ducanhdinh / jepa_proof_boyl_replace

ducanhdinh/jepa_proof_boyl_replace

BERT encoder pretrained from scratch với BYOL + Lexical Substitution augmentation .

Augmentation strategy

Thay vì span masking, model này dùng lexical substitution để tạo view 2:

Quá trình tạo view 2 được thực hiện offline 1 lần trước khi train.

Kiến trúc BYOL

View 1 ──► Online Encoder (θ) ──► Online Projector (θ) ──► Online Predictor (θ) ──► p1 ──┐
                                                                                           ├── loss = cosine(p1, z2) + cosine(p2, z1)
View 2 ──► Online Encoder (θ) ──► Online Projector (θ) ──► Online Predictor (θ) ──► p2 ──┘
View 1 ──► Target Encoder (ξ) ──► Target Projector (ξ) ──► z1  (stop grad)
View 2 ──► Target Encoder (ξ) ──► Target Projector (ξ) ──► z2  (stop grad)

Target update: ξ ← 0.996·ξ + 0.0040000000000000036·θ  (EMA)

Thông số huấn luyện

Cách dùng — Online BERT encoder (feature extraction)

from transformers import BertModel, BertTokenizerFast
import torch

tokenizer = BertTokenizerFast.from_pretrained("ducanhdinh/jepa_proof_boyl_replace")
bert      = BertModel.from_pretrained("ducanhdinh/jepa_proof_boyl_replace/encoder")

encoded = tokenizer(
    ["Hello world!", "BYOL with lexical substitution."],
    return_tensors="pt",
    padding=True,
    truncation=True,
)
with torch.no_grad():
    out = bert(**encoded)
    cls_emb = out.last_hidden_state[:, 0, :]   # [CLS] token → (B, 768)

Cách dùng — Full model

import torch
from text_byol_replace import TextBYOLReplace, BYOLReplacePretrainConfig

cfg   = BYOLReplacePretrainConfig()
model = TextBYOLReplace(cfg)
state = torch.load(
    hf_hub_download("ducanhdinh/jepa_proof_boyl_replace", "pytorch_model.bin"),
    map_location="cpu",
)
model.load_state_dict(state)
model.eval()