{"id":149,"date":"2026-10-07T11:03:00","date_gmt":"2026-10-07T03:03:00","guid":{"rendered":"https:\/\/luqimu.xyz\/?p=149"},"modified":"2026-10-07T11:04:02","modified_gmt":"2026-10-07T03:04:02","slug":"rnn%e8%ae%a9%e6%a8%a1%e5%9e%8b%e7%90%86%e8%a7%a3%e8%87%aa%e7%84%b6%e8%af%ad%e8%a8%80%e5%ba%8f%e5%88%97","status":"publish","type":"post","link":"https:\/\/luqimu.xyz\/index.php\/2026\/10\/07\/rnn%e8%ae%a9%e6%a8%a1%e5%9e%8b%e7%90%86%e8%a7%a3%e8%87%aa%e7%84%b6%e8%af%ad%e8%a8%80%e5%ba%8f%e5%88%97\/","title":{"rendered":"RNN\u67b6\u6784\u53calstm\u548cgru\u7684\u6539\u8fdb"},"content":{"rendered":"\n

\u4e00.\u89e3\u51b3\u95ee\u9898:<\/h3>\n\n\n\n

\u6587\u672c\u6a21\u578b\u4e0d\u4ec5\u8981\u77e5\u9053\u6709\u54ea\u4e9b\u8bcd,\u8fd8\u8981\u5229\u7528\u8bcd\u5e8f\u548c\u4e0a\u4e0b\u6587<\/p>\n\n\n\n

\"image-20261001144253653\"\/<\/div><\/figure>\n\n\n\n

\u4e3a\u4ec0\u4e48\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u4e0d\u9002\u5408\u5904\u7406\u81ea\u7136\u8bed\u8a00\u5e8f\u5217:<\/p>\n\n\n\n

fnn\u662f\u4e00\u79cd\u4fe1\u606f\u5355\u5411\u4f20\u64ad\u7684\u7f51\u7edc:\u6570\u636e\u4ece\u8f93\u5165\u5c42\u51fa\u53d1,\u4f9d\u6b21\u901a\u8fc7\u9690\u85cf\u5c42\u5230\u8fbe\u8f93\u51fa\u5c42,\u7f51\u7edc\u5185\u90e8\u6ca1\u6709\u5faa\u73af\u8fde\u63a5.mlp\u5c31\u662f\u7ecf\u5178\u7684fnn<\/p>\n\n\n\n

1.\u524d\u540e\u4f9d\u8d56:\u4e00\u6b21\u6027\u5904\u7406\u6574\u6bb5\u8f93\u5165,\u6ca1\u6709\u6cbf\u65f6\u95f4\u5efa\u6a21\u4e0a\u4e0b\u6587\u7684\u673a\u5236<\/p>\n\n\n\n

\"image-20261001144706578\"\/<\/div><\/figure>\n\n\n\n

2.\u53d8\u957f\u5e8f\u5217:\u8f93\u5165\u7ef4\u5ea6\u5fc5\u987b\u5148\u56fa\u5b9a,\u957f\u77ed\u4e0d\u4e00\u7684\u53e5\u5b50\u5f88\u96be\u5bf9\u9f50.<\/p>\n\n\n\n

\"image-20261001144752483\"\/<\/div><\/figure>\n\n\n\n

\u4e8c.rnn\u7684\u89e3\u51b3\u601d\u8def:\u6539\u53d8\u8bfb\u53d6\u65b9\u5f0f<\/h3>\n\n\n\n

rnn\u4e0d\u4f1a\u628a\u6574\u4e2a\u53e5\u5b50\u62fc\u6210\u56fa\u5b9a\u5411\u91cf,\u800c\u662f\u4ece\u5de6\u5230\u53f3\u9010\u4e2a\u8bfb\u53d6token.<\/p>\n\n\n\n

1.\u8bfb\u7b2c\u4e00\u4e2a\u8bcd,\u53ea\u80fd\u6839\u636e\u5f53\u524d\u8f93\u5165\u5f62\u6210\u521d\u59cb\u72b6\u6001<\/p>\n\n\n\n

2.\u8bfb\u7b2c\u4e8c\u4e2a\u8bcd,\u628a\u65b0\u8bcd\u4e0e\u4e0a\u4e00\u6b65\u7559\u4e0b\u7684\u72b6\u6001\u7ed3\u5408<\/p>\n\n\n\n

3.\u8bfb\u7b2c\u4e09\u4e2a\u8bcd,\u524d\u4e24\u6b65\u7d2f\u79ef\u7684\u4fe1\u606f\u7ee7\u7eed\u53c2\u4e0e\u8ba1\u7b97<\/p>\n\n\n\n

4,\u91cd\u590d\u8fd9\u4e2a\u8fc7\u7a0b\u76f4\u5230\u6574\u4e2a\u5e8f\u5217\u5904\u7406\u5b8c\u6bd5<\/p>\n\n\n\n

\"image-20261001145147251\"\/<\/div><\/figure>\n\n\n\n

\u9690\u85cf\u72b6\u6001:\u662f\u524d\u6587\u7684\u538b\u7f29\u6458\u8981<\/p>\n\n\n\n

\u672c\u8d28\u662f\u4e00\u4e2a\u5411\u91cf,\u4fe1\u606f\u7531\u591a\u4e2a\u7ef4\u5ea6\u5171\u540c\u8868\u8fbe<\/p>\n\n\n\n

\"image-20261001145512542\"\/<\/div><\/figure>\n\n\n\n

RNN layer<\/h4>\n\n\n\n

\u672c\u8d28:\u540c\u4e00\u4e2acell\u6cbf\u5e8f\u5217\u5faa\u73af\u8c03\u7528<\/h4>\n\n\n\n
\"image-20261001151159747\"\/<\/div><\/figure>\n\n\n\n

\u4efb\u52a1\u5934\u653e\u5728\u5faa\u73af\u5185,\u5c31\u4f1a\u6bcf\u6b65\u8f93\u51fa;\u653e\u5728\u5faa\u73af\u5916,\u5c31\u53ea\u8bfb\u6700\u540e\u72b6\u6001<\/p>\n\n\n\n

\"image-20261001151249435\"\/<\/div><\/figure>\n\n\n\n

\u4e00\u4e2acell\u4e2d\u7684\u6d41\u7a0b<\/h4>\n\n\n\n

RNNCell(Xt,Ht-1)–>Ht<\/p>\n\n\n\n

Xt:\u5f53\u524dtoken\u4e2d\u7684\u8bcd\u5411\u91cf,\u662f\u5f53\u524d\u65f6\u95f4\u6b65\u7684\u65b0\u4fe1\u606f;Ht-1:\u5f53\u524d\u5df2\u7ecf\u79ef\u7d2f\u7684\u4e0a\u4e0b\u6587<\/p>\n\n\n\n

\"image-20261001151606631\"\/<\/div><\/figure>\n\n\n\n

\u53c2\u6570\u5171\u4eab\u7684\u4f18\u70b9:<\/h4>\n\n\n\n

RNN:\u6bcf\u4e00\u6b65\u590d\u7528\u540c\u4e00\u4e2af<\/p>\n\n\n\n

\"image-20261001151814603\"\/<\/div><\/figure>\n\n\n\n

\u597d\u5904:1.\u53c2\u6570\u91cf\u4e0e\u53e5\u957f\u89e3\u8026:\u53e5\u5b50\u53d8\u957f\u53ea\u7528\u591a\u5faa\u73af,\u53c2\u6570\u89c4\u6a21\u4fdd\u6301\u4e0d\u53d8<\/p>\n\n\n\n

2.\u4e00\u5957\u89c4\u5219\u901a\u5403\u4efb\u610f\u957f\u5ea6:\u90fd\u4f7f\u7528\u4e00\u5957\u89c4\u5219\u9010\u6b65\u5904\u7406<\/p>\n\n\n\n

\u53c2\u6570\u5171\u4eab\u4f7fRNN\u771f\u6b63\u83b7\u5f97\u5904\u7406\u4e0d\u5b9a\u957f\u5e8f\u5217\u7684\u80fd\u529b<\/p>\n\n\n\n

RNN\u7c7b\u578b<\/h4>\n\n\n\n
\"image-20261001152047833\"\/<\/div><\/figure>\n\n\n\n

1.\u6574\u4f53\u5206\u7c7b:\u8f93\u5165\u4e00\u6574\u6bb5\u5e8f\u5217,\u8f93\u51fa\u4e00\u4e2a\u6574\u4f53\u6807\u7b7e<\/p>\n\n\n\n

\u5e94\u7528:\u60c5\u611f\u5206\u7c7b,\u5783\u573e\u90ae\u4ef6\u8bc6\u522b,\u65b0\u95fb\u5206\u7c7b<\/p>\n\n\n\n

\u6838\u5fc3:\u8bfb\u5b8c\u6574\u6bb5\u5e8f\u5217,\u53ea\u53d6\u6700\u540e\u72b6\u6001\u4ea4\u7ed9\u5206\u7c7b\u5668,\u8f93\u51fa\u4e00\u4e2a\u6574\u4f53\u6807\u7b7e<\/p>\n\n\n\n

\u200b
import torch
import torch.nn as nn
\u200b
class RNNClassifier(nn.Module):
    def __init__(self, vocab_size, embed_dim, hidden_dim, num_classes):
        super().__init__()
        # token id -> \u8bcd\u5411\u91cf
        self.embedding = nn.Embedding(vocab_size, embed_dim)
        # \u6cbf\u5e8f\u5217\u66f4\u65b0\u72b6\u6001
        self.rnn = nn.RNN(embed_dim, hidden_dim, batch_first=True)
        # \u72b6\u6001 -> \u7c7b\u522b\u5206\u6570
        self.classifier = nn.Linear(hidden_dim, num_classes)
\u200b
    def forward(self, token_ids):
        # token_ids: [batch, seq_len]
        x = self.embedding(token_ids)          # [batch, seq_len, embed_dim]
        # states: \u6bcf\u6b65\u72b6\u6001; h_n: \u6700\u540e\u4e00\u6b65\u72b6\u6001
        states, h_n = self.rnn(x)
        # \u53ea\u53d6\u6700\u540e\u72b6\u6001 [batch, hidden_dim]
        h_last = h_n[-1]
        # \u6574\u6bb5 -> \u4e00\u4e2a\u6807\u7b7e [batch, num_classes]
        logits = self.classifier(h_last)
        return logits
\u200b<\/pre>\n\n\n\n
2.\u5e8f\u5217\u6807\u6ce8:\u6bcf\u4e2a\u4f4d\u7f6e\u90fd\u8f93\u51fa\u6807\u7b7e<\/p>\n\n\n\n
\u5e94\u7528:\u8bcd\u6027\u6807\u6ce8,\u547d\u540d\u5b9e\u4f53\u8bc6\u522b<\/p>\n\n\n\n
\u6838\u5fc3:\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u72b6\u6001\u90fd\u8fc7\u540c\u4e00\u4e2a\u5206\u7c7b\u5668,\u9010\u4f4d\u7f6e\u8f93\u51fa\u6807\u7b7e`<\/p>\n\n\n\n
import torch
import torch.nn as nn
\u200b
class RNNLanguageModel(nn.Module):
    def __init__(self, vocab_size, embed_dim, hidden_dim):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim)
        self.rnn = nn.RNN(embed_dim, hidden_dim, batch_first=True)
        self.lm_head = nn.Linear(hidden_dim, vocab_size)
\u200b
    def forward(self, token_ids):
        x = self.embedding(token_ids)          # [batch, seq_len, embed_dim]
        states, h_n = self.rnn(x)              # \u4fdd\u7559\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u72b6\u6001
        logits = self.lm_head(states)          # \u6bcf\u4e2a\u4f4d\u7f6e\u9884\u6d4b\u4e0b\u4e00\u8bcd [batch, seq_len, vocab_size]
        return logits<\/pre>\n\n\n\n
3.\u5e8f\u5217\u751f\u6210:\u6bcf\u4e2a\u4f4d\u7f6e\u9884\u6d4b\u4e0b\u4e00\u4e2atoken<\/p>\n\n\n\n
\u6838\u5fc3:\u4e0e\u5e8f\u5217\u6807\u6ce8\u540c\u6784,\u4f46\u8f93\u51fa\u7ef4\u5ea6\u6362\u6210\u8bcd\u8868\u5927\u5c0f,\u6bcf\u4e2a\u4f4d\u7f6e\u90fd\u9884\u6d4b\u4e0b\u4e00\u4e2atoken<\/p>\n\n\n\n
import torch
import torch.nn as nn
\u200b
class RNNLanguageModel(nn.Module):
    def __init__(self, vocab_size, embed_dim, hidden_dim):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embed_dim)
        self.rnn = nn.RNN(embed_dim, hidden_dim, batch_first=True)
        self.lm_head = nn.Linear(hidden_dim, vocab_size)
\u200b
    def forward(self, token_ids):
        x = self.embedding(token_ids)          # [batch, seq_len, embed_dim]
        states, h_n = self.rnn(x)              # \u4fdd\u7559\u6bcf\u4e2a\u65f6\u95f4\u6b65\u7684\u72b6\u6001
        logits = self.lm_head(states)          # \u6bcf\u4e2a\u4f4d\u7f6e\u9884\u6d4b\u4e0b\u4e00\u8bcd [batch, seq_len, vocab_size]
        return logits<\/pre>\n\n\n\n
\u4e09\u7c7b\u4efb\u52a1\u8f93\u51fa\u65b9\u5f0f\u4e0d\u540c,\u8bad\u7ec3\u6d41\u7a0b\u662f\u4e00\u81f4\u7684:<\/p>\n\n\n\n
\u6cbf\u5e8f\u5217\u524d\u5411–>\u4efb\u52a1\u5934\u8f93\u51fa–>\u8ba1\u7b97\u6709\u6548\u4f4d\u7f6e\u635f\u5931–>\u6cbf\u72b6\u6001\u94fe\u53cd\u5411\u4f20\u64ad<\/p>\n\n\n\n
\u68af\u5ea6\u56de\u4f20\u8def\u5f84:\u635f\u5931\u4ece\u8f93\u51fa–>\u4efb\u52a1\u5934–>\u9690\u85cf\u72b6\u6001\u94fe–>Embedding<\/p>\n\n\n\n
\u6240\u6709\u4efb\u52a1\u52a8\u5458\u5f52\u7ed3\u4e8e\u4e00\u7c7b:\u628a\u68af\u5ea6\u6cbf\u65f6\u95f4\u94fe\u4f20\u56de\u5171\u4eab\u7684\u53c2\u6570<\/p>\n\n\n\n
\u5e8f\u5217\u751f\u6210\u5b9e\u5217<\/h3>\n\n\n\n
\"image-20261001153551885\"\/<\/div><\/figure>\n\n\n\n
1.\u51c6\u5907\u9884\u6599\u4e0e\u8bad\u7ec3\u5bf9<\/p>\n\n\n\n
import random
from pathlib import Path
import torch
import torch.nn as nn
from torch.nn.utils.rnn import pad_sequence  # \u628a\u53d8\u957f\u5e8f\u5217\u8865\u9f50\u6210\u540c\u4e00\u957f\u5ea6
\u200b
# ---------- 1) \u8bfb\u53d6\u8bed\u6599 ----------
# \u8bed\u6599\u6bcf\u884c\u4e00\u53e5\u8bdd\uff0c\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u5df2\u7528\u7a7a\u683c\u5206\u597d
lines = corpus_path.read_text(encoding=\"utf-8\").splitlines()
# \u4e22\u6389\u7a7a\u884c\uff0c\u518d\u6309\u7a7a\u683c\u5207\u6210 token \u5217\u8868: [\"\u6211\", \"\u559c\u6b22\", ...]
sentences = [s.split() for s in lines if s.strip()]
\u200b
# ---------- 2) \u5efa\u7acb\u8bcd\u8868 ----------
# \u4e09\u4e2a\u7279\u6b8a token \u56fa\u5b9a\u653e\u5728\u6700\u524d\uff0c\u4fdd\u8bc1\u5b83\u4eec\u7684 id \u7a33\u5b9a
PAD, UNK, EOS = \"<PAD>\", \"<UNK>\", \"<EOS>\"    # \u8865\u9f50 \/ \u672a\u77e5\u8bcd \/ \u53e5\u5b50\u7ed3\u675f
# \u5176\u4f59\u8bcd\u53bb\u91cd\u540e\u6392\u5e8f\uff0c\u63a5\u5728\u7279\u6b8a token \u4e4b\u540e\u7ec4\u6210\u8bcd\u8868
vocab = [PAD, UNK, EOS] + sorted(
            {w for s in sentences for w in s})  # \u8bcd \u2192 \u6574\u6570 id
word_to_id = {w: i for i, w in enumerate(vocab)}  
# id \u2192 \u8bcd (\u9884\u6d4b\u65f6\u53cd\u67e5)
id_to_word = {i: w for w, i in word_to_id.items()}  
# \u8bb0\u4e0b <PAD> \u7684 id
pad_id = word_to_id[PAD]                           
\u200b
# ---------- 3) \u5236\u4f5c\u8bad\u7ec3\u5bf9 ----------
# \u6bcf\u53e5\u8bdd\u672b\u5c3e\u8865\u4e00\u4e2a <EOS>\uff0c\u8ba9\u6a21\u578b\u5b66\u4f1a\"\u4f55\u65f6\u505c\u4e0b\"
encoded = [torch.tensor(
    [word_to_id[w] for w in s] + [word_to_id[EOS]])
    for s in sentences]
\u200b
# \u8f93\u5165\u53bb\u6389\u6700\u540e\u4e00\u4e2a\u8bcd\u3001\u76ee\u6807\u53bb\u6389\u7b2c\u4e00\u4e2a\u8bcd \u2192 \u4e24\u8005\u5929\u7136\u9519\u5f00\u4e00\u4f4d
# \u4f8b: \u8f93\u5165\u300c\u6211 \u559c\u6b22 \u5403 \u82f9\u679c\u300d\/ \u76ee\u6807\u300c\u559c\u6b22 \u5403 \u82f9\u679c <EOS>\u300d
inputs = pad_sequence([e[:-1] for e in encoded],
                      batch_first=True, padding_value=pad_id)  # \u540c\u4e00 batch \u5185\u77ed\u53e5\u7528 <PAD>
targets = pad_sequence([e[1:] for e in encoded],
                       batch_first=True, padding_value=pad_id)<\/pre>\n\n\n\n
2.\u5b9a\u4e49\u6a21\u578b:embedding+rnn+linear<\/p>\n\n\n\n
\u6838\u5fc3:\u628aembedding+rnn+linear\u4e32\u8d77\u6765:token\u5e8f\u5217–>\u8bcd\u5411\u91cf–>\u6bcf\u6b65\u9690\u85cf\u72b6\u6001–>\u6bcf\u6b65\u8bcd\u8868\u5206\u6570<\/p>\n\n\n\n
import torch
import torch.nn as nn
\u200b
class NextWordRNN(nn.Module):
    def __init__(self, vocab_size, embed_dim=48, hidden_dim=96, pad_id=0):
        super().__init__()
        # \u8bcd\u5d4c\u5165\u5c42: token id -> embed_dim \u7ef4\u8bcd\u5411\u91cf
        # padding_idx=pad_id \u8ba9 <PAD> \u7684\u5411\u91cf\u6052\u4e3a\u96f6\uff0c\u4e14\u4e0d\u53c2\u4e0e\u66f4\u65b0
        self.embedding = nn.Embedding(vocab_size, embed_dim, padding_idx=pad_id)
        
        # RNN \u5c42: \u6cbf\u5e8f\u5217\u9010\u6b65\u66f4\u65b0\u9690\u85cf\u72b6\u6001
        # batch_first=True -> \u8f93\u5165\u5f62\u72b6\u4e3a [batch, seq_len, embed_dim]
        self.rnn = nn.RNN(embed_dim, hidden_dim, batch_first=True)
        
        # \u8f93\u51fa\u5c42: \u628a\u6bcf\u4e00\u6b65\u7684\u9690\u85cf\u72b6\u6001\u6620\u5c04\u6210\u8bcd\u8868\u4e0a\u7684\u5206\u6570 logits
        self.output = nn.Linear(hidden_dim, vocab_size)
\u200b
    def forward(self, token_ids):
        # token_ids: [batch, seq_len]
        embeddings = self.embedding(token_ids)  # [batch, seq_len, embed_dim]
        states, _ = self.rnn(embeddings)        # [batch, seq_len, hidden_dim]
        return self.output(states)              # [batch, seq_len, vocab_size]<\/pre>\n\n\n\n
3.\u8bad\u7ec3<\/p>\n\n\n\n
\u6838\u5fc3:\u628a\u6bcf\u4e2a\u4f4d\u7f6e\u5f53\u6210\u4e00\u9053\u5206\u7c7b\u9898,\u7528\u4ea4\u53c9\u71b5\u53cd\u590d\u524d\u5411\u53cd\u5411\u66f4\u65b0\u53c2\u6570<\/p>\n\n\n\n
import torch
import torch.nn as nn
\u200b
# \u4ea4\u53c9\u71b5\u635f\u5931; ignore_index=pad_id \u8ba9\u8865\u9f50\u4f4d\u7f6e\u4e0d\u8ba1\u5165\u635f\u5931
criterion = nn.CrossEntropyLoss(ignore_index=pad_id)
\u200b
# Adam \u4f18\u5316\u5668\uff0c\u5b66\u4e60\u7387 0.01
optimizer = torch.optim.Adam(
    model.parameters(), lr=0.01)
\u200b
for epoch in range(1, 10001):
    logits = model(inputs)                # \u524d\u5411: [B, T, V] = [batch, seq_len, vocab_size]
    
    # \u628a logits \u644a\u5e73\u6210 [B*T, V]\u3001\u76ee\u6807\u644a\u5e73\u6210 [B*T]
    # \u4e8e\u662f\u6bcf\u4e2a\u65f6\u95f4\u6b65\u90fd\u6210\u4e3a\u4e00\u9053\u72ec\u7acb\u7684\u201c\u4e0b\u4e00\u8bcd\u5206\u7c7b\u9898\u201d
    loss = criterion(
        logits.reshape(-1, len(vocab)),
        targets.reshape(-1))
    
    optimizer.zero_grad()                 # \u6e05\u7a7a\u4e0a\u4e00\u8f6e\u7d2f\u79ef\u7684\u68af\u5ea6
    loss.backward()                       # \u53cd\u5411\u4f20\u64ad: \u68af\u5ea6\u6cbf\u65f6\u95f4\u94fe\u53cd\u5411\u4f20\u64ad
    optimizer.step()                      # \u6309\u68af\u5ea6\u66f4\u65b0\u53c2\u6570<\/pre>\n\n\n\n
4.\u9884\u6d4b\u4e0e\u751f\u6210:<\/p>\n\n\n\n
\u6838\u5fc3:\u53ea\u770b\u6700\u540e\u4e00\u6b65\u7684\u5206\u5e03,softmax\u540e\u53d6top-k,\u53cd\u590d\u53d6top-1\u8ffd\u52a0\u56de\u5e8f\u5217<\/p>\n\n\n\n
import torch
\u200b
@torch.no_grad()                         # \u9884\u6d4b\u9636\u6bb5\u4e0d\u9700\u8981\u8ba1\u7b97\u68af\u5ea6
def predict_next(prefix, top_k=5):
    # \u628a\u524d\u7f00\u91cc\u7684\u8bcd\u8f6c\u6210 id\uff0c\u4e0d\u5728\u8bcd\u8868\u91cc\u7684\u8bcd\u7528 <UNK> \u515c\u5e95
    ids = [word_to_id.get(w, word_to_id[UNK]) 
           for w in prefix.split()]
    
    # \u524d\u5411\u540e\u53ea\u53d6\u201c\u6700\u540e\u4e00\u4e2a\u65f6\u95f4\u6b65\u201d\u7684\u8f93\u51fa = \u7d27\u63a5\u524d\u7f00\u7684\u4e0b\u4e00\u8bcd\u5206\u5e03
    logits = model(torch.tensor([ids]))[0, -1].clone()
    
    logits[pad_id] = -torch.inf          # \u4e0d\u5141\u8bb8\u751f\u6210 <PAD>
    logits[word_to_id[UNK]] = -torch.inf # \u4e0d\u5141\u8bb8\u751f\u6210 <UNK>
    
    probs = torch.softmax(logits, dim=-1) # \u5206\u6570 logits \u8f6c\u6982\u7387
    v, i = torch.topk(probs, top_k)       # \u53d6\u6982\u7387\u6700\u9ad8\u7684 top_k
    
    return [(id_to_word[j.item()], p.item()) 
            for p, j in zip(v, i)]
\u200b
def generate(prefix, max_new_words=8):   # \u8d2a\u5fc3\u751f\u6210\uff1a\u6bcf\u6b21\u53d6\u6982\u7387\u6700\u9ad8\u7684\u8bcd
    words = prefix.split()
    for _ in range(max_new_words):
        # \u53d6\u5f53\u524d\u6982\u7387\u6700\u9ad8\u7684\u8bcd\uff0c\u8ffd\u52a0\u5230\u672b\u5c3e\uff0c\u518d\u7528\u65b0\u5e8f\u5217\u7ee7\u7eed\u9884\u6d4b
        nxt, _ = predict_next(\" \".join(words), 1)[0]
        if nxt == EOS: break             # \u751f\u6210\u5230 <EOS> \u505c\u6b62
        words.append(nxt)
    return \" \".join(words)<\/pre>\n\n\n\n
RNN\u7684\u5c40\u9650:<\/h3>\n\n\n\n
\u4e0d\u7ba1\u5e8f\u5217\u591a\u957f,rnn\u6bcf\u4e00\u6b65\u90fd\u8981\u628a\u5f53\u524d\u8f93\u5165\u548c\u5168\u90e8\u8fc7\u53bb\u91cd\u65b0\u538b\u7f29\u8fdb\u540c\u4e00\u4e2a\u56fa\u5b9a\u7ef4\u5ea6\u7684\u9690\u85cf\u72b6\u6001\u91cc<\/p>\n\n\n\n
\u5e8f\u5217\u53d8\u957f\u540e,\u65e9\u671f\u4fe1\u606f\u66f4\u5bb9\u6613\u88ab\u540e\u7eed\u8f93\u5165\u8986\u76d6<\/p>\n\n\n\n
\"image-20261001154933779\"\/<\/div><\/figure>\n\n\n\n
\u66ff\u4ee3\u65b9\u6848:<\/p>\n\n\n\n
\"image-20261001154959842\"\/<\/div><\/figure>\n\n\n\n
\"image-20261001155015303\"\/<\/div><\/figure>\n\n\n\n
rnn\u7f3a\u70b9\u6f14\u793a:<\/p>\n\n\n\n
\"image-20261001155253892\"\/<\/div><\/figure>\n\n\n\n
\u8ddd\u79bb\u6bcf\u589e\u52a0\u4e00\u6b65,\u72b6\u6001\u5c31\u591a\u7ecf\u5386\u4e00\u6b21\u91cd\u5199<\/p>\n\n\n\n
\u540c\u4e00\u6761\u65f6\u95f4\u94fe,\u540c\u65f6\u9650\u5236\u4fe1\u606f\u4e0e\u68af\u5ea6:\u65e9\u671f\u4fe1\u606f\u9010\u6e10\u51cf\u5f31;\u8bad\u7ec3\u4fe1\u53f7\u88ab\u53cd\u590d\u7f29\u653e<\/p>\n\n\n\n
lstm<\/h3>\n\n\n\n
\"image-20261004201400211\"\/<\/div><\/figure>\n\n\n\n
\u9057\u5fd8\u95e8\u51b3\u5b9a\u65e7\u8bb0\u5fc6\u4fdd\u7559\u591a\u5c11<\/h3>\n\n\n\n
\u9057\u5fd8\u95e8\uff08forget gate\uff09<\/strong>\u8bfb\u53d6\u5f53\u524d\u8f93\u5165\"img\"\u4e0e\u4e0a\u4e00\u6b65\u9690\u85cf\u72b6\u6001\"img\"\uff0c\u8ba1\u7b97\u9057\u5fd8\u95e8\u503c\"img\"<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u5176\u4e2d\"img\"\u8868\u793a\u5411\u91cf\u62fc\u63a5\uff0c\"img\"\u548c\"img\"\u662f\u9057\u5fd8\u95e8\u7684\u53c2\u6570\u3002\u9057\u5fd8\u95e8\u4e0d\u4f1a\u76f4\u63a5\u751f\u6210\u65b0\u5185\u5bb9\uff0c\u53ea\u8d1f\u8d23\u63a7\u5236\u65e7\u7ec6\u80de\u72b6\u6001\"img\"\u6709\u591a\u5c11\u53ef\u4ee5\u7ee7\u7eed\u8fdb\u5165\u5f53\u524d\u72b6\u6001\u3002<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\"img\"\u8868\u793a\u9010\u5143\u7d20\u4e58\u6cd5\uff0c\u5982\u679c\u67d0\u4e9b\u7ef4\u5ea6\u7684\"img\"\u63a5\u8fd1 1\uff0c\u5bf9\u5e94\u7684\u65e7\u4fe1\u606f\u4f1a\u88ab\u4fdd\u7559\uff1b\u5982\u679c\u63a5\u8fd1 0\uff0c\u5bf9\u5e94\u4fe1\u606f\u4f1a\u88ab\u5feb\u901f\u6e05\u9664\u3002\u4e0e\u666e\u901a RNN \u5bf9\u5168\u90e8\u72b6\u6001\u7edf\u4e00\u91cd\u5199\u76f8\u6bd4\uff0cLSTM \u5f00\u59cb\u5177\u5907\u9010\u7ef4\u5ea6\u4fdd\u7559\u5386\u53f2\u4fe1\u606f\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n
\u8f93\u5165\u95e8\u51b3\u5b9a\u65b0\u8bb0\u5fc6\u5199\u5165\u591a\u5c11<\/h3>\n\n\n\n
\u53ea\u4f1a\u5fd8\u8bb0\u8fd8\u4e0d\u591f\uff0c\u6a21\u578b\u8fd8\u9700\u8981\u51b3\u5b9a\u5f53\u524d\u8f93\u5165\u4e2d\u7684\u54ea\u4e9b\u5185\u5bb9\u503c\u5f97\u5199\u5165\u957f\u671f\u72b6\u6001\u3002LSTM \u5148\u751f\u6210\u4e00\u4efd\u5019\u9009\u8bb0\u5fc6<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u5019\u9009\u8bb0\u5fc6\"img\"\u5305\u542b\u5f53\u524d\u65f6\u95f4\u6b65\u53ef\u80fd\u5199\u5165\u7684\u65b0\u5185\u5bb9\u3002\u968f\u540e\uff0c\u8f93\u5165\u95e8\uff08input gate\uff09<\/strong>\u8ba1\u7b97\u6bcf\u4e00\u7ef4\u7684\u5199\u5165\u6bd4\u4f8b<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u771f\u6b63\u5199\u5165\u72b6\u6001\u7684\u5185\u5bb9\u662f<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u8f93\u5165\u95e8\u4e0e\u5019\u9009\u8bb0\u5fc6\u5206\u5de5\u660e\u786e\uff0c\u5019\u9009\u8bb0\u5fc6\u56de\u7b54\u5f53\u524d\u53ef\u4ee5\u5199\u5165\u4ec0\u4e48\uff0c\u8f93\u5165\u95e8\u56de\u7b54\u8fd9\u4e9b\u5185\u5bb9\u5b9e\u9645\u5199\u5165\u591a\u5c11\u3002\u4e24\u8005\u7ed3\u5408\u4ee5\u540e\uff0cLSTM \u624d\u80fd\u907f\u514d\u6bcf\u4e2a\u65b0 token \u90fd\u5927\u5e45\u6539\u53d8\u957f\u671f\u72b6\u6001\u3002<\/p>\n\n\n\n
\u7ec6\u80de\u72b6\u6001\u7528\u52a0\u6cd5\u5b8c\u6210\u66f4\u65b0<\/h3>\n\n\n\n
\u4fdd\u7559\u4e0b\u6765\u7684\u65e7\u8bb0\u5fc6\u4e0e\u7b5b\u9009\u540e\u7684\u65b0\u8bb0\u5fc6\u76f8\u52a0\uff0c\u5f62\u6210\u5f53\u524d\u7ec6\u80de\u72b6\u6001<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u8fd9\u4e2a\u516c\u5f0f\u662f LSTM \u6700\u91cd\u8981\u7684\u7ed3\u6784\uff0cRNN \u7528\u4e00\u4e2a\u975e\u7ebf\u6027\u51fd\u6570\u6574\u4f53\u751f\u6210\"img\"\uff0cLSTM \u5219\u628a\u72b6\u6001\u66f4\u65b0\u62c6\u6210\u4e24\u90e8\u5206\uff0c\u4e00\u90e8\u5206\u6cbf\u65e7\u72b6\u6001\u7ee7\u7eed\u4f20\u9012\uff0c\u53e6\u4e00\u90e8\u5206\u4f5c\u4e3a\u65b0\u5185\u5bb9\u52a0\u5165\u3002<\/p>\n\n\n\n
\u8f93\u51fa\u95e8\u51b3\u5b9a\u5f53\u524d\u66b4\u9732\u54ea\u4e9b\u8bb0\u5fc6<\/h3>\n\n\n\n
\u7ec6\u80de\u72b6\u6001\u4fdd\u5b58\u4e86\u957f\u671f\u4fe1\u606f\uff0c\u4f46\u5f53\u524d\u65f6\u95f4\u6b65\u4e0d\u4e00\u5b9a\u9700\u8981\u628a\u5176\u4e2d\u6240\u6709\u5185\u5bb9\u90fd\u4ea4\u7ed9\u4efb\u52a1\u5934\u6216\u4e0b\u4e00\u4e2a\u65f6\u95f4\u6b65\u3002\u8f93\u51fa\u95e8\uff08output gate\uff09<\/strong>\u8d1f\u8d23\u63a7\u5236\u5f53\u524d\u72b6\u6001\u5bf9\u5916\u66b4\u9732\u591a\u5c11<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u6700\u7ec8\u9690\u85cf\u72b6\u6001\u4e3a<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
Tanh \u628a\u7ec6\u80de\u72b6\u6001\u538b\u7f29\u5230\"img\"\uff0c\u8f93\u51fa\u95e8\u518d\u9010\u7ef4\u7b5b\u9009\uff0c\u5f97\u5230\u5f53\u524d\u9690\u85cf\u72b6\u6001\"img\"\u3002\u56e0\u6b64\uff0c\u67d0\u4efd\u4fe1\u606f\u53ef\u4ee5\u7ee7\u7eed\u4fdd\u5b58\u5728\"img\"\u4e2d\uff0c\u5374\u6682\u65f6\u4e0d\u51fa\u73b0\u5728\"img\"\u91cc\uff1b\u7b49\u540e\u9762\u4e0a\u4e0b\u6587\u9700\u8981\u5b83\u65f6\uff0c\u518d\u901a\u8fc7\u8f93\u51fa\u95e8\u91cd\u65b0\u53c2\u4e0e\u9884\u6d4b\u3002<\/p>\n\n\n\n
\u628a LSTM \u7684\u4e00\u6b21\u5b8c\u6574\u66f4\u65b0\u653e\u5728\u4e00\u8d77\uff0c\u53ef\u4ee5\u5199\u6210<\/p>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u4e09\u9053\u95e8\u548c\u5019\u9009\u8bb0\u5fc6\u5404\u6709\u81ea\u5df1\u7684\u53c2\u6570\uff0c\u4f46\u5b83\u4eec\u6bcf\u4e2a\u65f6\u95f4\u6b65\u4ecd\u7136\u5171\u4eab\u4f7f\u7528\u3002LSTM \u6ca1\u6709\u6539\u53d8 RNN \u6309\u987a\u5e8f\u9012\u63a8\u7684\u57fa\u672c\u6846\u67b6\uff0c\u5b83\u6539\u53d8\u7684\u662f\u6bcf\u4e00\u6b65\u5982\u4f55\u66f4\u65b0\u72b6\u6001\u3002<\/p>\n\n\n\n
GRU \u5de5\u4f5c\u6d41\u7a0b<\/h3>\n\n\n\n
GRU \u5728\u7b2c\"img\"\u4e2a\u65f6\u95f4\u6b65\u63a5\u6536\u5f53\u524d\u8f93\u5165\"img\"\u548c\u4e0a\u4e00\u6b65\u9690\u85cf\u72b6\u6001\"img\"\u3002\u5b83\u5148\u6839\u636e\u8fd9\u4e24\u4e2a\u5411\u91cf\u8ba1\u7b97\u91cd\u7f6e\u95e8\"img\"\u4e0e\u66f4\u65b0\u95e8\"img\"\uff0c\u518d\u5206\u4e24\u6b65\u5b8c\u6210\u72b6\u6001\u66f4\u65b0\u3002<\/p>\n\n\n\n
\u7b2c\u4e00\u6b65\u662f\u51c6\u5907\u4e00\u4efd\u5019\u9009\u72b6\u6001\u3002\u91cd\u7f6e\u95e8\u5148\u7b5b\u9009\"img\"\u4e2d\u54ea\u4e9b\u5386\u53f2\u4fe1\u606f\u9700\u8981\u53c2\u4e0e\u5019\u9009\u72b6\u6001\u7684\u8ba1\u7b97\uff0c\u518d\u628a\u7b5b\u9009\u540e\u7684\u5386\u53f2\u4e0e\"img\"\u7ed3\u5408\uff0c\u5f97\u5230\u53ef\u80fd\u5199\u5165\u7684\u65b0\u5185\u5bb9\"img\"\u3002\u7b2c\u4e8c\u6b65\u624d\u662f\u771f\u6b63\u66f4\u65b0\u72b6\u6001\uff0c\u66f4\u65b0\u95e8\u5728\u65e7\u72b6\u6001\"img\"\u548c\u5019\u9009\u72b6\u6001\"img\"\u4e4b\u95f4\u5206\u914d\u6bd4\u4f8b\uff0c\u5f97\u5230\u6700\u7ec8\u7684\"img\"\u3002<\/p>\n\n\n\n
\u91cd\u7f6e\u95e8\u51b3\u5b9a\u751f\u6210\u65b0\u5185\u5bb9\u65f6\u53c2\u8003\u591a\u5c11\u8fc7\u53bb\uff0c\u66f4\u65b0\u95e8\u51b3\u5b9a\u751f\u6210\u7684\u65b0\u5185\u5bb9\u6700\u7ec8\u5199\u5165\u591a\u5c11\u3002<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u91cd\u7f6e\u95e8\u63a7\u5236\u8ba1\u7b97\u5019\u9009\u72b6\u6001\u65f6\u53c2\u8003\u591a\u5c11\u8fc7\u53bb<\/h3>\n\n\n\n
GRU \u5148\u8ba1\u7b97\u91cd\u7f6e\u95e8<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u7136\u540e\u7528\u91cd\u7f6e\u95e8\u7b5b\u9009\u4e0a\u4e00\u6b65\u9690\u85cf\u72b6\u6001\uff0c\u518d\u7ed3\u5408\u5f53\u524d\u8f93\u5165\u751f\u6210\u5019\u9009\u72b6\u6001<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u5f53\"img\"\u7684\u67d0\u4e9b\u7ef4\u5ea6\u63a5\u8fd1 0 \u65f6\uff0c\u8ba1\u7b97\u5019\u9009\u72b6\u6001\u4f1a\u8f83\u5c11\u53c2\u8003\u5bf9\u5e94\u7684\u5386\u53f2\u4fe1\u606f\uff0c\u66f4\u96c6\u4e2d\u5730\u6839\u636e\u5f53\u524d\u8f93\u5165\u5f62\u6210\u65b0\u8868\u793a\uff1b\u5f53\u5b83\u63a5\u8fd1 1 \u65f6\uff0c\u8fc7\u53bb\u72b6\u6001\u4f1a\u66f4\u5145\u5206\u5730\u53c2\u4e0e\u5019\u9009\u72b6\u6001\u8ba1\u7b97\u3002<\/p>\n\n\n\n
\u91cd\u7f6e\u95e8\u53ea\u5f71\u54cd\u5019\u9009\u72b6\u6001\u5982\u4f55\u751f\u6210\uff0c\u5e76\u4e0d\u76f4\u63a5\u51b3\u5b9a\u6700\u7ec8\u72b6\u6001\u4e22\u5f03\u591a\u5c11\u5386\u53f2\u3002\u6700\u7ec8\u4fdd\u7559\u65e7\u72b6\u6001\u8fd8\u662f\u91c7\u7528\u65b0\u5019\u9009\uff0c\u7531\u66f4\u65b0\u95e8\u8d1f\u8d23\u3002<\/p>\n\n\n\n
\u66f4\u65b0\u95e8\u5728\u65e7\u72b6\u6001\u4e0e\u5019\u9009\u72b6\u6001\u4e4b\u95f4\u63d2\u503c<\/h3>\n\n\n\n
\u66f4\u65b0\u95e8\u7684\u8ba1\u7b97\u65b9\u5f0f\u662f<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
GRU \u7684\u5f53\u524d\u72b6\u6001\u5199\u6210<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u5f53\"img\"\u63a5\u8fd1 1 \u65f6\uff0cGRU \u4e3b\u8981\u4fdd\u7559\u65e7\u72b6\u6001\"img\"\uff1b\u5f53\"img\"\u63a5\u8fd1 0 \u65f6\uff0c\u5219\u4e3b\u8981\u91c7\u7528\u65b0\u5019\u9009\u72b6\u6001\"img\"\u3002\u66f4\u65b0\u95e8\u540c\u65f6\u627f\u62c5\u4e86 LSTM \u4e2d\u90e8\u5206\u9057\u5fd8\u95e8\u4e0e\u8f93\u5165\u95e8\u7684\u4f5c\u7528\uff0c\u56e0\u4e3a\u65e7\u72b6\u6001\u548c\u65b0\u72b6\u6001\u7684\u6743\u91cd\u88ab\u7ed1\u5b9a\u6210\"img\"\u4e0e\"img\"\uff0c\u4e24\u8005\u6b64\u6d88\u5f7c\u957f\u3002<\/p>\n\n\n\n
\u4e0d\u540c\u8d44\u6599\u6709\u65f6\u4f1a\u4ea4\u6362\"img\"\u4e0e\"img\"\u7684\u8bb0\u53f7\uff0c\u628a\"img\"\u5b9a\u4e49\u6210\u65b0\u4fe1\u606f\u7684\u5199\u5165\u6bd4\u4f8b\u3002\u4e24\u79cd\u5199\u6cd5\u8868\u8fbe\u7684\u662f\u540c\u4e00\u4ef6\u4e8b\uff0c\u9605\u8bfb\u516c\u5f0f\u65f6\u9700\u8981\u5148\u786e\u8ba4\u66f4\u65b0\u95e8\u7a76\u7adf\u63a7\u5236\u65e7\u72b6\u6001\u8fd8\u662f\u5019\u9009\u72b6\u6001\u3002\u672c\u6587\u6cbf\u7528 PyTorch \u5bf9\u6700\u7ec8\u72b6\u6001\u66f4\u65b0\u7684\u8bb0\u53f7\uff0c\u8ba9\"img\"\u8868\u793a\u65e7\u72b6\u6001\u7684\u4fdd\u7559\u6bd4\u4f8b\u3002<\/p>\n\n\n\n
\u628a GRU \u7684\u4e00\u6b21\u66f4\u65b0\u653e\u5728\u4e00\u8d77\uff0c\u5c31\u662f<\/p>\n\n\n\n
\"\"<\/div><\/figure>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
GRU \u6ca1\u6709\u72ec\u7acb\u7684\u8f93\u51fa\u95e8\uff0c\"img\"\u65e2\u627f\u62c5\u8de8\u65f6\u95f4\u6b65\u4f20\u9012\u4fe1\u606f\u7684\u4efb\u52a1\uff0c\u4e5f\u662f\u5f53\u524d\u65f6\u95f4\u6b65\u5411\u5916\u63d0\u4f9b\u7684\u72b6\u6001\u8868\u793a\u3002\u5b83\u5bf9\u72b6\u6001\u7684\u63a7\u5236\u6ca1\u6709 LSTM \u90a3\u6837\u7ec6\u5206\uff0c\u4f46\u8ba1\u7b97\u6b65\u9aa4\u66f4\u5c11\u3001\u53c2\u6570\u91cf\u901a\u5e38\u4e5f\u66f4\u5c0f\u3002<\/p>\n\n\n\n
GRU \u901a\u5e38\u80fd\u7528\u6bd4 LSTM \u66f4\u5c11\u7684\u53c2\u6570\u548c\u8ba1\u7b97\u53d6\u5f97\u63a5\u8fd1\u7684\u6548\u679c\uff0c\u56e0\u6b64\u6210\u4e3a\u5faa\u73af\u7f51\u7edc\u4e2d\u5e38\u7528\u7684\u8f7b\u91cf\u5316\u9009\u62e9\u3002<\/p>\n\n\n\n
Highway Network \u628a\u95e8\u63a7\u5e26\u5230\u524d\u9988\u7f51\u7edc<\/h3>\n\n\n\n
2015 \u5e74\u521d\u53d7 LSTM \u542f\u53d1\u8bde\u751f\u7684 Highway Network<\/strong> \u628a\u95e8\u63a7\u601d\u8def\u4ece\u65f6\u95f4\u7ef4\u5ea6\u5e26\u5230\u4e86\u6df1\u5c42\u524d\u9988\u7f51\u7edc\u3002\u4e00\u4e2a Highway \u5c42\u53ef\u4ee5\u5199\u6210<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u5176\u4e2d\"img\"\u662f\u53d8\u6362\u540e\u7684\u65b0\u7279\u5f81\uff0c\"img\"\u662f\u53d8\u6362\u95e8\uff08transform gate\uff09<\/strong>\uff0c\u8fd9\u4e2a\u5f62\u5f0f\u4e0e GRU \u7684\u72b6\u6001\u66f4\u65b0\u975e\u5e38\u63a5\u8fd1\u3002\u7f51\u7edc\u53ef\u4ee5\u6839\u636e\u5f53\u524d\u8f93\u5165\uff0c\u51b3\u5b9a\u54ea\u4e9b\u7ef4\u5ea6\u91c7\u7528\u65b0\u53d8\u6362\uff0c\u54ea\u4e9b\u7ef4\u5ea6\u7ee7\u7eed\u643a\u5e26\u539f\u6765\u7684\u7279\u5f81\u3002\u95e8\u63a7\u7684\u4fdd\u7559\u4e0e\u66f4\u65b0\u673a\u5236\u7531\u6b64\u4ece\u65f6\u95f4\u4e0a\u7684\u957f\u94fe\uff0c\u6269\u5c55\u5230\u4e86\u7f51\u7edc\u5c42\u4e4b\u95f4\u7684\u6df1\u94fe\u3002<\/p>\n\n\n\n
ResNet \u53bb\u6389\u95e8\u63a7\uff0c\u4fdd\u7559\u76f4\u63a5\u8def\u5f84<\/h3>\n\n\n\n
\u968f\u540e 2015 \u5e74\u5e95\u51fa\u73b0\u7684 ResNet \u8fdb\u4e00\u6b65\u7b80\u5316\u4e86\u8fd9\u6761\u601d\u8def\uff0c\u5b83\u4e0d\u518d\u4e3a\u4fdd\u7559\u8def\u5f84\u8ba1\u7b97\u95e8\u503c\uff0c\u800c\u662f\u7528\u6052\u7b49\u6620\u5c04\u4f5c\u4e3a\u6377\u5f84\uff0c\u53ea\u8ba9\u4e3b\u8def\u5f84\u5b66\u4e60\u9700\u8981\u8865\u5145\u7684\u6b8b\u5dee\"img\"<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
Highway Network \u9700\u8981\u5b66\u4e60\u539f\u4fe1\u606f\u4fdd\u7559\u591a\u5c11\uff0cResNet \u5219\u9ed8\u8ba4\u539f\u4fe1\u606f\u5b8c\u6574\u901a\u8fc7\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u5df2\u6709\u8868\u793a\u59cb\u7ec8\u62e5\u6709\u4e00\u6761\u4e0d\u7ecf\u8fc7\u590d\u6742\u53d8\u6362\u7684\u76f4\u63a5\u8def\u5f84\uff0c\u4e3b\u8def\u5f84\u53ea\u9700\u8981\u5728\u6b64\u57fa\u7840\u4e0a\u5b66\u4e60\u5fc5\u8981\u7684\u4fee\u6b63\u3002<\/p>\n\n\n\n
\u4ece LSTM\u3001GRU \u5230 Highway Network\uff0c\u518d\u5230 ResNet\uff0c\u80cc\u540e\u7684\u4f18\u5316\u601d\u8def\u975e\u5e38\u7c7b\u4f3c\uff0c\u6df1\u5c42\u6216\u957f\u94fe\u8ba1\u7b97\u4e0d\u5e94\u8be5\u5f3a\u8feb\u4fe1\u606f\u6bcf\u4e00\u6b65\u90fd\u88ab\u5b8c\u6574\u91cd\u5199\uff0c\u7ed9\u5df2\u6709\u8868\u793a\u4fdd\u7559\u4e00\u6761\u66f4\u76f4\u63a5\u7684\u8def\u5f84\uff0c\u53ef\u4ee5\u540c\u65f6\u6539\u5584\u4fe1\u606f\u4f20\u9012\u4e0e\u68af\u5ea6\u4f20\u64ad\u3002<\/p>\n\n\n\n
Highway Network \u7684\u63d0\u51fa\u8005 Srivastava \u7b49\u4eba\u5728\u540e\u7eed\u8bba\u6587\u4e2d\u660e\u786e\u6307\u51fa\uff0cResNet \u53ef\u4ee5\u88ab\u89c6\u4e3a Highway Network \u7684\u4e00\u79cd\u7279\u4f8b\u6216\u7b80\u5316\u7248\u672c\u3002\u4e24\u8005\u90fd\u662f\u4e3a\u4e86\u89e3\u51b3\u6df1\u5c42\u7f51\u7edc\u8bad\u7ec3\u56f0\u96be\uff08\u68af\u5ea6\u6d88\u5931\/\u9000\u5316\uff09\u7684\u95ee\u9898\uff0c\u6838\u5fc3\u601d\u60f3\u90fd\u662f\u8ba9\u4fe1\u606f\u80fd\u591f\u65e0\u635f\u5730\u7a7f\u8fc7\u591a\u4e2a\u5c42\u3002<\/p>\n\n\n\n
LSTM \u548c GRU \u4ecd\u7136\u53d7\u9650\u4e8e\u5faa\u73af\u7ed3\u6784<\/h2>\n\n\n\n
LSTM \u4e0e GRU \u663e\u8457\u6539\u5584\u4e86\u666e\u901a RNN \u7684\u957f\u671f\u4f9d\u8d56\u95ee\u9898\uff0c\u4f46\u5b83\u4eec\u4ecd\u7136\u5c5e\u4e8e\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u3002\u95e8\u63a7\u6539\u53d8\u4e86\u72b6\u6001\u5982\u4f55\u66f4\u65b0\uff0c\u6ca1\u6709\u6539\u53d8\u65f6\u95f4\u6b65\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb<\/p>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\u7b2c\"img\"\u6b65\u5fc5\u987b\u7b49\u7b2c\"img\"\u6b65\u5b8c\u6210\uff0c\u5355\u6761\u5e8f\u5217\u5185\u90e8\u65e0\u6cd5\u540c\u65f6\u8ba1\u7b97\u6240\u6709\u4f4d\u7f6e\u3002GPU \u64c5\u957f\u5927\u89c4\u6a21\u5e76\u884c\u77e9\u9635\u8fd0\u7b97\uff0c\u4f46\u5faa\u73af\u7ed3\u6784\u53ea\u80fd\u6cbf\u65f6\u95f4\u9010\u6b65\u63a8\u8fdb\uff1b\u5e8f\u5217\u8d8a\u957f\uff0c\u524d\u5411\u4e0e\u53cd\u5411\u4f20\u64ad\u6240\u9700\u7684\u4e32\u884c\u6b65\u9aa4\u5c31\u8d8a\u591a\uff0c\u8fd9\u6210\u4e3a LSTM \u548c GRU \u5728\u5927\u89c4\u6a21\u6587\u672c\u8bad\u7ec3\u4e2d\u7684\u4e3b\u8981\u6548\u7387\u74f6\u9888\u3002<\/p>\n\n\n\n
\u95e8\u63a7\u4e5f\u6ca1\u6709\u5f7b\u5e95\u6d88\u9664\u957f\u8ddd\u79bb\u4f20\u64ad\uff0cLSTM \u7ec6\u80de\u72b6\u6001\u7684\u68af\u5ea6\u4ecd\u4f1a\u8fde\u7eed\u4e58\u4e0a\u9057\u5fd8\u95e8\uff0cGRU \u7684\u65e7\u72b6\u6001\u8def\u5f84\u4e5f\u4f1a\u8fde\u7eed\u53d7\u5230\u66f4\u65b0\u95e8\u5f71\u54cd\u3002\u5982\u679c\u6709\u6548\u8ddd\u79bb\u8fbe\u5230\u6570\u767e\u751a\u81f3\u6570\u5343\u4e2a token\uff0c\u957f\u671f\u4fe1\u606f\u4ecd\u53ef\u80fd\u9010\u6e10\u8870\u51cf\u3002<\/p>\n\n\n\n
\u56fa\u5b9a\u7ef4\u5ea6\u72b6\u6001\u5e26\u6765\u7684\u4fe1\u606f\u74f6\u9888\u540c\u6837\u8fd8\u5728\uff0cLSTM \u628a\u957f\u671f\u72b6\u6001\u4e0e\u5f53\u524d\u8f93\u51fa\u5206\u5f00\uff0cGRU \u63d0\u4f9b\u4e86\u66f4\u7075\u6d3b\u7684\u72b6\u6001\u66f4\u65b0\uff0c\u5374\u4ecd\u8981\u628a\u622a\u81f3\u5f53\u524d\u4f4d\u7f6e\u7684\u5386\u53f2\u538b\u7f29\u8fdb\u56fa\u5b9a\u7ef4\u5ea6\u5411\u91cf\u3002\u5e8f\u5217\u4e0d\u65ad\u589e\u957f\u65f6\uff0c\u6a21\u578b\u4ecd\u7136\u9700\u8981\u51b3\u5b9a\u4fdd\u7559\u4ec0\u4e48\u3001\u820d\u5f03\u4ec0\u4e48\uff0c\u65e0\u6cd5\u968f\u65f6\u76f4\u63a5\u8bbf\u95ee\u4efb\u610f\u4e00\u4e2a\u65e9\u671f token \u7684\u539f\u59cb\u8868\u793a\u3002<\/p>\n\n\n\n
\u4e0e\u5176\u8981\u6c42\u4e00\u4e2a\u5b9a\u957f\u72b6\u6001\u628a\u6240\u6709\u5386\u53f2\u4e00\u8def\u643a\u5e26\u5230\u5f53\u524d\u4f4d\u7f6e\uff0c\u4e0d\u5982\u8ba9\u6a21\u578b\u5728\u9700\u8981\u65f6\u76f4\u63a5\u67e5\u770b\u76f8\u5173\u4f4d\u7f6e\u3002\u8fd9\u5c31\u662f\u540e\u7eed\u6f14\u8fdb\u7684 Attention \u673a\u5236\u3002Attention \u7f29\u77ed\u4e86\u4fe1\u606f\u8bbf\u95ee\u8def\u5f84\uff0cTransformer \u7684 Self Attention \u53c8\u8fdb\u4e00\u6b65\u53bb\u6389\u65f6\u95f4\u6b65\u9012\u63a8\uff0c\u8ba9\u540c\u4e00\u5c42\u4e2d\u7684\u6240\u6709 token \u53ef\u4ee5\u5e76\u884c\u8ba1\u7b97\u3002<\/p>\n\n\n\n
\u603b\u7ed3\uff1aRNN\u3001LSTM \u4e0e GRU \u5bf9\u6bd4<\/h2>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
\"img\"\/<\/div><\/figure>\n\n\n\n
RNN \u8bc1\u660e\u4e86\u795e\u7ecf\u7f51\u7edc\u53ef\u4ee5\u901a\u8fc7\u5171\u4eab\u53c2\u6570\u4e0e\u9690\u85cf\u72b6\u6001\u5904\u7406\u53d8\u957f\u5e8f\u5217\uff0c\u5373\u4f7f Transformer \u5df2\u7ecf\u5f88\u6210\u719f\u7684\u4eca\u5929\uff0c\u4e24\u8005\u5728\u8bed\u97f3\u3001\u4f20\u611f\u5668\u3001\u65f6\u95f4\u5e8f\u5217\u3001\u8fb9\u7f18\u8bbe\u5907\u548c\u4f4e\u5ef6\u8fdf\u6d41\u5f0f\u4efb\u52a1\u4e2d\uff0c\u4ecd\u7136\u5177\u6709\u660e\u786e\u4f18\u52bf\u3002\u5b83\u4eec\u7684\u5c40\u9650\u4e3b\u8981\u51fa\u73b0\u5728\u9700\u8981\u8d85\u957f\u4e0a\u4e0b\u6587\u3001\u5927\u89c4\u6a21\u5e76\u884c\u8bad\u7ec3\u548c\u4efb\u610f\u4f4d\u7f6e\u76f4\u63a5\u4ea4\u4e92\u7684\u573a\u666f<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e00.\u89e3\u51b3\u95ee\u9898: \u6587\u672c\u6a21\u578b\u4e0d\u4ec5\u8981\u77e5\u9053\u6709\u54ea\u4e9b\u8bcd,\u8fd8\u8981\u5229\u7528\u8bcd\u5e8f\u548c\u4e0a\u4e0b\u6587 \u4e3a\u4ec0\u4e48\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u4e0d\u9002\u5408\u5904\u7406\u81ea\u7136\u8bed\u8a00\u5e8f\u5217:  […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-149","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/149","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=149"}],"version-history":[{"count":2,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/149\/revisions"}],"predecessor-version":[{"id":153,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/149\/revisions\/153"}],"wp:attachment":[{"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=149"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=149"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/luqimu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=149"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}