第 22 章：Seq2Seq：机器翻译的旧时代

import torchimport torch.nn as nn# 1. 迷你训练数据（数字 → 拼写）pairs = [    ("123", "one two three"),    ("12", "one two"),    ("23", "two three"),    ("1", "one"),    ("3", "three"),    ("31", "three one"),    ("231", "two three one"),]# 2. 构建字符级词表src_chars = sorted(set(c for s, _ in pairs for c in s))tgt_chars = sorted(set(c for _, t in pairs for c in t) | {"<s>", "<e>"})src_vocab = {c: i for i, c in enumerate(src_chars)}tgt_vocab = {c: i for i, c in enumerate(tgt_chars)}tgt_idx = {i: c for c, i in tgt_vocab.items()}# 3. Encoder-Decoderclass Seq2Seq(nn.Module):    def __init__(self):        super().__init__()        self.enc = nn.GRU(8, 16, batch_first=True)        self.dec = nn.GRU(8, 16, batch_first=True)        self.out = nn.Linear(16, len(tgt_chars))    def forward(self, src, tgt):        _, h = self.enc(src)                    # 编码器：源序列 → 最后隐藏状态        out, _ = self.dec(tgt, h)               # 解码器：目标序列 + 编码器状态        return self.out(out)                    # 投影到词表分布model = Seq2Seq()opt = torch.optim.Adam(model.parameters(), lr=0.01)# 4. 训练（teacher forcing）for epoch in range(2000):    total_loss = 0    for src_str, tgt_str in pairs:        # 转为 one-hot 向量        src = torch.zeros(1, len(src_str), len(src_chars))        for i, c in enumerate(src_str):            src[0, i, src_vocab[c]] = 1        tgt_in = ["<s>"] + list(tgt_str)        tgt_out = list(tgt_str) + ["<e>"]        tgt = torch.zeros(1, len(tgt_in), len(tgt_chars))        for i, c in enumerate(tgt_in):            tgt[0, i, tgt_vocab[c]] = 1        tgt_label = torch.tensor([[tgt_vocab[c] for c in tgt_out]])        out = model(src, tgt)                   # 前向        loss = nn.functional.cross_entropy(            out.view(-1, len(tgt_chars)), tgt_label.view(-1))        opt.zero_grad()        loss.backward()        opt.step()        total_loss += loss.item()    if epoch % 500 == 0:        print(f"epoch {epoch:4d} loss {total_loss/len(pairs):.3f}")# 5. 推理（自由运行）model.eval()with torch.no_grad():    for src_str, _ in pairs[:3]:        src = torch.zeros(1, len(src_str), len(src_chars))        for i, c in enumerate(src_str):            src[0, i, src_vocab[c]] = 1        _, h = model.enc(src)                     # 编码        out_seq = []        token = torch.zeros(1, 1, len(tgt_chars))        token[0, 0, tgt_vocab["<s>"]] = 1        for _ in range(20):            out, h = model.dec(token, h)          # 逐步解码            nxt = out[0, -1].argmax().item()            if nxt == tgt_vocab["<e>"]:                break            out_seq.append(tgt_idx[nxt])            token = torch.zeros(1, 1, len(tgt_chars))            token[0, 0, nxt] = 1        print(f"'{src_str}' → '{' '.join(out_seq)}'")