Databalancer

Databalancer is the python library using in machine learning applications to balance the imbalanced text classification datasets before the model training.

Features

• Databalancer is able to balance any imbalanced text classification datasets
• If the given dataset is imbalanced then while balancing no existing data is removed, but new data will be generated and added to the dataset
• For a particular class the newly generated data will be the paraphrases of the existing data in that particular class
• By default, these paraphrases are generated using the ramsrigouthamg/t5_paraphraser model (You can read more about the model from Huggingface official documentation)
• The current version can generate the sentence paraphrases using multiple methods such as T5 models, NLPAUG and Textattack
• The user can select the balance method by passing the balance_method parameter while calling the balanceDataset method such as
  • balance_method=1 for ramsrigouthamg/t5_paraphraser T5 model based balancing (Default) ( For more info check t5_paraphraser)
  • balance_method=2 for ramsrigouthamg/t5-large-paraphraser-diverse-high-quality T5 model based balancing (For more info check t5-large-paraphraser-diverse-high-quality)
  • balance_method=3 for nlpaug based balancing (For more info check nlpaug)
  • balance_method=4 for textattack based balancing (For more info check textattack)
• The model argument in the balanceDataset method is only applicable when balance_method is set as 3, through which user can pass the transformer model name from Huggingface to generate paraphrases using NLPAUG .
• If the user enable quantize=True in balanceDataset then the T5 models(balance_method==1 and balance_method=2) will go through the quantization process using fastT5 before inference, so that the model inference time will be reduced.
• By default quantize parameter is set as False because quantization requires more RAM and more CPU Processing power
• Databalancer also provides another method called classCountVisualization to show the dataset class count distribution

Installation

Install the databalancer package with pip

 pip install databalancer

Compatibility

Databalancer is only compatable with python 3.6.9 or above.

Quick Start

The library databalancer provides two different functionalities.

1 - classCountVisualization

2 - balanceDataset

classCountVisualization

#Import the classCountVisualization from the 'databalancer' module
from databalancer import classCountVisualization
    
#Pass the required datasetname(here traindata.csv) to the function
classCountVisualization("traindata.csv")

Output

balanceDataset

#Import the balanceDataset from the 'databalancer' module
from databalancer import balanceDataset

#Pass the dataset name which is to be balanced(here traindata.csv) to the balanceDataset function
balanceDataset("traindata.csv",balance_method=1)

The above code will balance the dataset and store the balanced dataset('balanced_data.csv') in the local machine.

balanceDataset with model quantization

#Import the balanceDataset from the 'databalancer' module
from databalancer import balanceDataset

#Pass the dataset name which is to be balanced(here traindata.csv) to the balanceDataset function with balance_method=2 and enable quantization 
balanceDataset("traindata.csv",balance_method=2,quantize=True)

The above code will balance the dataset using balance_method=2 with quantization and store the balanced dataset('balanced_data.csv') in the local machine.

To show the balanced dataset class count distribution, run the code below.

from databalancer import classCountVisualization

classCountVisualization("balanced_data.csv")