{"id":22997,"date":"2024-12-16T10:01:57","date_gmt":"2024-12-16T02:01:57","guid":{"rendered":"http:\/\/139.9.1.231\/?p=22997"},"modified":"2025-06-20T17:36:20","modified_gmt":"2025-06-20T09:36:20","slug":"non-causal-convolutioncausal-convolution","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2024\/12\/16\/non-causal-convolutioncausal-convolution\/","title":{"rendered":"\u975e\u56e0\u679c\u5377\u79ef\/\u56e0\u679c\u5377\u79ef"},"content":{"rendered":"\n

\u56e0\u679c\u5377\u79ef:<\/strong><\/h2>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u56e0\u679c\u5377\u79ef\u53ef\u4ee5\u7528\u4e0a\u56fe\u76f4\u89c2\u8868\u793a\u3002\u5373\u5bf9\u4e8e\u4e0a\u4e00\u5c42t<\/em>\u65f6\u523b\u7684\u503c\uff0c\u53ea\u4f9d\u8d56\u4e8e\u4e0b\u4e00\u5c42t<\/em>\u65f6\u523b\u53ca\u5176\u4e4b\u524d\u7684\u503c\u3002\u548c\u4f20\u7edf\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\uff0c\u56e0\u679c\u5377\u79ef\u4e0d\u80fd\u770b\u5230\u672a\u6765\u7684\u6570\u636e\uff0c\u5b83\u662f\u5355\u5411\u7684\u7ed3\u6784\uff0c\u4e0d\u662f\u53cc\u5411\u7684\u3002\u4e5f\u5c31\u662f\u8bf4\u53ea\u6709\u6709\u4e86\u524d\u9762\u7684\u56e0\u624d\u6709\u540e\u9762\u7684\u679c\uff0c\u662f\u4e00\u79cd\u4e25\u683c\u7684\u65f6\u95f4\u7ea6\u675f\u6a21\u578b\uff0c\u56e0\u6b64\u88ab\u6210\u4e3a\u56e0\u679c\u5377\u79ef\u3002<\/p>\n\n\n\n

\u4e0a\u9762\u7684\u56fe\u7247\u53ef\u4ee5\u8be6\u7ec6\u7684\u89e3\u91ca\u56e0\u679c\u5377\u79ef\uff0c\u4f46\u662f\u95ee\u9898\u5c31\u6765\uff0c\u5982\u679c\u6211\u8981\u8003\u8651\u5f88\u4e45\u4e4b\u524d\u7684\u53d8\u91cfx\uff0c\u90a3\u4e48\u5377\u79ef\u5c42\u6570\u5c31\u5fc5\u987b\u589e\u52a0\uff08\u81ea\u884c\u4f53\u4f1a\uff09\u3002\u3002\u3002\u5377\u79ef\u5c42\u6570\u7684\u589e\u52a0\u5c31\u5e26\u6765\uff1a\u68af\u5ea6\u6d88\u5931\uff0c\u8bad\u7ec3\u590d\u6742\uff0c\u62df\u5408\u6548\u679c\u4e0d\u597d\u7684\u95ee\u9898\uff0c\u4e3a\u4e86\u51b3\u7edd\u8fd9\u4e2a\u95ee\u9898\uff0c\u51fa\u73b0\u4e86\u6269\u5c55\u5377\u79ef\uff08dilated\uff09<\/strong><\/p>\n\n\n\n

(1) \u6d41\u5f0f\u63a8\u7406\u4e2d\u7684\u5377\u79ef\u8981\u6c42<\/strong><\/p>\n\n\n\n

  • \u65e0\u672a\u6765\u4fe1\u606f\u4f9d\u8d56<\/strong>\uff1a\u5377\u79ef\u6838\u53ea\u80fd\u8bbf\u95ee\u5f53\u524d\u53ca\u4e4b\u524d\u7684\u8f93\u5165\uff0c\u4e0d\u5141\u8bb8\u8bbf\u95ee\u672a\u6765\u8f93\u5165\u3002<\/li>
  • \u56e0\u679c\u5377\u79ef\uff08Causal Convolution\uff09<\/strong>\uff1a\u901a\u8fc7\u8c03\u6574\u5377\u79ef\u6838\u7684 Padding\uff0c\u4f7f\u5377\u79ef\u64cd\u4f5c\u4ec5\u4f9d\u8d56\u5386\u53f2\u65f6\u95f4\u6b65\u7684\u6570\u636e\u3002<\/li><\/ul>\n\n\n\n

    (2) Padding \u8bbe\u8ba1<\/strong><\/p>\n\n\n\n

    • \u666e\u901a\u5377\u79ef\u7684 Padding<\/strong>\uff1a\u5728\u975e\u6d41\u5f0f\u6a21\u578b\u4e2d\uff0c\u901a\u5e38\u4f7f\u7528 SAME<\/code> Padding\uff08\u5982 TensorFlow \u6216 PyTorch \u7684\u5bf9\u79f0\u586b\u5145\uff09\uff0c\u586b\u5145\u65b9\u5f0f\u4f7f\u5f97\u8f93\u5165\u548c\u8f93\u51fa\u957f\u5ea6\u4e00\u81f4\u3002\u8fd9\u4f1a\u5bfc\u81f4\u5377\u79ef\u6838\u8bbf\u95ee\u672a\u6765\u65f6\u95f4\u6b65\u6570\u636e\uff0c\u65e0\u6cd5\u5b9e\u73b0\u6d41\u5f0f\u63a8\u7406\u3002<\/li>
    • \u56e0\u679c\u5377\u79ef\u7684 Padding<\/strong>\uff1a
      • \u5bf9\u5377\u79ef\u6838\u8fdb\u884c\u4e0d\u5bf9\u79f0\u586b\u5145\uff08\u5982\u53ea\u5728\u8f93\u5165\u524d\u4fa7\u586b\u5145\uff09\uff0c\u4f7f\u5f97\u5377\u79ef\u64cd\u4f5c\u4ec5\u4f9d\u8d56\u4e8e\u5f53\u524d\u53ca\u4e4b\u524d\u7684\u65f6\u95f4\u6b65\u3002<\/strong><\/li>
      • \u5177\u4f53\u586b\u5145\u91cf = \u5377\u79ef\u6838\u5927\u5c0f – 1\uff0c\u4f8b\u5982 3×1 \u5377\u79ef\u6838\u7684\u586b\u5145\u91cf\u662f 2\u3002<\/strong><\/li><\/ul><\/li><\/ul>\n\n\n\n
        \"\"<\/figure><\/div>\n\n\n\n
        import torch\nimport torch.nn as nn\nfrom torch.autograd import Variable\n\n__CUDA__ = torch.cuda.is_available()\n\nclass CausalConv1d(nn.Module):\n    \"\"\"\n    A causal 1D convolution.\n    \"\"\"\n    def __init__(self, kernel_size, in_channels, out_channels, dilation):\n        super(CausalConv1d, self).__init__(self)\n        \n        # attributes:\n        self.kernel_size = kernel_size\n        self.in_channels = in_channels\n        self.dilation = dilation\n        \n        # modules:\n        self.conv1d = torch.nn.Conv1d(in_channels, out_channels,\n                                      kernel_size, stride=1,\n                                      padding=padding = (kernel_size-1) * dilation,\n                                      dilation=dilation)\n\n    def forward(self, seq):\n        \"\"\"\n        Note that Conv1d expects (batch, in_channels, in_length).\n        We assume that seq ~ (len(seq), batch, in_channels), so we'll reshape it first.\n        \"\"\"\n        seq_ = seq.permute(1,2,0)\n        conv1d_out = self.conv1d(seq_).permute(2,0,1)\n        # remove k-1 values from the end:\n        return conv1d_out[0:-(self.kernel_size-1)]<\/code><\/pre>\n\n\n\n
        \u6269\u5c55\u56e0\u679c\u5377\u79ef\uff1a\u3010\u7a7a\u6d1e\u56e0\u679c\u5377\u79ef Dilated causal Conv\u3011<\/strong><\/h2>\n\n\n\n
        \u5bf9\u4e8e\u56e0\u679c\u5377\u79ef\uff0c\u5b58\u5728\u7684\u4e00\u4e2a\u95ee\u9898\u662f\u9700\u8981\u5f88\u591a\u5c42\u6216\u8005\u5f88\u5927\u7684filter\u6765\u589e\u52a0\u5377\u79ef\u7684\u611f\u53d7\u91ce\u3002<\/strong>\u6269\u5927\u5377\u79ef\uff08dilated convolution\uff09\u662f\u901a\u8fc7\u8df3\u8fc7\u90e8\u5206\u8f93\u5165\u6765\u4f7ffilter\u53ef\u4ee5\u5e94\u7528\u4e8e\u5927\u4e8efilter\u672c\u8eab\u957f\u5ea6\u7684\u533a\u57df\u3002\u7b49\u540c\u4e8e\u901a\u8fc7\u589e\u52a0\u96f6\u6765\u4ece\u539f\u59cbfilter\u4e2d\u751f\u6210\u66f4\u5927\u7684filter\u3002<\/strong><\/p>\n\n\n\n
        dilated\u7684\u597d\u5904\u662f\u4e0d\u505apooling\u635f\u5931\u4fe1\u606f\u7684\u60c5\u51b5\u4e0b\uff0c\u52a0\u5927\u4e86\u611f\u53d7\u91ce\uff0c\u8ba9\u6bcf\u4e2a\u5377\u79ef\u8f93\u51fa\u90fd\u5305\u542b\u8f83\u5927\u8303\u56f4\u7684\u4fe1\u606f\u3002\u5728\u56fe\u50cf\u9700\u8981\u5168\u5c40\u4fe1\u606f\u6216\u8005\u8bed\u97f3\u6587\u672c\u9700\u8981\u8f83\u957f\u7684sequence\u4fe1\u606f\u4f9d\u8d56\u7684\u95ee\u9898\u4e2d\uff0c\u90fd\u80fd\u5f88\u597d\u7684\u5e94\u7528dilated conv\uff0c\u6bd4\u5982\u56fe\u50cf\u5206\u5272\u3001\u8bed\u97f3\u5408\u6210WaveNet\u3001\u673a\u5668\u7ffb\u8bd1ByteNet\u4e2d.<\/p>\n\n\n\n
        \"\"<\/figure>\n\n\n\n
        Normalization \u5c42\u7684\u9009\u62e9\u4e0e\u8c03\u6574<\/h2>\n\n\n\n
        Normalization \u662f\u6d41\u5f0f\u63a8\u7406\u4e2d\u53e6\u4e00\u4e2a\u5173\u952e\u6311\u6218\u3002\u666e\u901a\u7684\u6279\u5f52\u4e00\u5316\uff08Batch Normalization, BN\uff09\u9700\u8981\u8ba1\u7b97\u5168\u5c40\u7edf\u8ba1\u91cf\uff08\u5982\u5747\u503c\u548c\u65b9\u5dee\uff09\uff0c\u8fd9\u5728\u6d41\u5f0f\u63a8\u7406\u4e2d\u662f\u4e0d\u53ef\u80fd\u5b9e\u73b0\u7684\u3002<\/strong><\/p>\n\n\n\n
        (1) Batch Normalization \u7684\u95ee\u9898<\/strong><\/h4>\n\n\n\n