{"id":1815,"date":"2025-02-17T15:30:04","date_gmt":"2025-02-17T07:30:04","guid":{"rendered":"https:\/\/www.forillusion.com\/?p=1815"},"modified":"2025-02-17T15:30:04","modified_gmt":"2025-02-17T07:30:04","slug":"4-4-custom-layer","status":"publish","type":"post","link":"https:\/\/www.forillusion.com\/index.php\/4-4-custom-layer\/","title":{"rendered":"4.4 \u81ea\u5b9a\u4e49\u5c42"},"content":{"rendered":"\n

<\/div><\/p>\n\n\n\n

\u4e0d\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42<\/h2>\n\n\n\n

\u5148\u4ecb\u7ecd\u5982\u4f55\u5b9a\u4e49\u4e00\u4e2a\u4e0d\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42\u3002\u4e8b\u5b9e\u4e0a\uff0c\u8fd9\u548c4.1\u8282\uff08\u6a21\u578b\u6784\u9020\uff09\u4e2d\u4ecb\u7ecd\u7684\u4f7f\u7528Module<\/code>\u7c7b\u6784\u9020\u6a21\u578b\u7c7b\u4f3c\u3002\u4e0b\u9762\u7684CenteredLayer<\/code>\u7c7b\u901a\u8fc7\u7ee7\u627fModule<\/code>\u7c7b\u81ea\u5b9a\u4e49\u4e86\u4e00\u4e2a\u5c06\u8f93\u5165\u51cf\u6389\u5747\u503c\u540e\u8f93\u51fa\u7684\u5c42\uff0c\u5e76\u5c06\u5c42\u7684\u8ba1\u7b97\u5b9a\u4e49\u5728\u4e86forward<\/code>\u51fd\u6570\u91cc\u3002\u8fd9\u4e2a\u5c42\u91cc\u4e0d\u542b\u6a21\u578b\u53c2\u6570\u3002<\/p>\n\n\n\n

import torch\nfrom torch import nn\n\nclass CenteredLayer(nn.Module):\n    def __init__(self, **kwargs):\n        super(CenteredLayer, self).__init__(**kwargs)\n    def forward(self, x):\n        return x - x.mean() # x.mean()\u6c42\u5747\u503c<\/code><\/pre>\n\n\n\n
\u53ef\u4ee5\u5b9e\u4f8b\u5316\u8fd9\u4e2a\u5c42\uff0c\u7136\u540e\u505a\u524d\u5411\u8ba1\u7b97\u3002<\/p>\n\n\n\n
layer = CenteredLayer()\nlayer(torch.tensor([1, 2, 3, 4, 5], dtype=torch.float))<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
tensor([-2., -1.,  0.,  1.,  2.])<\/code><\/pre>\n\n\n\n
\u4e5f\u53ef\u4ee5\u7528\u5b83\u6765\u6784\u9020\u66f4\u590d\u6742\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n
net = nn.Sequential(nn.Linear(8, 128), CenteredLayer())<\/code><\/pre>\n\n\n\n
\u4e0b\u9762\u6253\u5370\u81ea\u5b9a\u4e49\u5c42\u5404\u4e2a\u8f93\u51fa\u7684\u5747\u503c\u3002\u56e0\u4e3a\u5747\u503c\u662f\u6d6e\u70b9\u6570\uff0c\u6240\u4ee5\u5b83\u7684\u503c\u662f\u4e00\u4e2a\u5f88\u63a5\u8fd10\u7684\u6570\u3002<\/p>\n\n\n\n
y = net(torch.rand(4, 8))\ny.mean().item()<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
0.0<\/code><\/pre>\n\n\n\n
\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42<\/h2>\n\n\n\n
\u8fd8\u53ef\u4ee5\u81ea\u5b9a\u4e49\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42\u3002\u5176\u4e2d\u7684\u6a21\u578b\u53c2\u6570\u53ef\u4ee5\u901a\u8fc7\u8bad\u7ec3\u5b66\u51fa\u3002<\/p>\n\n\n\n
\u57284.2\u8282\uff08\u6a21\u578b\u53c2\u6570\u7684\u8bbf\u95ee\u3001\u521d\u59cb\u5316\u548c\u5171\u4eab\uff09\u4e2d\u4ecb\u7ecd\u4e86Parameter<\/code>\u7c7b\u5176\u5b9e\u662fTensor<\/code>\u7684\u5b50\u7c7b\uff0c\u5982\u679c\u4e00\u4e2aTensor<\/code>\u662fParameter<\/code>\uff0c\u90a3\u4e48\u5b83\u4f1a\u81ea\u52a8\u88ab\u6dfb\u52a0\u5230\u6a21\u578b\u7684\u53c2\u6570\u5217\u8868\u91cc\u3002\u6240\u4ee5\u5728\u81ea\u5b9a\u4e49\u542b\u6a21\u578b\u53c2\u6570\u7684\u5c42\u65f6\uff0c\u5e94\u8be5\u5c06\u53c2\u6570\u5b9a\u4e49\u6210Parameter<\/code>\uff0c\u9664\u4e86\u50cf4.2.1\u8282\u90a3\u6837\u76f4\u63a5\u5b9a\u4e49\u6210Parameter<\/code>\u7c7b\u5916\uff0c\u8fd8\u53ef\u4ee5\u4f7f\u7528ParameterList<\/code>\u548cParameterDict<\/code>\u5206\u522b\u5b9a\u4e49\u53c2\u6570\u7684\u5217\u8868\u548c\u5b57\u5178\u3002<\/p>\n\n\n\n
ParameterList<\/code>\u63a5\u6536\u4e00\u4e2aParameter<\/code>\u5b9e\u4f8b\u7684\u5217\u8868\u4f5c\u4e3a\u8f93\u5165\u7136\u540e\u5f97\u5230\u4e00\u4e2a\u53c2\u6570\u5217\u8868\uff0c\u4f7f\u7528\u7684\u65f6\u5019\u53ef\u4ee5\u7528\u7d22\u5f15\u6765\u8bbf\u95ee\u67d0\u4e2a\u53c2\u6570\uff0c\u53e6\u5916\u4e5f\u53ef\u4ee5\u4f7f\u7528append<\/code>\u548cextend<\/code>\u5728\u5217\u8868\u540e\u9762\u65b0\u589e\u53c2\u6570\u3002<\/p>\n\n\n\n
class MyDense(nn.Module):\n    def __init__(self):\n        super(MyDense, self).__init__()\n        self.params = nn.ParameterList([nn.Parameter(torch.randn(4, 4)) for i in range(3)]) # for\u5faa\u73af\u521b\u5efa3\u4e2a\u53c2\u6570,\u6bcf\u4e2a\u53c2\u6570\u662f4x4\u7684\u77e9\u9635\n        self.params.append(nn.Parameter(torch.randn(4, 1))) # \u6dfb\u52a0\u4e00\u4e2a\u53c2\u6570,\u662f4x1\u7684\u77e9\u9635\n\n    def forward(self, x):\n        for i in range(len(self.params)):\n            x = torch.mm(x, self.params[i])\n        return x\nnet = MyDense()\nprint(net)<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
MyDense(\n  (params): ParameterList(\n      (0): Parameter containing: [torch.FloatTensor of size 4x4]\n      (1): Parameter containing: [torch.FloatTensor of size 4x4]\n      (2): Parameter containing: [torch.FloatTensor of size 4x4]\n      (3): Parameter containing: [torch.FloatTensor of size 4x1]\n  )\n)<\/code><\/pre>\n\n\n\n
\u800cParameterDict<\/code>\u63a5\u6536\u4e00\u4e2aParameter<\/code>\u5b9e\u4f8b\u7684\u5b57\u5178\u4f5c\u4e3a\u8f93\u5165\u7136\u540e\u5f97\u5230\u4e00\u4e2a\u53c2\u6570\u5b57\u5178\uff0c\u7136\u540e\u53ef\u4ee5\u6309\u7167\u5b57\u5178\u7684\u89c4\u5219\u4f7f\u7528\u4e86\u3002\u4f8b\u5982\u4f7f\u7528update()<\/code>\u65b0\u589e\u53c2\u6570\uff0c\u4f7f\u7528keys()<\/code>\u8fd4\u56de\u6240\u6709\u952e\u503c\uff0c\u4f7f\u7528items()<\/code>\u8fd4\u56de\u6240\u6709\u952e\u503c\u5bf9\u7b49\u7b49\uff0c\u53ef\u53c2\u8003\u5b98\u65b9\u6587\u6863<\/a>\u3002<\/p>\n\n\n\n
class MyDictDense(nn.Module):\n    def __init__(self):\n        super(MyDictDense, self).__init__()\n        self.params = nn.ParameterDict({\n                'linear1': nn.Parameter(torch.randn(4, 4)),\n                'linear2': nn.Parameter(torch.randn(4, 1))\n        })\n        self.params.update({'linear3': nn.Parameter(torch.randn(4, 2))}) # \u65b0\u589e\n\n    def forward(self, x, choice='linear1'):\n        return torch.mm(x, self.params[choice]) # \u8fd4\u56dex\u548cchoice\u5bf9\u5e94\u7684\u53c2\u6570\u7684\u77e9\u9635\u4e58\u6cd5\n\nnet = MyDictDense()\nprint(net)<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
MyDictDense(\n  (params): ParameterDict(\n      (linear1): Parameter containing: [torch.FloatTensor of size 4x4]\n      (linear2): Parameter containing: [torch.FloatTensor of size 4x1]\n      (linear3): Parameter containing: [torch.FloatTensor of size 4x2]\n  )\n)<\/code><\/pre>\n\n\n\n
\u8fd9\u6837\u5c31\u53ef\u4ee5\u6839\u636e\u4f20\u5165\u7684\u952e\u503c\u6765\u8fdb\u884c\u4e0d\u540c\u7684\u524d\u5411\u4f20\u64ad\uff1a<\/p>\n\n\n\n
x = torch.ones(1, 4)\nprint(net(x, 'linear1'))\nprint(net(x, 'linear2'))\nprint(net(x, 'linear3'))<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
tensor([[1.5082, 1.5574, 2.1651, 1.2409]], grad_fn=<MmBackward>)\ntensor([[-0.8783]], grad_fn=<MmBackward>)\ntensor([[ 2.2193, -1.6539]], grad_fn=<MmBackward>)<\/code><\/pre>\n\n\n\n
\u4e5f\u53ef\u4ee5\u4f7f\u7528\u81ea\u5b9a\u4e49\u5c42\u6784\u9020\u6a21\u578b\u3002\u5b83\u548cPyTorch\u7684\u5176\u4ed6\u5c42\u5728\u4f7f\u7528\u4e0a\u5f88\u7c7b\u4f3c\u3002<\/p>\n\n\n\n
net = nn.Sequential(\n    MyDictDense(),\n    MyListDense(),\n)\nprint(net)\nprint(net(x))<\/code><\/pre>\n\n\n\n
\u8f93\u51fa\uff1a<\/p>\n\n\n\n
Sequential(\n  (0): MyDictDense(\n    (params): ParameterDict(\n        (linear1): Parameter containing: [torch.FloatTensor of size 4x4]\n        (linear2): Parameter containing: [torch.FloatTensor of size 4x1]\n        (linear3): Parameter containing: [torch.FloatTensor of size 4x2]\n    )\n  )\n  (1): MyListDense(\n    (params): ParameterList(\n        (0): Parameter containing: [torch.FloatTensor of size 4x4]\n        (1): Parameter containing: [torch.FloatTensor of size 4x4]\n        (2): Parameter containing: [torch.FloatTensor of size 4x4]\n        (3): Parameter containing: [torch.FloatTensor of size 4x1]\n    )\n  )\n)\ntensor([[-101.2394]], grad_fn=<MmBackward>)<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
\u4e0d\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42 \u5148\u4ecb\u7ecd\u5982\u4f55\u5b9a\u4e49\u4e00\u4e2a\u4e0d\u542b\u6a21\u578b\u53c2\u6570\u7684\u81ea\u5b9a\u4e49\u5c42\u3002\u4e8b\u5b9e\u4e0a\uff0c\u8fd9\u548c4.1\u8282\uff08\u6a21\u578b\u6784\u9020\uff09\u4e2d\u4ecb\u7ecd\u7684\u4f7f\u7528Module\u7c7b\u6784\u9020\u6a21 …<\/p>","protected":false},"author":1,"featured_media":1816,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46,3],"tags":[45,44,12,22],"class_list":["post-1815","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-46","category-3","tag-45","tag-44","tag-12","tag-22"],"_links":{"self":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1815","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/comments?post=1815"}],"version-history":[{"count":1,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1815\/revisions"}],"predecessor-version":[{"id":1817,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1815\/revisions\/1817"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/media\/1816"}],"wp:attachment":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/media?parent=1815"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/categories?post=1815"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/tags?post=1815"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}