jev-training-data/train_standalone.py

152 lines
6.8 KiB
Python

#!/usr/bin/env python3
"""train_all.py — 一体化: 下载 mmBERT + 训练 + 评估 (noul 二分类)"""
import json, os, sys, urllib.request
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from transformers import AutoTokenizer, AutoConfig, AutoModel
from peft import LoraConfig, get_peft_model
import bisect
INSTR = "Does this document contain the information needed to answer the user's question? Answer based only on the document content."
def load_jsonl(p):
return [json.loads(l) for l in open(p, encoding="utf-8")]
def build_seq(tok, state, q, max_len, head_len):
mask_tok = tok.mask_token
opts = ["false", "true"]
ins = str(q["ins"]).replace(mask_tok, " ")
head_ids = tok("%s question: %s" % (q["t"], ins), add_special_tokens=False)["input_ids"]
opt_ids = []
for o in opts:
opt_ids.append([tok.mask_token_id] + tok(" "+o.replace(mask_tok," "), add_special_tokens=False)["input_ids"][:48])
budget = head_len - sum(len(o) for o in opt_ids)
if budget < 16:
per = max(4, (head_len-16)//2)
opt_ids = [o[:per] for o in opt_ids]
budget = head_len - sum(len(o) for o in opt_ids)
head_ids = head_ids[:max(8, budget)]
ids = [tok.cls_token_id] + head_ids + [tok.sep_token_id]
markers = [len(ids)]
ids.extend(opt_ids[1])
ids.append(tok.sep_token_id)
room = max(0, max_len - len(ids) - 1)
st = tok(state.replace(mask_tok, " "), add_special_tokens=False)["input_ids"]
st = st[:room]
ids = ids + st + [tok.sep_token_id]
return ids[:max_len], [m for m in markers if m < max_len]
class NoulDS(Dataset):
def __init__(self, rows, tok, max_len=512):
self.rows, self.tok, self.max_len = rows, tok, max_len
self.q = {"t": "noul", "ins": INSTR, "crit": None}
def __len__(self): return len(self.rows)
def __getitem__(self, i):
r = self.rows[i]
state = f"用户问题: {r['query']}\n\n文档内容:\n{r['doc_text'][:2500]}"
ids, markers = build_seq(self.tok, state, self.q, self.max_len, 192)
pad = self.max_len - len(ids)
mm = [True]*len(markers) + [False]*(8-len(markers))
return {"input_ids": torch.tensor(ids + [0]*pad),
"attention_mask": torch.tensor([1]*len(ids) + [0]*pad),
"marker_pos": torch.tensor(markers + [0]*(8-len(markers))),
"marker_mask": torch.tensor(mm),
"label": torch.tensor(float(r["hard_label"])),
"soft": torch.tensor(float(r.get("soft_label") or r["hard_label"]))}
def collate_noul(batch):
keys = batch[0].keys()
return {k: torch.stack([b[k] for b in batch]) for k in keys}
def ece(probs, labels):
bins = [[] for _ in range(15)]
for p, y in zip(probs, labels):
bins[min(int(p*15), 14)].append((p, y))
e = 0.0
for b in bins:
if not b: continue
e += len(b)/len(probs) * abs(sum(p for p,_ in b)/len(b) - sum(y for _,y in b)/len(b))
return e
def main():
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("--data", default="/tmp/jd")
ap.add_argument("--base", default="/tmp/mmbert")
ap.add_argument("--out", default="/tmp/laya_out")
ap.add_argument("--epochs", type=int, default=2)
ap.add_argument("--bs", type=int, default=16)
ap.add_argument("--lr", type=float, default=1e-4)
args = ap.parse_args()
os.makedirs(args.out, exist_ok=True)
dev = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[jev] device={dev}", flush=True)
tok = AutoTokenizer.from_pretrained(args.base)
enc_cfg = AutoConfig.from_pretrained(args.base)
enc = AutoModel.from_config(enc_cfg)
from laya.common import DecisionModel
dm = DecisionModel(enc, head_layers=2, n_act=2)
lora = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, target_modules=["Wqkv", "Wo"], task_type=None)
dm.encoder = get_peft_model(dm.encoder, lora)
dm.to(dev); dm.train()
print(f"[jev] model ready on {dev}", flush=True)
train = [r for r in load_jsonl(os.path.join(args.data, "train.jsonl")) if r.get("hard_label") in (0, 1)]
calib = load_jsonl(os.path.join(args.data, "calibration.jsonl"))
print(f"[jev] train={len(train)} calib={len(calib)}", flush=True)
dl = DataLoader(NoulDS(train, tok), batch_size=args.bs, shuffle=True, num_workers=2, drop_last=True, collate_fn=collate_noul)
dl_cal = DataLoader(NoulDS(calib, tok), batch_size=32, collate_fn=collate_noul)
steps = len(dl) * args.epochs
opt = torch.optim.AdamW([p for p in dm.parameters() if p.requires_grad], lr=args.lr, weight_decay=0.01)
sch = get_cosine_schedule_with_warmup(opt, int(0.06*steps), steps)
best_auc = 0.0
gstep = 0
for ep in range(args.epochs):
dm.train()
for b in dl:
with torch.autocast("cuda", enabled=(dev=="cuda")):
logits, act = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
qtype=b["qtype"].to(dev))
true_logit = logits[..., 1].float()
hard = b["label"].to(dev); soft = b["soft"].to(dev)
loss = F.binary_cross_entropy_with_logits(true_logit, hard) + 0.5*F.mse_loss(torch.sigmoid(true_logit), soft)
opt.zero_grad(); loss.backward()
torch.nn.utils.clip_grad_norm_([p for p in dm.parameters() if p.requires_grad], 1.0)
opt.step(); sch.step()
gstep += 1
if gstep % 100 == 0:
print(f"[jev] ep{ep} step{gstep}/{steps} loss={loss.item():.4f}", flush=True)
dm.eval()
ps, ls = [], []
with torch.no_grad():
for b in dl_cal:
lg, _ = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
qtype=b["qtype"].to(dev))
p = torch.softmax(lg.float(), -1)[..., 1]
ps += p.tolist(); ls += b["label"].tolist()
pos = sorted(p for p, y in zip(ps, ls) if y == 1)
neg = sorted(p for p, y in zip(ps, ls) if y == 0)
auc = sum(bisect.bisect_left(neg, p) for p in pos) / (len(pos)*len(neg)) if pos and neg else 0
e = ece(ps, ls)
print(f"[jev] ep{ep} calib AUC={auc:.4f} ECE={e:.4f}", flush=True)
if auc > best_auc:
best_auc = auc
torch.save(dm.state_dict(), os.path.join(args.out, "model.pt"))
print(f"[jev] saved best (auc={auc:.4f})", flush=True)
dm.train()
print(f"[jev] TRAINING-DONE best_auc={best_auc:.4f}", flush=True)
def collate_noul(batch):
keys = batch[0].keys()
return {k: torch.stack([b[k] for b in batch]) for k in keys}
if __name__ == "__main__":
main()