A ELM sentiment classifier

This is a sentiment classifier implemented via Extreme Learning machine and Pretrained Language Model.

1. Requirements

For ELM

• numpy==1.19.4
• matplotlib==3.3.3(optional)
• tqdm==4.54.0(optional)
• sklearn==0.23.2(optional)

For sentiment classification

• pytorch==1.7.0
• transformers==4.0.0
• matplotlib==3.3.3
• pandas==1.1.4
• tqdm==4.54.0

2. Getting Started

2.1 Using ELM

There are 3 types of elm in elm.py:

• basic_elm: basic implementation of elm(single layer, binary classification only)
• normal_elm: normalize with a parameter before calculating Moore–Penrose inverse(single layer, binary classification only)
• classic_elm(recommend): single layer elm for multi-classes

The detailed usage can be found in elm_example.py

You can also use the wrapper ELM from elm.py.

from elm import ELM
from argparse import ArgumentParser

def main():
    # parse args
    parser = ArgumentParser()
    parser.add_argument('--type', type=str, default='classic')
    parser.add_argument('--input_shape', type=int, default=256)
    parser.add_argument('--hidden_dim', type=int, default=64)
    parser.add_argument('--activation', type=str, default='sigmoid')
    parser.add_argument('--normalize', action=float, default=1.0)
    parser.add_argument('--classes', type=int, default=2)
    args = parser.parse_args()
    
    # load ELM
    elm = ELM(args)

2.2 Using sentiment classification

To run sentiment classification task simply run sentiment.py in your command line.

python sentiment.py --training_type finetune_classifier_elm \
 --batch_size 64 \
 --epoch_num 6 \
 --learning_rate 1e-5 \
 --eval_epoch 1

For detailed usage, run:

python sentiment.py --help

3. TODO

• add rbf kernel for elm
• add multi-layer elm
• add chinese dataset support for sentiment classify