corners=False\n)\n\n# Second, apply argmax on the class dimension\npred_seg = upsampled_logits.argmax(dim=1)[0]\n<\/code><\/pre>\n\u73b0\u5728\u662f\u65f6\u5019\u663e\u793a\u7ed3\u679c\u4e86\u3002\u6211\u4eec\u5c06\u7ed3\u679c\u663e\u793a\u5728\u771f\u5b9e\u8499\u7248\u65c1\u8fb9\u3002<\/p>\n
\u4f60\u89c9\u5f97\u5982\u4f55\uff1f\u4f60\u4f1a\u5e26\u7740\u8fd9\u4e9b\u5206\u5272\u4fe1\u606f\u8ba9\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u4e0a\u8def\u5417\uff1f<\/p>\n
\u7ed3\u679c\u53ef\u80fd\u8fd8\u4e0d\u5b8c\u7f8e\uff0c\u4f46\u6211\u4eec\u59cb\u7ec8\u53ef\u4ee5\u6269\u5c55\u6570\u636e\u96c6\u4ee5\u4f7f\u6a21\u578b\u66f4\u52a0\u7a33\u5065\u3002\u6211\u4eec\u73b0\u5728\u8fd8\u53ef\u4ee5\u8bad\u7ec3\u66f4\u5927\u7684 SegFormer \u6a21\u578b\uff0c\u770b\u770b\u5b83\u7684\u8868\u73b0\u5982\u4f55\u3002<\/p>\n
5\u3001\u7ed3\u675f\u8bed<\/h2>\n
\u5c31\u662f\u8fd9\u6837\uff01<\/p>\n
\u4f60\u73b0\u5728\u77e5\u9053\u5982\u4f55\u521b\u5efa\u81ea\u5df1\u7684\u56fe\u50cf\u5206\u5272\u6570\u636e\u96c6\u4ee5\u53ca\u5982\u4f55\u4f7f\u7528\u5b83\u6765\u5fae\u8c03\u8bed\u4e49\u5206\u5272\u6a21\u578b\u3002<\/p>\n
\n\n","protected":false},"excerpt":{"rendered":"
\u672c\u6307\u5357\u5c55\u793a\u4e86\u5982\u4f55\u5fae\u8c03 Segformer\uff0c\u8fd9\u662f\u4e00\u79cd\u6700\u5148\u8fdb\u7684\u8bed\u4e49\u5206\u5272\u6a21\u578b\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u4e3a\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u5efa\u7acb\u4e00\u4e2a\u6a21\u578b\uff0c\u8fd9\u6837\u5b83\u5c31\u53ef\u4ee5\u770b\u5230\u8981\u884c\u9a76\u7684\u65b9\u5411\u5e76\u8bc6\u522b\u969c\u788d\u7269 \ud83e\udd16\u3002 \u6211\u4eec\u5c06\u9996\u5148\u5728 Segments.ai \u4e0a\u6807\u8bb0\u4e00\u7ec4\u4eba\u884c\u9053\u56fe\u50cf\u3002\u7136\u540e\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 \ud83e\udd17 transformers \u5fae\u8c03\u9884\u5148\u8bad\u7ec3\u7684 SegFormer \u6a21\u578b\uff0ctransformers \u662f\u4e00\u4e2a\u5f00\u6e90\u5e93\uff0c\u63d0\u4f9b\u6700\u5148\u8fdb\u6a21\u578b\u7684\u6613\u4e8e\u4f7f\u7528\u7684\u5b9e\u73b0\u3002\u5728\u6b64\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u4f7f\u7528 Hugging Face Hub\uff0c\u8fd9\u662f\u6700\u5927\u7684\u5f00\u6e90\u6a21\u578b\u548c\u6570\u636e\u96c6\u76ee\u5f55\u3002 \u8bed\u4e49\u5206\u5272\u662f\u5bf9\u56fe\u50cf\u4e2d\u7684\u6bcf\u4e2a\u50cf\u7d20\u8fdb\u884c\u5206\u7c7b\u7684\u4efb\u52a1\u3002\u4f60\u53ef\u4ee5\u5c06\u5176\u89c6\u4e3a\u5bf9\u56fe\u50cf\u8fdb\u884c\u66f4\u7cbe\u786e\u5206\u7c7b\u7684\u65b9\u6cd5\u3002\u5b83\u5728\u533b\u5b66\u6210\u50cf\u548c\u81ea\u52a8\u9a7e\u9a76\u7b49\u9886\u57df\u6709\u5e7f\u6cdb\u7684\u7528\u4f8b\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u6211\u4eec\u7684\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u6765\u8bf4\uff0c\u91cd\u8981\u7684\u662f\u8981\u786e\u5207\u5730\u77e5\u9053\u4eba\u884c\u9053\u5728\u56fe\u50cf\u4e2d\u7684\u4f4d\u7f6e\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u662f\u5426\u6709\u4eba\u884c\u9053\u3002 \u56e0\u4e3a\u8bed\u4e49\u5206\u5272\u662f\u4e00\u79cd\u5206\u7c7b\uff0c\u6240\u4ee5\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\u548c\u8bed\u4e49\u5206\u5272\u7684\u7f51\u7edc\u67b6\u6784\u975e\u5e38\u76f8\u4f3c\u3002 2014 \u5e74\uff0cLong \u7b49\u4eba\u53d1\u8868\u4e86\u4e00\u7bc7\uff0c\u4f7f\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u8bed\u4e49\u5206\u5272\u3002 \u6700\u8fd1\uff0cTransformers \u5df2\u7528\u4e8e\u56fe\u50cf\u5206\u7c7b\uff08\u4f8b\u5982 ViT\uff09\uff0c\u73b0\u5728\u5b83\u4eec\u4e5f\u7528\u4e8e\u8bed\u4e49\u5206\u5272\uff0c\u8fdb\u4e00\u6b65\u63a8\u52a8\u4e86\u6700\u5148\u8fdb\u7684\u6280\u672f\u3002 \u662f Xie \u7b49\u4eba\u4e8e 2021 \u5e74\u5f15\u5165\u7684\u8bed\u4e49\u5206\u5272\u6a21\u578b\u3002 \u5b83\u6709\u4e00\u4e2a\u4e0d\u4f7f\u7528\u4f4d\u7f6e\u7f16\u7801\u7684\u5206\u5c42 Transformer \u7f16\u7801\u5668\uff08\u4e0e ViT \u76f8\u53cd\uff09\u548c\u4e00\u4e2a\u7b80\u5355\u7684\u591a\u5c42\u611f\u77e5\u5668\u89e3\u7801\u5668\u3002 SegFormer \u5728\u591a\u4e2a\u5e38\u89c1\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002 \u8ba9\u6211\u4eec\u770b\u770b\u6211\u4eec\u7684\u62ab\u8428\u9001\u8d27\u673a\u5668\u4eba\u5728\u4eba\u884c\u9053\u56fe\u50cf\u4e0a\u7684\u8868\u73b0\u5982\u4f55\u3002 \u62ab\u8428\u5916\u9001\u673a\u5668\u4eba\u5206\u5272\u573a\u666f \u8ba9\u6211\u4eec\u4ece\u5b89\u88c5\u5fc5\u8981\u7684\u4f9d\u8d56\u9879\u5f00\u59cb\u3002\u56e0\u4e3a\u6211\u4eec\u8981\u5c06\u6570\u636e\u96c6\u548c\u6a21\u578b\u63a8\u9001\u5230 Hugging Face Hub\uff0c\u6240\u4ee5\u6211\u4eec\u9700\u8981\u5b89\u88c5 Git LFS \u5e76\u767b\u5f55 Hugging Face\u3002 git-lfs \u7684\u5b89\u88c5\u5728\u60a8\u7684\u7cfb\u7edf\u4e0a\u53ef\u80fd\u6709\u6240\u4e0d\u540c\u3002\u8bf7\u6ce8\u610f\uff0cGoogle Colab \u5df2\u9884\u88c5 Git LFS\u3002 pip install -q […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53766","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53766","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53766"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53766\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53766"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53766"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53766"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}