{"id":4718,"date":"2022-07-11T15:55:41","date_gmt":"2022-07-11T07:55:41","guid":{"rendered":"http:\/\/139.9.1.231\/?p=4718"},"modified":"2022-09-20T10:48:52","modified_gmt":"2022-09-20T02:48:52","slug":"mlp-mixer-an-all-mlp-architecture-for-vision","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/07\/11\/mlp-mixer-an-all-mlp-architecture-for-vision\/","title":{"rendered":"Vision MLP\u7cfb\u5217–MLP-Mixer: An all-MLP Architecture for Vision"},"content":{"rendered":"\n

MLP-Mixer\u662fViT\u56e2\u961f\u7684\u53e6\u4e00\u4e2a\u7eafMLP\u67b6\u6784\u7684\u5c1d\u8bd5\u3002\u5982\u679cMLP-Mixer\u91cd\u65b0\u5f15\u9886CV\u9886\u57df\u4e3b\u6d41\u67b6\u6784\u7684\u8bdd\uff0c\u90a3\u4e48CV\u9886\u57df\u4e3b\u6d41\u67b6\u6784\u7684\u6f14\u53d8\u8fc7\u7a0b\u5c31\u662fMLP->CNN->Transformer->MLP? \u8981\u56de\u5230\u6700\u521d\u7684\u8d77\u70b9\u4e86\u5417\uff1f\uff1f\uff1f<\/strong>( Transformer<\/strong>\u79fb\u9664\u4e86\u6ce8\u610f\u529b\u4ee5\u540e\u5c31\u5269MLP\u4e86)<\/p>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u201d\u7eaf\u201cMLP\u7ed3\u6784\u7684\u89c6\u89c9\u67b6\u6784\u3002<\/p>\n\n\n\n

\u5148\u5c06\u8f93\u5165\u56fe\u7247\u62c6\u5206\u6210patches\uff0c\u7136\u540e\u901a\u8fc7Per-patch Fully-connected\u5c06\u6bcf\u4e2apatch\u8f6c\u6362\u6210feature embedding\uff0c\u7136\u540e\u9001\u5165N\u4e2aMixer Layer\uff0c\u6700\u540e\u901a\u8fc7Fully-connected\u8fdb\u884c\u5206\u7c7b\u3002<\/p>\n\n\n\n

Mixer\u5206\u4e3achannel-mixing MLP\u548ctoken-mixing MLP\u4e24\u7c7b\u3002channel-mixing MLP\u5141\u8bb8\u4e0d\u540c\u901a\u9053\u4e4b\u95f4\u8fdb\u884c\u4ea4\u6d41\uff1btoken-mixing MLP\u5141\u8bb8\u4e0d\u540c\u7a7a\u95f4\u4f4d\u7f6e(tokens)\u8fdb\u884c\u4ea4\u6d41\u3002\u8fd9\u4e24\u79cd\u7c7b\u578b\u7684layer\u662f\u4ea4\u66ff\u5806\u53e0\u7684\uff0c\u65b9\u4fbf\u652f\u6301\u4e24\u4e2a\u8f93\u5165\u7ef4\u5ea6\u7684\u4ea4\u6d41\u3002\u6bcf\u4e2aMLP\u7531\u4e24\u5c42fully-connected\u548c\u4e00\u4e2aGELU\u6784\u6210\u3002<\/p>\n\n\n\n

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

\u4ece\u4e0a\u56fe\u6211\u4eec\u53ef\u4ee5\u770b\u51fa\uff0cMLP -Mixer \u9996\u5148\u4f7f\u7528\u56fe\u7247\u5206\u6210\u5f88\u591a\u4e2a\u5c0f\u6b63\u65b9\u5f62\u7684patch,\u6bcf\u4e2apatch\u7684\u5927\u5c0f\u5b9a\u4e49\u4e3apatch_size\u3002\u8bba\u6587\u4e2d\u5b9e\u73b0\u8fd9\u4e00\u6b65\u9aa4\u4f7f\u7528\u7684\u662f\u524d\u9762\u63d0\u5230\u7684\u5377\u79ef\uff0c\u5377\u79ef\u6838\u7684\u5927\u5c0f\u548c\u6b65\u957f\u5747patch_size\u3002\u8bba\u6587\u4e2d\u7ed9\u7684\u53c2\u6570\uff0c\u4e5f\u662f2\u7684\u5e42\u3002
\u7f51\u7edc\u4e0d\u518d\u4f7f\u7528\u4f20\u7edf\u7684RELU\u6fc0\u6d3b\u51fd\u6570\uff0c\u800c\u662f\u4f7f\u7528\u4e86GELU<\/a>\u6fc0\u6d3b\u51fd\u6570\u3002
<\/p>\n\n\n\n

\u5c06\u56fe\u7247\u5206\u6210\u5c0f\u5757\u540e\uff0c\u5728\u5c06\u5b83\u8f6c\u6362\u4e3a\u4e00\u7ef4\u7ed3\u6784\u3002\u5982\u56fe\uff1a<\/p>\n\n\n\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n

\u7136\u540e\u5c06\u6bcf\u4e00\u4e2apatch\u8fdb\u884c\u8f6c\u6362\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"\/<\/figure>\n\n\n\n

\u901a\u8fc7\u8fd9\u6837\u4e00\u79cd\u65b9\u5f0f\u5462\uff0c\u5c31\u5c06\u4e00\u5f20\u56fe\u7247\u8f6c\u6362\u4e3a\u4e86\u4e00\u4e2a\u5927\u77e9\u9635\uff0c\u5c31\u53ef\u4ee5\u8f93\u5165\u5230Mixer Layer \u4e2d\u8fdb\u884c\u8ba1\u7b97\u3002<\/p>\n\n\n\n

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

MLP \u662f\u4e24\u4e2a\u5168\u8fde\u63a5\u5c42\u7684\u611f\u77e5\u673a,W1,W2,\u5bf9\u5e94token_mixer\u4e2d\u4e24\u4e2a\u5168\u8fde\u63a5\u7684\u6743\u91cd\uff0cW3,W4\u5219\u8868\u793achannel_mixer\u4e24\u4e2a\u5168\u8fde\u63a5\u7684\u6743\u91cd\u3002\u03c3\u8868\u793aGELU\u6fc0\u6d3b\u51fd\u6570\u3002\u90a3\u4e48\u516c\u793a\u5c31\u5f88\u7b80\u5355\u4e86\uff0c\u8f93\u5165X\u7ecf\u8fc7Layer Normalize,\u518d\u4e58\u4ee5W1\uff0c\u518d\u7ecf\u8fc7\u6fc0\u6d3b\u51fd\u6570\u540e\u4e58\u4ee5W2\uff0c\u518d\u52a0\u4e0aX\u3002\u7b2c\u4e8c\u4e2a\u516c\u5f0f\u4e5f\u662f\u76f8\u540c\u7684\u8ba1\u7b97\u8fc7\u7a0b\u3002
\u5c06\u524d\u9762\u901a\u8fc7\u7f16\u7801\u5f97\u5230\u7684\u77e9\u9635\u7ecf\u8fc7Layer Norm \u5728\u5c06\u77e9\u9635\u8fdb\u884c\u65cb\u8f6c\uff08T \u8868\u793a\u65cb\u8f6c\uff09\u8fde\u63a5MLP1,MLP1 \u5c31\u662f\u6587\u7ae0token_mixer \u7528\u6765\u5bfb\u627e\u50cf\u7d20\u4e0e\u50cf\u7d20\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u5176\u4e2d\uff0cMLP1\u4e2d\u7684\u6743\u503c\u5171\u4eab\u3002\u8ba1\u7b97\u5b8c\u4e4b\u540e\uff0c\u518d\u5c06\u77e9\u9635\u65cb\u8f6c\u56de\u6765\uff0c\u901a\u8fc7Layer Norm \u540e\u518d\u63a5\u4e00\u4e2achannel_mixer \u7528\u4e8e\u5bfb\u627e\u901a\u9053\u4e0e\u901a\u9053\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u5176\u4e2dMixerLayer \u8fd8\u542f\u7528\u4e86ResNet\u4e2d\u7684\u8de8\u8fde\u7ed3\u6784\uff0c\u8de8\u8fde\u7ed3\u6784\u7684\u4f5c\u7528\u53ef\u4ee5\u53c2\u8003[ResNet\u539f\u7406\u8bb2\u89e3\u548c\u590d\u73b0]<\/a>\uff0c\u770b\u5230\u8fd9\u91cc\uff0c\u662f\u4e0d\u662f\u611f\u89c9\u5b83\u8ddf\u5377\u79ef\u7684\u539f\u7406\u5f88\u7c7b\u4f3c\u3002
\u4ece\u4e0a\u56fe\u53ef\u4ee5\u770b\u51faMixer Layer\u7684\u8f93\u5165\u7ef4\u5ea6\u548c\u8f93\u51fa\u7ef4\u5ea6\u76f8\u540c\uff0c\u5e76\u4e14\u901a\u8fc7MLP\u7684\u65b9\u5f0f\u6765\u5bfb\u627e\u56fe\u7247\u50cf\u7d20\u4e0e\u50cf\u7d20\uff0c\u901a\u9053\u4e0e\u901a\u9053\u7684\u5173\u7cfb\u3002
\u8fd9\u5c31\u662fMLP-MIXER\u7684\u7f51\u7edc\u7ed3\u6784\u4e86<\/p>\n\n\n\n

\u5b9e\u73b0\u7684\u96be\u70b9\u5728\u4e8e\uff0c\u77e9\u9635\u65cb\u8f6c\uff0c\u6211\u4eec\u4f7f\u7528einops<\/strong>\u4e2d\u7684Rearrange\u5b9e\u73b0\u77e9\u9635\u65cb\u8f6c<\/p>\n\n\n\n

\u4f7f\u7528Rearrange \u5b9e\u73b0\u65cb\u8f6c<\/h1>\n\n\n\n

Rearrange(‘b n d -> b d n’) #\u8fd9\u91cc\u662f[batch_size, num_patch, dim] -> [batch_size, dim, num_patch]<\/p>\n\n\n\n

#\u5b9a\u4e49\u591a\u5c42\u611f\u77e5\u673a\nimport torch\nimport numpy as np\nfrom torch import nn\nfrom einops.layers.torch import Rearrange\nfrom torchsummary import summary\nimport torch.nn.functional as F\n\nclass FeedForward(nn.Module):\n    def __init__(self,dim,hidden_dim,dropout=0.):\n        super().__init__()\n        self.net=nn.Sequential(\n            #\u7531\u6b64\u53ef\u4ee5\u770b\u51fa FeedForward \u7684\u8f93\u5165\u548c\u8f93\u51fa\u7ef4\u5ea6\u662f\u4e00\u81f4\u7684\n            nn.Linear(dim,hidden_dim),\n            #\u6fc0\u6d3b\u51fd\u6570\n            nn.GELU(),\n            #\u9632\u6b62\u8fc7\u62df\u5408\n            nn.Dropout(dropout),\n            #\u91cd\u590d\u4e0a\u8ff0\u8fc7\u7a0b\n            nn.Linear(hidden_dim,dim),\n\n            nn.Dropout(dropout)\n        )\n    def forward(self,x):\n        x=self.net(x)\n        return x\n\n\nclass MixerBlock(nn.Module):\n    def __init__(self,dim,num_patch,token_dim,channel_dim,dropout=0.):\n        super().__init__()\n        self.token_mixer=nn.Sequential(\n            nn.LayerNorm(dim),\n            Rearrange('b n d -> b d n'),   #\u8fd9\u91cc\u662f[batch_size, num_patch, dim] -> [batch_size, dim, num_patch]\n            FeedForward(num_patch,token_dim,dropout),\n            Rearrange('b d n -> b n d')    #[batch_size, dim, num_patch] -> [batch_size, num_patch, dim]\n\n         )\n        self.channel_mixer=nn.Sequential(\n            nn.LayerNorm(dim),\n            FeedForward(dim,channel_dim,dropout)\n        )\n    def forward(self,x):\n\n        x=x+self.token_mixer(x)\n\n        x=x+self.channel_mixer(x)\n\n        return x\n\nclass MLPMixer(nn.Module):\n    def __init__(self,in_channels,dim,num_classes,patch_size,image_size,depth,token_dim,channel_dim,dropout=0.):\n        super().__init__()\n        assert image_size%patch_size==0\n        self.num_patches=(image_size\/\/patch_size)**2\n        #embedding \u64cd\u4f5c\uff0c\u7528\u5377\u79ef\u6765\u5206\u6210\u4e00\u5c0f\u5757\u4e00\u5c0f\u5757\u7684\n        self.to_embedding=nn.Sequential(nn.Conv2d(in_channels=in_channels,out_channels=dim,kernel_size=patch_size,stride=patch_size),\n            Rearrange('b c h w -> b (h w) c')\n        )\n        #\u7ecf\u8fc7Mixer Layer \u7684\u6b21\u6570\n        self.mixer_blocks=nn.ModuleList([])\n        for _ in range(depth):\n            self.mixer_blocks.append(MixerBlock(dim,self.num_patches,token_dim,channel_dim,dropout))\n        self.layer_normal=nn.LayerNorm(dim)\n\n        self.mlp_head=nn.Sequential(\n            nn.Linear(dim,num_classes)\n        )\n    def forward(self,x):\n        x=self.to_embedding(x)\n        for mixer_block in self.mixer_blocks:\n            x=mixer_block(x)\n        x=self.layer_normal(x)\n        x=x.mean(dim=1)\n\n        x=self.mlp_head(x)\n\n        return x\n<\/code><\/pre>\n\n\n\n
MLP-Mixer\u7528Mixer\u7684MLP\u6765\u66ff\u4ee3ViT\u7684Transformer\uff0c\u51cf\u5c11\u4e86\u7279\u5f81\u63d0\u53d6\u7684\u81ea\u7531\u5ea6\uff0c\u5e76\u4e14\u5de7\u5999\u7684\u53ef\u4ee5\u4ea4\u66ff\u8fdb\u884cpatch\u95f4\u4fe1\u606f\u4ea4\u6d41\u548cpatch\u5185\u4fe1\u606f\u4ea4\u6d41\uff0c\u4ece\u7ed3\u679c\u4e0a\u6765\u770b\uff0c\u7eafMLP\u8c8c\u4f3c\u4e5f\u662f\u53ef\u884c\u7684\uff0c\u800c\u4e14\u7701\u53bb\u4e86Transformer\u590d\u6742\u7684\u7ed3\u6784\uff0c\u53d8\u7684\u66f4\u52a0\u7b80\u6d01\uff0c\u6709\u70b9\u671f\u5f85\u540e\u7eedViT\u548cMLP-Mixer\u5982\u4f55\u9488\u950b\u76f8\u5bf9<\/strong>\u7684\uff0c\u611f\u89c9\u5927\u7ec4\u5c31\u662f\u4e1c\u6316\u4e00\u4e2a\u897f\u6316\u4e00\u4e2a\u7684\uff0c\u53c8\u628a\u5c18\u5c01\u591a\u5e74\u7684MLP\u7ed9\u6316\u51fa\u6765\u4e86<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"
MLP-Mixer\u662fViT\u56e2\u961f\u7684\u53e6\u4e00\u4e2a\u7eafMLP\u67b6\u6784\u7684\u5c1d\u8bd5\u3002\u5982\u679cMLP-Mixer\u91cd\u65b0\u5f15\u9886CV\u9886\u57df\u4e3b\u6d41\u67b6\u6784\u7684\u8bdd\uff0c … \u7ee7\u7eed\u9605\u8bfbVision MLP\u7cfb\u5217–MLP-Mixer: An all-MLP Architecture for Vision<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4718"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=4718"}],"version-history":[{"count":14,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4718\/revisions"}],"predecessor-version":[{"id":7955,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/4718\/revisions\/7955"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=4718"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=4718"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=4718"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}