src/train.py

"""모두의 빛길 — 학습 스크립트.

세 단계로 학습한다.

Stage 1) 슬롯 추출기 fine-tune        (KLUE/RoBERTa)
Stage 2) GNN 자기지도 사전학습          (링크 예측)
Stage 3) 추천 fine-tune                 (BPR + 멀티태스크)

GPU 설정은 사용자가 별도 셀에서 모델을 .to(device)로 옮긴 뒤 호출한다.
이 모듈 내부에서는 device를 직접 지정하지 않는다.
"""
from __future__ import annotations
from typing import Dict, List, Tuple, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F


# ============================================================
# Stage 1) 슬롯 추출기 학습
# ============================================================

def train_slot_extractor(
    model,                             # RobertaSlotExtractor
    train_examples: List[Dict],        # [{"text": ..., "labels": [...]}]
    epochs: int = 5,
    batch_size: int = 8,
    lr: float = 3e-5,
):
    """KLUE/RoBERTa fine-tuning.

    train_examples 형식
    -------------------
    [
      {
        "text": "아이와 저녁 공연을 보고 안전하게 귀가할 수 있는 동선을 추천해줘",
        "labels": ["O", "B-USER_TYPE", "O", "B-TIME_WINDOW", "B-CULTURE_PREF", ...]
      },
      ...
    ]
    """
    from .slot_extractor import LABEL2ID

    optim = torch.optim.AdamW(model.parameters(), lr=lr)

    def encode(batch):
        texts  = [b["text"] for b in batch]
        labels = [b["labels"] for b in batch]
        enc = model.tokenizer(
            texts, is_split_into_words=False,
            return_tensors="pt", padding=True, truncation=True, max_length=128,
            return_offsets_mapping=True,
        )
        offsets = enc.pop("offset_mapping")
        # offset → 라벨 정렬 (BIO를 토큰 레벨로 펴는 작업)
        # 단순화: 라벨이 토큰 단위로 이미 펴져 있다고 가정
        # 실제로는 char offset → token offset 매핑이 필요. 데이터 빌드 시 처리.
        label_ids = torch.zeros_like(enc["input_ids"]).long().fill_(-100)
        for i, lab in enumerate(labels):
            for j in range(min(len(lab), label_ids.size(1))):
                label_ids[i, j] = LABEL2ID.get(lab[j], 0)
        return enc, label_ids

    model.train()
    device = next(model.parameters()).device
    for ep in range(epochs):
        total = 0.0
        for i in range(0, len(train_examples), batch_size):
            batch = train_examples[i:i + batch_size]
            enc, lab = encode(batch)
            for k in enc:
                enc[k] = enc[k].to(device)
            lab = lab.to(device)
            out = model(input_ids=enc["input_ids"],
                        attention_mask=enc["attention_mask"],
                        labels=lab)
            loss = out.loss
            optim.zero_grad()
            loss.backward()
            optim.step()
            total += loss.item()
        print(f"[Stage1] epoch {ep+1}/{epochs}  loss={total / max(1, len(train_examples)//batch_size):.4f}")


# ============================================================
# Stage 2) GNN 자기지도 사전학습 (링크 예측)
# ============================================================

def train_gnn_link_pred(
    gnn: nn.Module,
    hetero_data,
    relations: List[Tuple[str, str, str]] = None,
    epochs: int = 50,
    lr: float = 1e-3,
    neg_ratio: int = 3,
):
    """엣지를 마스킹하고 복원하는 자기지도 학습.

    BPR 손실:  sigmoid(z_pos · z_neg)을 최소화.
    """
    if relations is None:
        relations = [
            ("venue", "near", "transit"),
            ("venue", "has_amenity", "amenity"),
            ("venue", "hosts", "event"),
        ]
    optim = torch.optim.Adam(gnn.parameters(), lr=lr)
    gnn.train()
    device = next(gnn.parameters()).device

    x_dict = {k: v.to(device) for k, v in hetero_data.x_dict.items()}
    eidx = {k: v.to(device) for k, v in hetero_data.edge_index_dict.items()}

    for ep in range(epochs):
        out = gnn(x_dict, eidx)
        loss = torch.tensor(0.0, device=device)
        for rel in relations:
            if rel not in eidx:
                continue
            src_t, _, dst_t = rel
            ei = eidx[rel]
            if ei.size(1) == 0:
                continue
            pos_src = out[src_t][ei[0]]
            pos_dst = out[dst_t][ei[1]]
            pos_score = (pos_src * pos_dst).sum(-1)

            # 음성 샘플
            num_dst = out[dst_t].size(0)
            neg_dst = torch.randint(0, num_dst, (ei.size(1) * neg_ratio,), device=device)
            neg_src = ei[0].repeat_interleave(neg_ratio)
            neg_score = (out[src_t][neg_src] * out[dst_t][neg_dst]).sum(-1)

            # BPR
            loss = loss + -F.logsigmoid(pos_score.repeat_interleave(neg_ratio) - neg_score).mean()

        optim.zero_grad()
        loss.backward()
        optim.step()
        if (ep + 1) % 10 == 0:
            print(f"[Stage2] epoch {ep+1}/{epochs}  loss={loss.item():.4f}")


# ============================================================
# Stage 3) 추천 fine-tune (BPR)
# ============================================================

def train_recommender(
    gnn: nn.Module,
    venue_ranker,                       # VenueRanker
    hetero_data,
    triples: List[Tuple[torch.Tensor, int, int]],
    # (slot_vec, pos_venue_id, neg_venue_id) 형태
    epochs: int = 30,
    lr: float = 1e-3,
):
    """슬롯 → venue ranking BPR 학습."""
    params = list(gnn.parameters()) + list(venue_ranker.parameters())
    optim = torch.optim.Adam(params, lr=lr)
    gnn.train()
    venue_ranker.train()
    device = next(gnn.parameters()).device

    x_dict = {k: v.to(device) for k, v in hetero_data.x_dict.items()}
    eidx = {k: v.to(device) for k, v in hetero_data.edge_index_dict.items()}

    for ep in range(epochs):
        total = 0.0
        for slot_vec, pos_id, neg_id in triples:
            slot_vec = slot_vec.to(device).unsqueeze(0)

            x = gnn(x_dict, eidx)
            venue_emb = x["venue"]

            scores = venue_ranker(slot_vec, venue_emb).squeeze(0)
            pos_score = scores[pos_id]
            neg_score = scores[neg_id]
            loss = -F.logsigmoid(pos_score - neg_score)

            optim.zero_grad()
            loss.backward()
            optim.step()
            total += loss.item()
        if (ep + 1) % 5 == 0:
            print(f"[Stage3] epoch {ep+1}/{epochs}  loss={total / max(1, len(triples)):.4f}")


__all__ = [
    "train_slot_extractor",
    "train_gnn_link_pred",
    "train_recommender",
]