Multi-layer Perceptron (MLP) Neural Network

This project involves training and testing a Multi-layer Perceptron (MLP) Neural Network using two datasets: the iris dataset and an autoencoder dataset. The implemented MLP uses backpropagation for training, and supports various command-line parameters to customize the training and testing process.

Usage

You can run the main script main.py from the command line and customize the behavior with several options.

Example run:

python main.py -i 4 -hn 2 -o 4 -b 1 -p irises --lr 0.6 --epochs 800 --momentum 0.5 --train-file irises_train_log.txt --test-file irises_test_log.txt --test-size 0.9 --save-network --network-file irises.pickle

Parameters

• -i, --input-nodes: Number of input nodes in the MLP (Required).
• -hn, --hidden-layer-neurons: Number of neurons in the hidden layer (Required).
• -o, --output-nodes: Number of output nodes in the MLP (Required).
• -b, --bias: Whether to use bias nodes in the MLP (Required).
• -p, --process: Which dataset to process. Choices are "autoencoder" or "irises" (Required).

Training parameters:

• --lr, --learning-rate: Learning rate for MLP training (Default: 0.5).
• --epochs, --max-epochs: Maximum number of epochs for MLP training (Default: 1000).
• --err, --desired-error: Desired error for MLP training (Default: 0.05).
• --momentum: Momentum factor for MLP training (Default: 0.5).
• --record-interval: Interval (in epochs) at which to record error during MLP training (Default: 10).
• --train-file: File to save the error records during MLP training (Default: 'train_log.txt').

Test parameters:

• --record-items: A comma-separated list of items to record during training. If None save all statistics. Choose from: input_pattern, total_error, desired_output, output_errors, output_values, output_weights, hidden_values, hidden_weights (Default: None).
• --test-file: File to save the statistics during MLP testing (Default: 'test_log.txt').

Irises dataset specific options:

• --test-size: Proportion of the iris data to use as a test set. Ignored if --process is "autoencoder" (Default: 0.6).
• --result-file: File to save the results of testing the MLP on the iris data. Ignored if --process is "autoencoder" (Default: 'irises_results.txt').

Save/load network options:

• --save-network: Whether to save the trained MLP to a file (Default: False).
• --load-network: If specified, load the MLP from this file instead of training a new one (Default: None).
• --network-file: File to save or load the trained MLP (Default: 'mlp.pickle').
• --model-file: The file to save and load the trained model from for training sessions (Default: False).

License

MIT