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:

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}
}