序列到序列模型

import torch
import torch.nn as nn


class Encoder(nn.Module):
    def __init__(self, src_vocab_size, embed_dim, hidden_dim):
        super().__init__()
        self.embed = nn.Embedding(src_vocab_size, embed_dim, padding_idx=0)
        self.gru = nn.GRU(embed_dim, hidden_dim, batch_first=True)

    def forward(self, src):
        e = self.embed(src)
        outputs, hidden = self.gru(e)
        return outputs, hidden

class Decoder(nn.Module):
    def __init__(self, tgt_vocab_size, embed_dim, hidden_dim):
        super().__init__()
        self.embed = nn.Embedding(tgt_vocab_size, embed_dim, padding_idx=0)
        self.gru = nn.GRU(embed_dim, hidden_dim, batch_first=True)
        self.fc = nn.Linear(hidden_dim, tgt_vocab_size)

    def forward(self, token, hidden):
        e = self.embed(token)
        out, hidden = self.gru(e, hidden)
        logits = self.fc(out)
        return logits, hidden

def train_batch(encoder, decoder, src, tgt, bos_id, optimizer, teacher_forcing_ratio=0.9):
    optimizer.zero_grad()
    _, hidden = encoder(src)
    batch_size, tgt_len = tgt.shape
    input_token = torch.full((batch_size, 1), bos_id, dtype=torch.long)
    loss = 0.0
    loss_fn = nn.CrossEntropyLoss(ignore_index=0)

    for t in range(tgt_len):
        logits, hidden = decoder(input_token, hidden)
        step_loss = loss_fn(logits.squeeze(1), tgt[:, t])
        loss += step_loss
        use_teacher = torch.rand(1).item() < teacher_forcing_ratio
        if use_teacher:
            input_token = tgt[:, t].unsqueeze(1)
        else:
            input_token = logits.argmax(dim=-1)

    loss.backward()
    optimizer.step()
    return loss.item() / tgt_len

@torch.no_grad()
def greedy_decode(encoder, decoder, src, bos_id, eos_id, max_len=50):
    _, hidden = encoder(src)
    batch_size = src.shape[0]
    input_token = torch.full((batch_size, 1), bos_id, dtype=torch.long)
    output_ids = []
    for _ in range(max_len):
        logits, hidden = decoder(input_token, hidden)
        next_token = logits.argmax(dim=-1)
        output_ids.append(next_token)
        input_token = next_token
        if (next_token == eos_id).all():
            break
    return torch.cat(output_ids, dim=1)

seq_len=5   copy accuracy: 98%
seq_len=10  copy accuracy: 91%
seq_len=20  copy accuracy: 62%
seq_len=40  copy accuracy: 23%

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("facebook/bart-base")
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/bart-base")

src = tok("Translate this to French: Hello, how are you?", return_tensors="pt")
out = model.generate(**src, max_new_tokens=50, num_beams=4)
print(tok.decode(out[0], skip_special_tokens=True))

---
name: seq2seq-design
description: Design a sequence-to-sequence pipeline for a given task.
phase: 5
lesson: 09
---

Given a task (translation, summarization, paraphrase, question rewrite), output:

1. Architecture. Pretrained transformer encoder-decoder (BART, T5, mBART, NLLB) is the default. RNN-based seq2seq only for specific constraints.
2. Starting checkpoint. Name it (`facebook/bart-base`, `google/flan-t5-base`, `facebook/nllb-200-distilled-600M`). Match the checkpoint to task and language coverage.
3. Decoding strategy. Greedy for deterministic output, beam search (width 4-5) for quality, sampling with temperature for diversity. One sentence justification.
4. One failure mode to verify before shipping. Exposure bias manifests as generation drift on longer outputs; sample 20 outputs at the 90th-percentile length and eyeball.

Refuse to recommend training a seq2seq from scratch for under a million parallel examples. Flag any pipeline that uses greedy decoding for user-facing content as fragile (greedy repeats and loops).