FastPLM

An Empirical Study on Adaptive Inference for Pre-trained Language Model.

Model Zoo

FastPLM is supported by the UER project, and all of UER high-quality models can be accelerated in the Adaptive inference way.

FastPLM object supports the following models:

Models (kernel_name)	URL	Description
google_bert_base_en	https://share.weiyun.com/fpdOtcmz	Google pretrained English BERT-base model on Wiki corpus.
google_bert_base_zh	https://share.weiyun.com/AykBph9V	Google pretrained Chinese BERT-base model on Wiki corpus.
uer_bert_large_zh	https://share.weiyun.com/chx2VhGk	UER pretrained Chinese BERT-large model on mixed corpus.
uer_bert_small_zh	https://share.weiyun.com/wZuVBM5g	UER pretrained Chinese BERT-small model on mixed corpus.
uer_bert_tiny_zh	https://share.weiyun.com/VJ3JEN9Z	UER pretrained Chinese BERT-tiny model on mixed corpus.
uer_roberta_base_zh	https://share.weiyun.com/2gdpc4P0	Facebook pretrained Chinese RoBerta-base model.
uer_roberta_base_en	https://share.weiyun.com/xHrPjgEK	Facebook pretrained English RoBerta-base model.
uer_gpt_zh	https://share.weiyun.com/Pzn5Iob2	UER pretrained Chinese GPT model on mixed corpus.
uer_gpt_en	https://share.weiyun.com/Kc7KlgBs	UER pretrained English GPT model on Wikien corpus.
uer_gcnn_9_zh	https://share.weiyun.com/AIL6xiPa	UER pretrained Chinese GCNN model on Wiki corpus.
uer_gcnn_9_en	https://share.weiyun.com/TzjugXLH	UER pretrained English GCNN model on Wikien corpus.
google_albert_base_en	https://share.weiyun.com/artkxjB1	Google pretrained English ALBERT-base model.
google_albert_base_zh	https://share.weiyun.com/ewaiqKdR	Google pretrained Chinese ALBERT-base model.
huawei_tinybert_4_en	https://share.weiyun.com/HaulZbd2	Huawei pretrained English TinyBERT_4 model.
huawei_tinybert_6_en	https://share.weiyun.com/jPg6FkwA	Huawei pretrained English TinyBERT_6 model.

In fact, you don't have to download the model yourself. FastPLM will download the corresponding model file automatically at the first time you use it. If the automatically downloading failed, you can download these model files from the above URLs, and saving them to the directory of "~/.fastplm/".

Quick Start

Single-sentence classification

An example of single-sentence classification is shown in single_sentence_classification.

from fastplm import FastPLM

# Loading your dataset
labels = ['T', 'F']
sents_train = [
    'Do you like FastPLM?',
    'Yes, it runs faster than original PLM!',
    ...
]
labels_train = [
    'T',
    'F',
    ...
]

# Creating a model
model = FastPLM(
    kernel_name="google_bert_base_en",  # "google_bert_base_zh" for Chinese
    labels=labels,
    device='cuda:0'
)

# Training the model
model.fit(
    sents_train,
    labels_train,
    model_saving_path='./fastplm.bin',
)

# Loading the model and making inference
model.load_model('./fastplm.bin')
label, exec_layers = model('I like FastPLM', speed=0.7)

Double-sentence classification

An example of double-sentence classification is shown in double_sentence_classification.

Acknowledgement

This work is supported by Tencent Rhino-Bird Program. A portion of this work was presented at the 58th Annual Meeting of the Association for Computational Linguistics (ACL) in July 2020, FastBERT.

@article{weijieliu2021fastplm,
  title={An Empirical Study on Adaptive Inference for Pre-trained Language Model},
  author={Weijie Liu, Xin Zhao, Zhe Zhao, Qi Ju*, Xuefeng Yang, and Wei Lu},
  journal={IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS},
  year={2021}
}

autoliuweijie / FastPLM