File size: 17,031 Bytes
7ed86c3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
"""Fine-tune Stoicheia into a joint lemmatizer + XPOS tagger.

Single-process or torchrun/DDP (data-parallel over packed rows; the dataset is tiny,
every rank holds it all in RAM and takes a disjoint slice of rows each epoch).

  torchrun --nproc_per_node=4 -m tagger.train --config configs/tagger_fold0.json
"""
from __future__ import annotations

import argparse, json, math, os, sys, time
from pathlib import Path

import torch

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from tagger.backbone import load_backbone_auto
from tagger.conllu import Sentence, Token, read_conllu
from tagger.dataset import TaggerDataset, batch_chunk, encode_word, pack_dev_items
from tagger.edits import LabelVocab
from tagger.model import TaggerConfig, TaggerModel


def read_silver(path, limit_words):
    """Qwen-teacher silver lemmas (one word/line jsonl grouped by 'sentence') ->
    Sentence objects with dummy XPOS/UPOS; used for lemma-distillation pretraining."""
    sents, cur_s, cur = [], None, []
    n = 0
    with open(path, encoding="utf-8") as f:
        for line in f:
            try:
                r = json.loads(line)
            except Exception:
                continue
            s, fo, le = r.get("sentence", ""), r.get("form", ""), r.get("lemma", "")
            if not fo or not le:
                continue
            if s != cur_s:
                if cur:
                    sents.append(Sentence(tokens=cur))
                cur_s, cur = s, []
            cur.append(Token(str(len(cur) + 1), fo, le, "-", "-" * 9,
                             "_", "_", "_", "_", "_"))
            n += 1
            if n >= limit_words:
                break
    if cur:
        sents.append(Sentence(tokens=cur))
    return sents


def mask_non_lemma_labels(ds: TaggerDataset):
    """Silver has trustworthy lemmas only: keep y_script, ignore xpos/upos/full-tag."""
    for e in ds.encs:
        if e is None:
            continue
        e.y_xpos[:] = -100
        e.y_upos[:] = -100
        e.y_tag[:] = -100


def ddp_setup():
    if "RANK" in os.environ:
        import datetime
        import torch.distributed as dist
        dist.init_process_group("nccl", timeout=datetime.timedelta(minutes=60))
        rank = dist.get_rank(); world = dist.get_world_size()
        torch.cuda.set_device(rank % torch.cuda.device_count())
        return rank, world, True
    return 0, 1, False


def param_groups(model: TaggerModel, cfg):
    """Head LR flat; encoder LR with layerwise decay from the top block down.
    Decay only matrices (same convention as pretraining).

    Dispatches on the encoder type: an HF backbone (tagger.hf_backbone.HFBackboneWithHidden)
    has none of CharBertWithHidden's e_char/e_bnd/e_dia/e_punct embeddings or
    head_char/head_bnd/... pretraining heads, so it gets its own (simpler) walk in
    tagger.hf_backbone.param_groups_hf; this keeps one call site for both tagger/train.py and
    parser/joint_train.py (via `from tagger.train import param_groups as tagger_param_groups`)
    regardless of which backbone a config selects."""
    from tagger.hf_backbone import HFBackboneWithHidden, param_groups_hf
    if isinstance(model.encoder, HFBackboneWithHidden):
        return param_groups_hf(model, cfg)

    llrd = cfg.get("llrd", 0.95)
    lr_enc, lr_head, wd = cfg["lr_enc"], cfg["lr_head"], cfg.get("wd", 0.01)
    depth = len(model.encoder.blocks)
    groups = {}

    def add(p, lr, is_enc=True):
        key = (lr, 0.0 if p.ndim < 2 else wd, is_enc)
        groups.setdefault(key, []).append(p)

    enc = model.encoder
    emb_lr = lr_enc * llrd ** depth
    for m in (enc.e_char, enc.e_bnd, enc.e_dia, enc.e_punct):
        for p in m.parameters():
            add(p, emb_lr)
    for i, blk in enumerate(enc.blocks):
        for p in blk.parameters():
            add(p, lr_enc * llrd ** (depth - 1 - i))
    for p in enc.norm_out.parameters():
        add(p, lr_enc)
    # pretraining output heads ride along untrained-into-loss; give them enc LR (harmless)
    for m in (enc.head_char, enc.head_bnd, enc.head_dia, enc.head_cap, enc.head_punct):
        for p in m.parameters():
            add(p, lr_enc)
    heads = [model.xpos_heads, model.head_script, model.head_upos]
    if model.head_flat is not None:
        heads.append(model.head_flat)
    for m in heads:
        for p in m.parameters():
            add(p, lr_head, is_enc=False)
    if hasattr(model, "mix_w"):
        add(model.mix_w, lr_head, is_enc=False)
    cap_emb = getattr(enc, "cap_emb", None)
    if cap_emb is not None:                    # fine-tune-only channel: learns at head LR
        for p in cap_emb.parameters():
            add(p, lr_head)
    return [dict(params=ps, lr=lr, weight_decay=w, base_lr=lr, is_enc=e)
            for (lr, w, e), ps in groups.items()]


@torch.no_grad()
def evaluate_dev(model, dev_rows, micro, device, T, W, tokenizer=None):
    """Unconstrained dev accuracies (early-stop signal; full constrained decode is
    evaluate.py's job). Returns counts tensor for cross-rank reduction."""
    model.eval()
    # counts: [xpos_exact_ok, script_ok, upos_ok, n_words, pos1_ok]
    cnt = torch.zeros(5, dtype=torch.long, device=device)
    for i in range(0, len(dev_rows), micro):
        batch = batch_chunk(dev_rows[i:i + micro], T, W, tokenizer)
        batch = {k: (v.to(device) if torch.is_tensor(v) else v) for k, v in batch.items()}
        with torch.autocast("cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
            out = model(batch)
        m = batch["y_script"] != -100
        if not m.any():
            continue
        if "flat" in out:
            cnt[0] += int(((out["flat"].argmax(-1) == batch["y_tag"]) & m).sum())
            cnt[4] += int(((out["xpos"][0].argmax(-1) == batch["y_xpos"][:, :, 0]) & m).sum())
        else:
            ok = torch.ones_like(m)
            for p, lg in enumerate(out["xpos"]):
                pred = lg.argmax(-1)
                good = pred == batch["y_xpos"][:, :, p]
                ok &= good | (batch["y_xpos"][:, :, p] == -100)
                if p == 0:
                    cnt[4] += int((good & m).sum())
            cnt[0] += int((ok & m).sum())
        cnt[1] += int(((out["script"].argmax(-1) == batch["y_script"]) & m).sum())
        cnt[2] += int(((out["upos"].argmax(-1) == batch["y_upos"]) & m).sum())
        cnt[3] += int(m.sum())
    model.train()
    return cnt


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--config", required=True)
    ap.add_argument("--ckpt", default=None, help="override backbone checkpoint path")
    ap.add_argument("--fold", type=int, default=None)
    a = ap.parse_args()
    cfg = json.loads(Path(a.config).read_text())
    if a.ckpt:
        cfg["ckpt"] = a.ckpt
    if a.fold is not None:
        cfg["fold"] = a.fold
    for k in ("out_dir", "kfold_dir"):
        cfg[k] = os.path.expandvars(cfg[k])
    if "ckpt" in cfg:
        cfg["ckpt"] = os.path.expandvars(cfg["ckpt"])

    rank, world, is_ddp = ddp_setup()
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    torch.manual_seed(cfg.get("seed", 0) + rank)
    torch.set_float32_matmul_precision("high")

    fold = cfg["fold"]
    T, W, micro = cfg.get("T", 2048), cfg.get("W", 384), cfg["micro_batch"]
    out = Path(cfg["out_dir"]); out.mkdir(parents=True, exist_ok=True)

    # ---- data + vocab (every rank builds identically; deterministic)
    t0 = time.time()
    kdir = Path(cfg["kfold_dir"])
    train_sents = list(read_conllu(kdir / f"train{fold}.conllu"))
    dev_sents = list(read_conllu(kdir / f"dev{fold}.conllu"))
    if cfg.get("max_sents"):
        train_sents = train_sents[:cfg["max_sents"]]
        dev_sents = dev_sents[:max(cfg["max_sents"] // 4, 8)]
    vocab = LabelVocab.build(train_sents, lambda f: encode_word(f) is not None)
    if rank == 0:
        vocab.save(out / "vocab.json")
        print(f"fold={fold} train_sents={len(train_sents)} dev_sents={len(dev_sents)} "
              f"scripts={vocab.n_scripts} tags={len(vocab.tags)} "
              f"vocab_build={time.time()-t0:.0f}s", flush=True)

    # ---- model (loaded before the datasets since the HF path needs its tokenizer to encode)
    encoder, pcfg, tokenizer = load_backbone_auto(cfg, device)
    is_hf = tokenizer is not None
    hf_max_len = cfg.get("hf_max_len", 512)
    use_cap = cfg.get("use_cap", False)
    if is_hf and use_cap:
        # the fine-tune-only capitalization channel is injected into CharBERT's char-plane
        # embeddings (see CharBertWithHidden.forward); there's no equivalent injection point
        # for subword input (casing is already whatever the subword vocab encodes), so it's
        # forced off here rather than silently leaving a dead, ungraded nn.Embedding around
        print("  [hf backbone] ignoring use_cap=true (no char-plane to inject into)",
              flush=True)
        use_cap = False

    train_ds = TaggerDataset(train_sents, vocab, T, W, tokenizer=tokenizer, hf_max_len=hf_max_len)
    dev_ds = TaggerDataset(dev_sents, vocab, T, W, tokenizer=tokenizer, hf_max_len=hf_max_len)
    dev_rows, _ = pack_dev_items(dev_ds.encs, W, tokenizer, T)
    dev_shard = dev_rows[rank::world]
    if rank == 0:
        print(f"encoded in {time.time()-t0:.0f}s  dev_rows={len(dev_rows)}", flush=True)

    tcfg = TaggerConfig(pool=cfg.get("pool", "mean"),
                        head_dropout=cfg.get("head_dropout", 0.1),
                        w_xpos=cfg.get("w_xpos", 1.0), w_script=cfg.get("w_script", 1.0),
                        w_upos=cfg.get("w_upos", 0.2), use_cap=use_cap,
                        w_flat=cfg.get("w_flat", 0.0),
                        scalar_mix=cfg.get("scalar_mix", False))
    model = TaggerModel(encoder, vocab, tcfg, W=W).to(device)
    if cfg.get("freeze_encoder", False):
        for p in model.encoder.parameters():
            p.requires_grad_(False)
    if is_ddp:
        from torch.nn.parallel import DistributedDataParallel as DDP
        model = DDP(model, device_ids=[rank % torch.cuda.device_count()])
    core = model.module if is_ddp else model

    opt = torch.optim.AdamW(
        [g for g in param_groups(core, cfg) if any(p.requires_grad for p in g["params"])],
        betas=(0.9, 0.95), fused=(device.type == "cuda"))

    # steps/epoch from a reference packing; schedule = warmup + cosine over the plan
    ref_rows, trunc = pack_dev_items(train_ds.encs, W, tokenizer, T)
    n_rows = len(ref_rows) // world * world
    steps_per_epoch = max(n_rows // world // micro, 1)
    total_steps = steps_per_epoch * cfg["epochs"]
    warmup = max(int(total_steps * cfg.get("warmup_frac", 0.05)), 1)
    if rank == 0:
        print(f"rows/epoch={n_rows} steps/epoch={steps_per_epoch} total={total_steps} "
              f"truncated_sents={trunc}", flush=True)

    def lr_scale(step):
        if step < warmup:
            return step / warmup
        f = (step - warmup) / max(total_steps - warmup, 1)
        return 0.5 * (1 + math.cos(math.pi * min(f, 1.0)))

    metrics_f = out / f"metrics_rank{rank}.jsonl"
    t0 = time.time()

    def do_step(batch, lr_s, step, tag, enc_scale=1.0):
        for pg in opt.param_groups:
            pg["lr"] = pg["base_lr"] * lr_s * (enc_scale if pg.get("is_enc") else 1.0)
        batch = {k: (v.to(device) if torch.is_tensor(v) else v) for k, v in batch.items()}
        with torch.autocast("cuda", dtype=torch.bfloat16, enabled=device.type == "cuda"):
            out_h = model(batch)
            loss, logs = core.loss(out_h, batch)
        opt.zero_grad(set_to_none=True)
        loss.backward()
        gnorm = torch.nn.utils.clip_grad_norm_(model.parameters(), cfg.get("clip", 1.0))
        if torch.isfinite(gnorm):
            opt.step()
        elif rank == 0:
            print(f"  [skip] non-finite grad at {tag} step {step}", flush=True)
        if rank == 0 and step % cfg.get("log_every", 20) == 0:
            rec = dict(phase=tag, step=step, lr=round(opt.param_groups[0]["lr"], 7),
                       loss=round(loss.item(), 4), gnorm=round(float(gnorm), 3),
                       sps=round(step / max(time.time() - t0, 1e-9), 2), **logs)
            print("  " + " ".join(f"{k}={v}" for k, v in rec.items()), flush=True)
            with open(metrics_f, "a") as mf:
                mf.write(json.dumps(rec) + "\n")

    # ---- phase 1 (optional): silver lemma-distillation pretrain (script loss only)
    if cfg.get("silver"):
        t1 = time.time()
        silver_sents = read_silver(os.path.expandvars(cfg["silver"]),
                                   cfg.get("silver_limit", 4_000_000))
        silver_ds = TaggerDataset(silver_sents, vocab, T, W, tokenizer=tokenizer,
                                  hf_max_len=hf_max_len)
        mask_non_lemma_labels(silver_ds)
        if rank == 0:
            print(f"silver: {len(silver_sents)} sents, encoded in {time.time()-t1:.0f}s",
                  flush=True)
        sstep = 0
        for sep in range(cfg.get("silver_epochs", 1)):
            rows, _ = pack_dev_items(silver_ds.encs, W, tokenizer, T,
                                     order=torch.randperm(
                                         len(silver_ds.encs),
                                         generator=torch.Generator().manual_seed(123 + sep)
                                         ).tolist())
            n = len(rows) // world * world
            shard = rows[rank:n:world]
            nsteps = len(shard) // micro * micro
            warm = max((nsteps // micro) // 10, 1)
            for i in range(0, nsteps, micro):
                # warmup then constant LR (distillation phase, no decay); optionally keep
                # the encoder frozen so silver noise cannot drift the tagging features
                do_step(batch_chunk(shard[i:i + micro], T, W, tokenizer),
                        min((sstep + 1) / warm, 1.0), sstep, "silver",
                        enc_scale=cfg.get("silver_enc_scale", 0.0))
                sstep += 1
        cnt = evaluate_dev(model, dev_shard, cfg.get("eval_micro", micro), device, T, W,
                           tokenizer)
        if is_ddp:
            import torch.distributed as dist
            dist.all_reduce(cnt)
        if rank == 0:
            n_w = max(int(cnt[3]), 1)
            print(f"  SILVER DONE steps={sstep} dev_script_acc={int(cnt[1])/n_w:.4f}",
                  flush=True)

    best_score, best_epoch = -1.0, -1
    step = 0
    for epoch in range(cfg["epochs"]):
        rows, _ = pack_dev_items(train_ds.encs, W, tokenizer, T,
                                 order=torch.randperm(
                                     len(train_ds.encs),
                                     generator=torch.Generator().manual_seed(
                                         cfg.get("seed", 0) * 1000 + epoch)).tolist())
        n = len(rows) // world * world
        shard = rows[rank:n:world]
        nsteps = len(shard) // micro * micro
        for i in range(0, nsteps, micro):
            do_step(batch_chunk(shard[i:i + micro], T, W, tokenizer), lr_scale(step), step,
                    "gold")
            step += 1

        # ---- dev eval (each rank its shard, reduce counts)
        cnt = evaluate_dev(model, dev_shard, cfg.get("eval_micro", micro), device, T, W,
                           tokenizer)
        if is_ddp:
            import torch.distributed as dist
            dist.all_reduce(cnt)
        n_w = max(int(cnt[3]), 1)
        m = dict(epoch=epoch, dev_xpos_exact=round(int(cnt[0]) / n_w, 4),
                 dev_script_acc=round(int(cnt[1]) / n_w, 4),
                 dev_upos_acc=round(int(cnt[2]) / n_w, 4),
                 dev_pos1_acc=round(int(cnt[4]) / n_w, 4), n_words=n_w)
        score = (m["dev_xpos_exact"] + m["dev_script_acc"]) / 2
        if rank == 0:
            print("  EVAL " + json.dumps(m), flush=True)
            with open(out / "eval.jsonl", "a") as ef:
                ef.write(json.dumps(m) + "\n")
            if score > best_score:
                tmp = out / "best.pt.tmp"
                torch.save(dict(model=core.state_dict(), tcfg=vars(tcfg), cfg=cfg,
                                pretrain_cfg=pcfg, epoch=epoch, dev=m, W=W, T=T), tmp)
                os.replace(tmp, out / "best.pt")
                print(f"  new best.pt (score {score:.4f})", flush=True)
        # keep best/stop decisions consistent across ranks (same reduced counts)
        if score > best_score:
            best_score, best_epoch = score, epoch
        if epoch - best_epoch >= cfg.get("patience", 6):
            if rank == 0:
                print(f"EARLY STOP at epoch {epoch} (best epoch {best_epoch}, "
                      f"score {best_score:.4f})", flush=True)
            break

    if rank == 0:
        print(f"DONE best_score={best_score:.4f} best_epoch={best_epoch} "
              f"({(time.time()-t0)/60:.1f} min)", flush=True)
    if is_ddp:
        import torch.distributed as dist
        dist.destroy_process_group()


if __name__ == "__main__":
    main()