GAU-alpha-pytorch

pytorch版本https://github.com/ZhuiyiTechnology/GAU-alpha

News

• 2022/05/11 感谢苏神提醒，添加了一个注释，其中RoFormerV2^*表示未经多任务学习的RoFormerV2模型。
• 2022/04/22 初始化仓库，添加初步的代码, 添加paddle版本gau_alpha代码。
• 2022/04/30 添加CLUE分类任务代码。

Install

pip install git+https://github.com/JunnYu/GAU-alpha-pytorch.git
or
pip install gau_alpha

精度对齐

python compare.py
# bert4keras vs pytorch
# mean diff : tensor(6.9320e-07)
# max diff : tensor(3.9101e-05)

torch版本使用

依赖：

• torch
• transformers

import torch
from gau_alpha import GAUAlphaForMaskedLM, GAUAlphaTokenizer

text = "今天[MASK]很好，我[MASK]去公园玩。"
tokenizer = GAUAlphaTokenizer.from_pretrained(
    "junnyu/chinese_GAU-alpha-char_L-24_H-768"
)
pt_model = GAUAlphaForMaskedLM.from_pretrained(
    "junnyu/chinese_GAU-alpha-char_L-24_H-768"
)
pt_model.eval()
pt_inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    pt_outputs = pt_model(**pt_inputs).logits[0]
pt_outputs_sentence = "pytorch: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        val, idx = pt_outputs[i].softmax(-1).topk(k=5)
        tokens = tokenizer.convert_ids_to_tokens(idx)
        new_tokens = []
        for v, t in zip(val.cpu(), tokens):
            new_tokens.append(f"{t}+{round(v.item(),4)}")
        pt_outputs_sentence += "[" + "||".join(new_tokens) + "]"
    else:
        pt_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(pt_outputs_sentence)
# pytorch: 今天[天+0.8657||气+0.0535||阳+0.0165||，+0.0126||晴+0.0111]很好，我[要+0.4619||想+0.4352||又+0.0252||就+0.0157||跑+0.0064]去公园玩。

Paddle版本使用

依赖：

• paddlepaddle>=2.2.0
• paddlenlp

import paddle
from transformers import BertTokenizer as GAUAlphaTokenizer
from gau_alpha_paddle import GAUAlphaForMaskedLM

text = "今天[MASK]很好，我[MASK]去公园玩。"
tokenizer = GAUAlphaTokenizer.from_pretrained(
    "junnyu/chinese_GAU-alpha-char_L-24_H-768"
)
pd_model = GAUAlphaForMaskedLM.from_pretrained("chinese_GAU-alpha-char_L-24_H-768")
pd_model.eval()
pd_inputs = tokenizer(text)
pd_inputs = {k: paddle.to_tensor([v]) for k, v in pd_inputs.items()}

with paddle.no_grad():
    pd_outputs = pd_model(**pd_inputs)[0][0]

pd_outputs_sentence = "paddle: "
for i, id in enumerate(tokenizer.encode(text)):
    if id == tokenizer.mask_token_id:
        val, idx = paddle.nn.functional.softmax(pd_outputs[i], -1).topk(k=5)
        tokens = tokenizer.convert_ids_to_tokens(idx)
        new_tokens = []
        for v, t in zip(val.cpu(), tokens):
            new_tokens.append(f"{t}+{round(v.item(),4)}")
        pd_outputs_sentence += "[" + "||".join(new_tokens) + "]"
    else:
        pd_outputs_sentence += "".join(
            tokenizer.convert_ids_to_tokens([id], skip_special_tokens=True)
        )
print(pd_outputs_sentence)
# paddle: 今天[天+0.8657||气+0.0535||阳+0.0165||，+0.0126||晴+0.0111]很好，我[要+0.4619||想+0.4352||又+0.0252||就+0.0157||跑+0.0064]去公园玩。

介绍

• GAU-α：https://kexue.fm/archives/9052
• GAU：https://kexue.fm/archives/8934
• 原始论文：https://arxiv.org/abs/2202.10447

评测对比

CLUE-dev榜单分类任务结果，base版本。

	iflytek	tnews	afqmc	cmnli	ocnli	wsc	csl
BERT	60.06	56.80	72.41	79.56	73.93	78.62	83.93
RoBERTa	60.64	58.06	74.05	81.24	76.00	87.50	84.50
RoFormer	60.91	57.54	73.52	80.92	76.07	86.84	84.63
RoFormerV2*	60.87	56.54	72.75	80.34	75.36	80.92	84.67
GAU-α	61.41	57.76	74.17	81.82	75.86	79.93	85.67
RoFormerV2-pytorch	62.87	59.03	76.20	80.85	79.73	87.82	91.87
GAU-α-pytorch（Adafactor）	61.18	57.52	73.42	80.91	75.69	80.59	85.5
GAU-α-pytorch（AdamW wd0.01 warmup0.1）	60.68	57.95	73.08	81.02	75.36	81.25	83.93

CLUE-test榜单分类任务结果，base版本。

	iflytek	tnews	afqmc	cmnli	ocnli	wsc	csl
RoFormerV2-pytorch	63.15	58.24	75.42	80.59	74.17	83.79	83.73
GAU-α-pytorch（Adafactor）	61.38	57.08	74.05	80.37	73.53	74.83	85.6
GAU-α-pytorch（AdamW wd0.01 warmup0.1）	60.54	57.67	72.44	80.32	72.97	76.55	84.13

CLUE-dev集榜单阅读理解和NER结果

	cmrc2018	c3	chid	cluener
BERT	56.17	60.54	85.69	79.45
RoBERTa	56.54	67.66	86.71	79.47
RoFormer	56.26	67.24	86.57	79.72
RoFormerV2*	57.91	64.62	85.09	81.08
GAU-α	58.09	68.24	87.91	80.01

注：

• 其中RoFormerV2^*表示的是未进行多任务学习的RoFormerV2模型，该模型苏神并未开源，感谢苏神的提醒。
• 其中不带有pytorch后缀结果都是从GAU-alpha仓库复制过来的。
• 其中带有pytorch后缀的结果都是自己训练得出的。

引用

Bibtex：

@techreport{gau-alpha,
  title={GAU-α: GAU-based Transformers for NLP - ZhuiyiAI},
  author={Jianlin Su, Shengfeng Pan, Bo Wen, Yunfeng Liu},
  year={2022},
  url="https://github.com/ZhuiyiTechnology/GAU-alpha",
}

Tips:

• 感谢苏神提供的模型和代码！