This project uses a combination of TensorFlow, scikit-learn,and Pandas to construct a multi-class classifier.
A feed-forward neural network is a type of neural network architecture consisting of a series of sequential layers that do not loop. The number of hidden layers determines whether it can be classified as a single-layer perceptron or a multi-layer perceptron. In this project, such an architecture is used to establish multi-class classification where data from a provided dataset is classified into one of multiple possible classes.
The provided dataset contains cartographic variables that pertain to different forest cover types. In order to be used for training and prediction, it must be run through a series of pre-processing tasks to make it viable for use by the neural network.
Initially, it is split into data and label datasets, containing raw data, and prediction attributes respectively. Using scikit-learn, each is then split further into training and validation datasets, to train the neural network. After this is completed, all datasets are then normalised before being converted into binary vectors.
The previous datasets are then fed into the model. Using a grid-searching algorithm, the best hyperparameters are found, before a newly-generated model is trained with them. Upon completion of training, the latter is then saved into an HDF5 file.
To make predictions, the previously trained neural network is loaded. The preprocessed data and labels are again loaded for prediction purposes. Finally, a classification report of the results is displayed.