SIGRNN

SIGRNN: Synthetic Minority Instances Generation in Imbalanced Datasets Using a Recurrent Neural Network Approach

Transform data

• Before generating any data first we transform the input data to a paragraph. transform_data.py is an example of a code snippiet to convert a csv file of features to a text file. This text file if fed as input to the SIGRNN in the next step.

Train SIGRNN

• To train SIGRNN run main.py. The script takes a set of inputs and can be executed as the following example: python main.py --cuda --nhid 1024 --dropout 0.5 --epochs 20 --batch_size 8 --nlayers 2 --data ./data/seer --bptt 11
  • Flag descriptions:
    • --code is used for GPU execution. The GPU id can be set within the code.
    • --nhid is the number of hidden units.
    • --droput is the percentage of drouped out connections.
    • --batch_size is to specify the batch size.
    • --data is the folder where the data is saved.
      • 3 files need to be present under the directory train.txt, test.txt, and valid.txt
    • --nlayers is the number of hidden layers. The default cell is LSTM.
    • --bptt is for the length of the sequence.
    • The model will be saved in model.pt

Generate using SIGRNN

• To generate synthetic SIGRNN run generate.py. The script takes a set of inputs and can be executed as the following example: python generate.py --cuda --words 1500000 --data ./data/seer --outf ./data/seer/generated.txt --temperature 1.3
  • Flag descriptions:
    • --cuda is used for GPU execution. The GPU id can be set within the code.
    • --words is the number of words to be generated.
    • --outf is to specify the output file.
    • --temperature is to increase/decrese the varience in the generated examples.
    • --data is the folder where the data is saved.
      • 3 files need to be present under the directory train.txt, test.txt, and valid.txt