{"id":2882,"date":"2022-03-11T16:55:23","date_gmt":"2022-03-11T08:55:23","guid":{"rendered":"http:\/\/139.9.1.231\/?p=2882"},"modified":"2022-03-11T17:10:11","modified_gmt":"2022-03-11T09:10:11","slug":"pack_padded_sequence-pad_packed_sequence","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/03\/11\/pack_padded_sequence-pad_packed_sequence\/","title":{"rendered":"torch.nn.utils.rnn\u4e2d\u7684pack_padded_sequence\u548c pad_packed_sequence and pad_sequence"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u8fd9\u4e24\u4e2a\u662f\u4e92\u9006\u7684\u64cd\u4f5c\u3002<\/p>\n\n\n\n

\u5148\u6765\u770b\u4e3a\u4ec0\u4e48\u9700\u8981pad\u548cpack\u64cd\u4f5c\uff1a<\/p>\n\n\n\n

\u5148\u770b\u4e00\u4e2a\u4f8b\u5b50\uff0c\u8fd9\u4e2abatch\u4e2d\u67095\u4e2asample<\/p>\n\n\n\n

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

\u5982\u679c\u4e0d\u7528pack\u548cpad\u64cd\u4f5c\u4f1a\u6709\u4e00\u4e2a\u95ee\u9898\uff0c\u4ec0\u4e48\u95ee\u9898\u5462\uff1f\u6bd4\u5982\u4e0a\u56fe\uff0c\u53e5\u5b50\u201cYes\u201d\u53ea\u6709\u4e00\u4e2a\u5355\u8bcd\uff0c\u4f46\u662fpadding\u4e86\u591a\u4f59\u7684pad\u7b26\u53f7\uff0c\u8fd9\u6837\u4f1a\u5bfc\u81f4LSTM\u5bf9\u5b83\u7684\u8868\u793a\u901a\u8fc7\u4e86\u975e\u5e38\u591a\u65e0\u7528\u7684\u5b57\u7b26\uff0c\u8fd9\u6837\u5f97\u5230\u7684\u53e5\u5b50\u8868\u793a\u5c31\u4f1a\u6709\u8bef\u5dee\uff0c\u66f4\u76f4\u89c2\u7684\u5982\u4e0b\u56fe\uff1a<\/p>\n\n\n\n

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

\u90a3\u4e48\u6211\u4eec\u6b63\u786e\u7684\u505a\u6cd5\u5e94\u8be5\u662f\u600e\u4e48\u6837\u5462\uff1f<\/p>\n\n\n\n

\u5728\u4e0a\u9762\u8fd9\u4e2a\u4f8b\u5b50\uff0c\u6211\u4eec\u60f3\u8981\u5f97\u5230\u7684\u8868\u793a\u4ec5\u4ec5\u662fLSTM\u8fc7\u5b8c\u5355\u8bcd”Yes”\u4e4b\u540e\u7684\u8868\u793a\uff0c\u800c\u4e0d\u662f\u901a\u8fc7\u4e86\u591a\u4e2a\u65e0\u7528\u7684\u201cPad\u201d\u5f97\u5230\u7684\u8868\u793a\uff1a\u5982\u4e0b\u56fe\uff1a<\/p>\n\n\n\n

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

torch.nn.utils.rnn.pack_padded_sequence()<\/h3>\n\n\n\n

\u8fd9\u91cc\u7684pack<\/code>\uff0c\u7406\u89e3\u6210\u538b\u7d27\u6bd4\u8f83\u597d\u3002 \u5c06\u4e00\u4e2a \u586b\u5145\u8fc7\u7684\u53d8\u957f\u5e8f\u5217 \u538b\u7d27\u3002\uff08\u586b\u5145\u65f6\u5019\uff0c\u4f1a\u6709\u5197\u4f59\uff0c\u6240\u4ee5\u538b\u7d27\u4e00\u4e0b\uff09<\/p>\n\n\n\n

\u5176\u4e2dpack\u7684\u8fc7\u7a0b\u4e3a\uff1a\uff08\u6ce8\u610fpack\u7684\u5f62\u5f0f\uff0c\u4e0d\u662f\u6309\u884c\u538b\uff0c\u800c\u662f\u6309\u5217\u538b\uff09<\/p>\n\n\n\n

pack\u4e4b\u540e\uff0c\u539f\u6765\u586b\u5145\u7684 PAD\uff08\u4e00\u822c\u521d\u59cb\u5316\u4e3a0\uff09\u5360\u4f4d\u7b26\u88ab\u5220\u6389\u4e86\u3002<\/p>\n\n\n\n

\u8f93\u5165\u7684\u5f62\u72b6\u53ef\u4ee5\u662f(T\u00d7B\u00d7* )\u3002T<\/code>\u662f\u6700\u957f\u5e8f\u5217\u957f\u5ea6\uff0cB<\/code>\u662fbatch size<\/code>\uff0c*<\/code>\u4ee3\u8868\u4efb\u610f\u7ef4\u5ea6(\u53ef\u4ee5\u662f0)\u3002\u5982\u679cbatch_first=True<\/code>\u7684\u8bdd\uff0c\u90a3\u4e48\u76f8\u5e94\u7684 input size<\/code> \u5c31\u662f (B\u00d7T\u00d7*)<\/code>\u3002<\/p>\n\n\n\n

Variable<\/code>\u4e2d\u4fdd\u5b58\u7684\u5e8f\u5217\uff0c\u5e94\u8be5\u6309\u5e8f\u5217\u957f\u5ea6\u7684\u957f\u77ed\u6392\u5e8f\uff0c\u957f\u7684\u5728\u524d\uff0c\u77ed\u7684\u5728\u540e\u3002\u5373input[:,0]<\/code>\u4ee3\u8868\u7684\u662f\u6700\u957f\u7684\u5e8f\u5217\uff0cinput[:, B-1]<\/code>\u4fdd\u5b58\u7684\u662f\u6700\u77ed\u7684\u5e8f\u5217\u3002<\/p>\n\n\n\n

NOTE\uff1a<\/code> \u53ea\u8981\u662f\u7ef4\u5ea6\u5927\u4e8e\u7b49\u4e8e2\u7684input<\/code>\u90fd\u53ef\u4ee5\u4f5c\u4e3a\u8fd9\u4e2a\u51fd\u6570\u7684\u53c2\u6570\u3002\u4f60\u53ef\u4ee5\u7528\u5b83\u6765\u6253\u5305labels<\/code>\uff0c\u7136\u540e\u7528RNN<\/code>\u7684\u8f93\u51fa\u548c\u6253\u5305\u540e\u7684labels<\/code>\u6765\u8ba1\u7b97loss<\/code>\u3002\u901a\u8fc7PackedSequence<\/code>\u5bf9\u8c61\u7684.data<\/code>\u5c5e\u6027\u53ef\u4ee5\u83b7\u53d6 Variable<\/code>\u3002<\/p>\n\n\n\n

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

\u53c2\u6570\u8bf4\u660e:<\/p>\n\n\n\n

  • input (Variable) \u2013 \u53d8\u957f\u5e8f\u5217 \u88ab\u586b\u5145\u540e\u7684 batch<\/li>
  • lengths (list[int]) \u2013 Variable<\/code> \u4e2d \u6bcf\u4e2a\u5e8f\u5217\u7684\u957f\u5ea6\u3002<\/li>
  • batch_first (bool, optional) \u2013 \u5982\u679c\u662fTrue<\/code>\uff0cinput\u7684\u5f62\u72b6\u5e94\u8be5\u662fB*T*size<\/code>\u3002<\/li>
  • \u53c2\u6570 enforce_sorted<\/strong> \uff0c\u5982\u679c\u662f True \uff0c\u5219\u8f93\u5165\u5e94\u8be5\u662f\u6309\u957f\u5ea6\u964d\u5e8f\u6392\u5e8f\u7684\u5e8f\u5217\u3002\u5982\u679c\u662f False \uff0c\u4f1a\u5728\u51fd\u6570\u5185\u90e8\u8fdb\u884c\u6392\u5e8f\u3002\u9ed8\u8ba4\u503c\u4e3a True \u3002\u4e5f\u5c31\u662f\u8bf4\u5728\u8f93\u5165 pack_padded_sequence \u524d\uff0c\u6211\u4eec\u4e5f\u53ef\u4ee5\u4e0d\u5bf9\u6570\u636e\u8fdb\u884c\u6392\u5e8f\u3002<\/li><\/ul>\n\n\n\n

    \u8fd4\u56de\u503c:<\/p>\n\n\n\n

    \u4e00\u4e2aPackedSequence<\/code> \u5bf9\u8c61\u3002<\/p>\n\n\n\n

    torch.nn.utils.rnn.pad_packed_sequence()<\/h3>\n\n\n\n

    \u586b\u5145packed_sequence<\/code>\u3002<\/p>\n\n\n\n

    \u4e0a\u9762\u63d0\u5230\u7684\u51fd\u6570\u7684\u529f\u80fd\u662f\u5c06\u4e00\u4e2a\u586b\u5145\u540e\u7684\u53d8\u957f\u5e8f\u5217\u538b\u7d27\u3002 \u8fd9\u4e2a\u64cd\u4f5c\u548cpack_padded_sequence()\u662f\u76f8\u53cd\u7684\u3002\u628a\u538b\u7d27\u7684\u5e8f\u5217\u518d\u586b\u5145\u56de\u6765\u3002\u586b\u5145\u65f6\u4f1a\u521d\u59cb\u5316\u4e3a0\u3002<\/p>\n\n\n\n

    \u8fd4\u56de\u7684Varaible\u7684\u503c\u7684size<\/code>\u662f T\u00d7B\u00d7*<\/code>, T<\/code> \u662f\u6700\u957f\u5e8f\u5217\u7684\u957f\u5ea6\uff0cB<\/code> \u662f batch_size,\u5982\u679c batch_first=True<\/code>,\u90a3\u4e48\u8fd4\u56de\u503c\u662fB\u00d7T\u00d7*<\/code>\u3002<\/p>\n\n\n\n

    Batch\u4e2d\u7684\u5143\u7d20\u5c06\u4f1a\u4ee5\u5b83\u4eec\u957f\u5ea6\u7684\u9006\u5e8f\u6392\u5217\u3002<\/p>\n\n\n\n

    \u53c2\u6570\u8bf4\u660e:<\/p>\n\n\n\n

    • sequence (PackedSequence) \u2013 \u5c06\u8981\u88ab\u586b\u5145\u7684 batch<\/li>
    • batch_first (bool, optional) \u2013 \u5982\u679c\u4e3aTrue\uff0c\u8fd4\u56de\u7684\u6570\u636e\u7684\u683c\u5f0f\u4e3a B\u00d7T\u00d7*<\/code>\u3002<\/li><\/ul>\n\n\n\n

      \u8fd4\u56de\u503c: \u4e00\u4e2atuple\uff0c\u5305\u542b\u88ab\u586b\u5145\u540e\u7684\u5e8f\u5217\uff0c\u548cbatch\u4e2d\u5e8f\u5217\u7684\u957f\u5ea6\u5217\u8868<\/p>\n\n\n\n

      \u5b9e\u4f8b \u4ee3\u7801\uff1a<\/p>\n\n\n\n

      \n\n>>> from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence\n>>> seq = torch.tensor([[1,2,0], [3,0,0], [4,5,6]])\n#seq\u7684\u7ef4\u5ea6\u662f3*3\uff0c\u7b2c\u4e00\u4e2a \u7ef4\u5ea6\u8868\u793aT\uff0cis the length of the longest sequence \n#\u7b2c\u4e8c\u7ef4\u8868\u793aB\uff0c\u6279\u6b21\u5927\u5c0f\uff0c\u4e5f\u5c31\u662f\u8bf4\u67093\u4e2a\u957f\u5ea6\u4e3a3\u7684\u5411\u91cf\uff0c\u5176\u4e2d\u6bcf\u5217\u8868\u793a\u4e00\u4e2a\u5e8f\u5217\uff0c\u5e8f\u5217\u90fd\u662f\u7b49\u957f\u7684\uff0c\u56e0\u4e3a\u77ed\u7684\u5e8f\u5217\u5df2\u7ecf\u75280\u8865\u9f50\u4e86\n>>> lens = [2, 1, 3]\n>>> packed = pack_padded_sequence(seq, lens, batch_first=True, enforce_sorted=False)\n#\u5c06\u8865\u9f50\u7684\u5e8f\u5217\u538b\u7d27\u6210\u4e00\u4e2a\u5e8f\u5217\uff0c\u5c060\u53bb\u6389\n>>> packed\nPackedSequence(data=tensor([4, 1, 3, 5, 2, 6]), batch_sizes=tensor([3, 2, 1]),\n               sorted_indices=tensor([2, 0, 1]), unsorted_indices=tensor([1, 2, 0]))\n>>> seq_unpacked, lens_unpacked = pad_packed_sequence(packed, batch_first=True)\n>>> seq_unpacked\ntensor([[1, 2, 0],\n        [3, 0, 0],\n        [4, 5, 6]])\n>>> lens_unpacked\ntensor([2, 1, 3])<\/code><\/pre>\n\n\n\n
      pad_sequence<\/h2>\n\n\n\n
      \u53c2\u6570<\/strong><\/h3>\n\n\n\n
      sequences<\/strong>\uff1a\u8868\u793a\u8f93\u5165\u6837\u672c\u5e8f\u5217\uff0c\u4e3a list \u7c7b\u578b\uff0clist \u4e2d\u7684\u5143\u7d20\u4e3a tensor \u7c7b\u578b\u3002 tensor \u7684 size \u4e3a L * F \u3002\u5176\u4e2d\uff0cL \u4e3a\u5355\u4e2a\u5e8f\u5217\u7684\u957f\u5ea6\uff0cF \u4e3a\u5e8f\u5217\u4e2d\u6bcf\u4e2a\u65f6\u95f4\u6b65\uff08time step\uff09\u7279\u5f81\u7684\u4e2a\u6570\uff0c\u6839\u636e\u4efb\u52a1\u7684\u4e0d\u540c F \u7684\u7ef4\u5ea6\u4f1a\u6709\u6240\u4e0d\u540c\u3002<\/p>\n\n\n\n
      batch_first<\/strong>\uff1a\u4e3a True \u5bf9\u5e94 [batch_size, seq_len, feature]\uff1bFalse \u5bf9\u5e94[seq_len, batch_size, feature]\uff0c\u4ece\u4e60\u60ef\u4e0a\u6765\u8bb2\u4e00\u822c\u8bbe\u7f6e\u4e3a True \u6bd4\u8f83\u7b26\u5408\u6211\u4eec\u7684\u8ba4\u77e5\u3002<\/p>\n\n\n\n
      padding_value<\/strong>\uff1a\u586b\u5145\u503c\uff0c\u9ed8\u8ba4\u503c\u4e3a 0 \u3002<\/p>\n\n\n\n
      \u8bf4\u660e<\/strong><\/h3>\n\n\n\n
      \u4e3b\u8981\u7528\u6765\u5bf9\u6837\u672c\u8fdb\u884c\u586b\u5145\uff0c\u586b\u5145\u503c\u4e00\u822c\u4e3a 0 \u3002\u6211\u4eec\u5728\u8bad\u7ec3\u7f51\u7edc\u65f6\uff0c\u4e00\u822c\u4f1a\u91c7\u7528\u4e00\u4e2a\u4e00\u4e2a mini-batch \u7684\u65b9\u5f0f\uff0c\u5c06\u8bad\u7ec3\u6837\u672c\u6570\u636e\u5582\u7ed9\u7f51\u7edc\u3002\u5728 PyTorch \u91cc\u9762\u6570\u636e\u90fd\u662f\u4ee5 tensor \u7684\u5f62\u5f0f\u5b58\u5728\uff0c\u4e00\u4e2a mini-batch \u5b9e\u9645\u4e0a\u5c31\u662f\u4e00\u4e2a\u9ad8\u7ef4\u7684 tensor \uff0c\u6bcf\u4e2a\u5e8f\u5217\u6570\u636e\u7684\u957f\u5ea6\u5fc5\u987b\u76f8\u540c\u624d\u80fd\u7ec4\u6210\u4e00\u4e2a tensor \u3002\u4e3a\u4e86\u4f7f\u7f51\u7edc\u53ef\u4ee5\u5904\u7406 mini-batch \u5f62\u5f0f\u7684\u6570\u636e\uff0c\u5c31\u5fc5\u987b\u5bf9\u5e8f\u5217\u6837\u672c\u8fdb\u884c\u586b\u5145\uff0c\u4fdd\u8bc1\u4e00\u4e2a mini-batch \u91cc\u9762\u7684\u6570\u636e\u957f\u5ea6\u662f\u76f8\u540c\u7684\u3002<\/p>\n\n\n\n
      \u5728 PyTorch \u91cc\u9762\u4e00\u822c\u662f\u4f7f\u7528 DataLoader \u8fdb\u884c\u6570\u636e\u52a0\u8f7d\uff0c\u8fd4\u56de mini-batch \u5f62\u5f0f\u7684\u6570\u636e\uff0c\u518d\u5c06\u6b64\u6570\u636e\u5582\u7ed9\u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\u3002\u6211\u4eec\u4e00\u822c\u4f1a\u81ea\u5b9a\u4e49\u4e00\u4e2a collate_fn \u51fd\u6570\uff0c\u5b8c\u6210\u5bf9\u6570\u636e\u7684\u586b\u5145\u3002<\/p>\n\n\n\n
      import torch\nfrom torch.utils.data import Dataset, DataLoader\nfrom torch.nn.utils.rnn import pad_sequence,pack_padded_sequence,pack_sequence,pad_packed_sequence\n\nclass MyData(Dataset):\n    def __init__(self, data):\n        self.data = data\n\n    def __len__(self):\n        return len(self.data)\n\n    def __getitem__(self, idx):\n        return self.data[idx]\n\ndef collate_fn(data):\n    data.sort(key=lambda x: len(x), reverse=True)\n    data = pad_sequence(data, batch_first=True, padding_value=0)\n    return data\n\na = torch.tensor([1,2,3,4])\nb = torch.tensor([5,6,7])\nc = torch.tensor([7,8])\nd = torch.tensor([9])\ntrain_x = [a, b, c, d]\n\ndata = MyData(train_x)\ndata_loader = DataLoader(data, batch_size=2, shuffle=True, collate_fn=collate_fn)\n# \u91c7\u7528\u9ed8\u8ba4\u7684 collate_fn \u4f1a\u62a5\u9519<\/em>\n#data_loader = DataLoader(data, batch_size=2, shuffle=True) <\/em>\nbatch_x = iter(data_loader).next()<\/code><\/pre>\n\n\n\n
      \u8fd0\u884c\u7a0b\u5e8f\uff0c\u5f97\u5230 batch_x \u7684\u503c\uff1a<\/p>\n\n\n\n
      # batch_x<\/em>\ntensor([[1, 2, 3, 4],\n        [9, 0, 0, 0]])<\/code><\/pre>\n\n\n\n
      \u4ece batch_x \u7684\u503c\u53ef\u4ee5\u770b\u51fa\uff0c\u7b2c\u4e8c\u884c\u586b\u5145\u4e86\u4e09\u4e2a 0 \uff0c\u4f7f\u5176\u957f\u5ea6\u548c\u7b2c\u4e00\u884c\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n\n\n\n
      \u9700\u8981\u8bf4\u660e\u7684\u662f\uff0c\u5bf9\u4e8e\u957f\u5ea6\u4e0d\u540c\u7684\u5e8f\u5217\uff0c\u4f7f\u7528\u9ed8\u8ba4\u7684 collate_fn \u51fd\u6570\uff0c\u4e0d\u81ea\u5b9a\u4e49 collate_fn \u51fd\u6570\u5b8c\u6210\u5bf9\u5e8f\u5217\u7684\u586b\u5145\uff0c\u4e0a\u9762\u7684\u7a0b\u5e8f\u5c31\u4f1a\u62a5\u9519\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
      \u8fd9\u4e24\u4e2a\u662f\u4e92\u9006\u7684\u64cd\u4f5c\u3002 \u5148\u6765\u770b\u4e3a\u4ec0\u4e48\u9700\u8981pad\u548cpack\u64cd\u4f5c\uff1a \u5148\u770b\u4e00\u4e2a\u4f8b\u5b50\uff0c\u8fd9\u4e2abatch\u4e2d\u67095\u4e2asample … \u7ee7\u7eed\u9605\u8bfbtorch.nn.utils.rnn\u4e2d\u7684pack_padded_sequence\u548c pad_packed_sequence and pad_sequence<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[11],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2882"}],"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=2882"}],"version-history":[{"count":18,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2882\/revisions"}],"predecessor-version":[{"id":2907,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/2882\/revisions\/2907"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=2882"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=2882"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=2882"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}