{"id":4670,"date":"2022-07-08T15:34:49","date_gmt":"2022-07-08T07:34:49","guid":{"rendered":"http:\/\/139.9.1.231\/?p=4670"},"modified":"2022-07-08T15:34:51","modified_gmt":"2022-07-08T07:34:51","slug":"pytorch-image-models-timm","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/07\/08\/pytorch-image-models-timm\/","title":{"rendered":"Pytorch Image Models –timm\u5feb\u901f\u4f7f\u7528"},"content":{"rendered":"\n

\u539f\u6587\uff1aGetting Started with PyTorch Image Models (timm): A Practitioner\u2019s Guide – 2022.02.02<\/a><\/p>\n\n\n\n

\u4e2d\u6587\u6559\u7a0b: https:\/\/www.aiuai.cn\/aifarm1967.html<\/a><\/p>\n\n\n\n

Github\uff1a\u00a0rwightman\/pytorch-image-models<\/a><\/p>\n\n\n\n

PyTorch Image Models\uff08timm\uff09 \u662f\u4e00\u4e2a\u4f18\u79c0\u7684\u56fe\u50cf\u5206\u7c7b Python \u5e93\uff0c\u5176\u5305\u542b\u4e86\u5927\u91cf\u7684\u56fe\u50cf\u6a21\u578b\uff08Image Models\uff09\u3001Optimizers\u3001Schedulers\u3001Augmentations \u7b49\u7b49.\u91cc\u9762\u63d0\u4f9b\u4e86\u8bb8\u591a\u8ba1\u7b97\u673a\u89c6\u89c9\u7684SOTA\u6a21\u578b\uff0c\u53ef\u4ee5\u5f53\u4f5c\u662ftorchvision\u7684\u6269\u5145\u7248\u672c\uff0c\u5e76\u4e14\u91cc\u9762\u7684\u6a21\u578b\u5728\u51c6\u786e\u5ea6\u4e0a\u4e5f\u8f83\u9ad8\u3002<\/p>\n\n\n\n

timm \u63d0\u4f9b\u4e86\u53c2\u8003\u7684\u00a0training<\/a>\u00a0\u548c\u00a0validation<\/a>\u00a0\u811a\u672c\uff0c\u7528\u4e8e\u590d\u73b0\u5728\u00a0ImageNet<\/a>\u00a0\u4e0a\u7684\u8bad\u7ec3\u7ed3\u679c\uff1b\u4ee5\u53ca\u66f4\u591a\u7684\u00a0\u5b98\u65b9\u6587\u6863<\/a>\u00a0\u548c\u00a0timmdocs project<\/a>.<\/p>\n\n\n\n

timm\u7684\u5b89\u88c5<\/a><\/h2>\n\n\n\n

\u5173\u4e8etimm\u7684\u5b89\u88c5\uff0c\u6211\u4eec\u53ef\u4ee5\u9009\u62e9\u4ee5\u4e0b\u4e24\u79cd\u65b9\u5f0f\u8fdb\u884c\uff1a<\/p>\n\n\n\n

  1. \u901a\u8fc7pip\u5b89\u88c5<\/li><\/ol>\n\n\n\n
    pip install timm\n<\/pre>\n\n\n\n
    1. \u901a\u8fc7git\u4e0epip\u8fdb\u884c\u5b89\u88c5<\/li><\/ol>\n\n\n\n
      git clone https:\/\/github.com\/rwightman\/pytorch-image-models\ncd pytorch-image-models &&<\/strong> pip install -e .<\/pre>\n\n\n\n
      \u5982\u4f55\u67e5\u770b\u9884\u8bad\u7ec3\u6a21\u578b\u79cd\u7c7b<\/a><\/h2>\n\n\n\n
      1. \u67e5\u770btimm\u63d0\u4f9b\u7684\u9884\u8bad\u7ec3\u6a21\u578b \u622a\u6b62\u52302022.3.27\u65e5\u4e3a\u6b62\uff0ctimm\u63d0\u4f9b\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u5df2\u7ecf\u8fbe\u5230\u4e86592\u4e2a\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7timm.list_models()<\/code>\u65b9\u6cd5\u67e5\u770btimm\u63d0\u4f9b\u7684\u9884\u8bad\u7ec3\u6a21\u578b\uff08\u6ce8\uff1a\u672c\u7ae0\u6d4b\u8bd5\u4ee3\u7801\u5747\u662f\u5728jupyter notebook\u4e0a\u8fdb\u884c\uff09<\/li><\/ol>\n\n\n\n
        import<\/strong> timm\navail_pretrained_models =<\/strong> timm.<\/strong>list_models(<\/strong>pretrained=<\/strong>True<\/strong>)<\/strong>\nlen(<\/strong>avail_pretrained_models)<\/strong><\/pre>\n\n\n\n
        1. \u67e5\u770b\u7279\u5b9a\u6a21\u578b\u7684\u6240\u6709\u79cd\u7c7b \u6bcf\u4e00\u79cd\u7cfb\u5217\u53ef\u80fd\u5bf9\u5e94\u7740\u4e0d\u540c\u65b9\u6848\u7684\u6a21\u578b\uff0c\u6bd4\u5982Resnet\u7cfb\u5217\u5c31\u5305\u62ec\u4e86ResNet18\uff0c50\uff0c101\u7b49\u6a21\u578b\uff0c\u6211\u4eec\u53ef\u4ee5\u5728timm.list_models()<\/code>\u4f20\u5165\u60f3\u67e5\u8be2\u7684\u6a21\u578b\u540d\u79f0\uff08\u6a21\u7cca\u67e5\u8be2\uff09\uff0c\u6bd4\u5982\u6211\u4eec\u60f3\u67e5\u8be2densenet\u7cfb\u5217\u7684\u6240\u6709\u6a21\u578b\u3002<\/li><\/ol>\n\n\n\n
          all_densnet_models =<\/strong> timm.<\/strong>list_models(<\/strong>\"*densenet*\")<\/strong>\nall_densnet_models\n<\/pre>\n\n\n\n
          \u6211\u4eec\u53d1\u73b0\u4ee5\u5217\u8868\u7684\u5f62\u5f0f\u8fd4\u56de\u4e86\u6240\u6709densenet\u7cfb\u5217\u7684\u6240\u6709\u6a21\u578b\u3002<\/p>\n\n\n\n
          [<\/strong>'densenet121',\n 'densenet121d',\n 'densenet161',\n 'densenet169',\n 'densenet201',\n 'densenet264',\n 'densenet264d_iabn',\n 'densenetblur121d',\n 'tv_densenet121']<\/strong><\/pre>\n\n\n\n
          1. \u67e5\u770b\u6a21\u578b\u7684\u5177\u4f53\u53c2\u6570 \u5f53\u6211\u4eec\u60f3\u67e5\u770b\u4e0b\u6a21\u578b\u7684\u5177\u4f53\u53c2\u6570\u7684\u65f6\u5019\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u6a21\u578b\u7684default_cfg<\/code>\u5c5e\u6027\u6765\u8fdb\u884c\u67e5\u770b\uff0c\u5177\u4f53\u64cd\u4f5c\u5982\u4e0b<\/li><\/ol>\n\n\n\n
            model =<\/strong> timm.<\/strong>create_model(<\/strong>'resnet34',<\/strong>num_classes=<\/strong>10<\/strong>,<\/strong>pretrained=<\/strong>True<\/strong>)<\/strong>\nmodel.<\/strong>default_cfg\n<\/pre>\n\n\n\n
            {<\/strong>'url':<\/strong> 'https:\/\/github.com\/rwightman\/pytorch-image-models\/releases\/download\/v0.1-weights\/resnet34-43635321.pth',<\/strong>\n 'num_classes':<\/strong> 1000<\/strong>,<\/strong>\n 'input_size':<\/strong> (<\/strong>3<\/strong>,<\/strong> 224<\/strong>,<\/strong> 224<\/strong>),<\/strong>\n 'pool_size':<\/strong> (<\/strong>7<\/strong>,<\/strong> 7<\/strong>),<\/strong>\n 'crop_pct':<\/strong> 0.875<\/strong>,<\/strong>\n 'interpolation':<\/strong> 'bilinear',<\/strong>\n 'mean':<\/strong> (<\/strong>0.485<\/strong>,<\/strong> 0.456<\/strong>,<\/strong> 0.406<\/strong>),<\/strong>\n 'std':<\/strong> (<\/strong>0.229<\/strong>,<\/strong> 0.224<\/strong>,<\/strong> 0.225<\/strong>),<\/strong>\n 'first_conv':<\/strong> 'conv1',<\/strong>\n 'classifier':<\/strong> 'fc',<\/strong>\n 'architecture':<\/strong> 'resnet34'}<\/strong>\n<\/pre>\n\n\n\n
            \u9664\u6b64\u4e4b\u5916\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u8fd9\u4e2a\u94fe\u63a5<\/a>\u00a0\u67e5\u770b\u63d0\u4f9b\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u51c6\u786e\u5ea6\u7b49\u4fe1\u606f\u3002<\/p>\n\n\n\n
            \u4f7f\u7528\u548c\u4fee\u6539\u9884\u8bad\u7ec3\u6a21\u578b<\/a><\/h2>\n\n\n\n
            \u5728\u5f97\u5230\u6211\u4eec\u60f3\u8981\u4f7f\u7528\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7timm.create_model()<\/code>\u7684\u65b9\u6cd5\u6765\u8fdb\u884c\u6a21\u578b\u7684\u521b\u5efa\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u4f20\u5165\u53c2\u6570pretrained=True<\/code>\uff0c\u6765\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u540c\u6837\u7684\uff0c\u6211\u4eec\u4e5f\u53ef\u4ee5\u4f7f\u7528\u8ddftorchvision\u91cc\u9762\u7684\u6a21\u578b\u4e00\u6837\u7684\u65b9\u6cd5\u67e5\u770b\u6a21\u578b\u7684\u53c2\u6570\uff0c\u7c7b\u578b\/<\/p>\n\n\n\n
            import<\/strong> timm\nimport<\/strong> torch\n\nmodel =<\/strong> timm.<\/strong>create_model(<\/strong>'resnet34',<\/strong>pretrained=<\/strong>True<\/strong>)<\/strong>\nx =<\/strong> torch.<\/strong>randn(<\/strong>1<\/strong>,<\/strong>3<\/strong>,<\/strong>224<\/strong>,<\/strong>224<\/strong>)<\/strong>\noutput =<\/strong> model(<\/strong>x)<\/strong>\noutput.<\/strong>shape\n<\/pre>\n\n\n\n
            torch.Size([<\/strong>1<\/strong>, 1000<\/strong>])<\/strong>\n<\/pre>\n\n\n\n