A short and simple 3 layers Neural Network Tensorflow code to classification
Just specify your train and test/validation sets and let the algorithm classify your data!
python neural_network.py does it all.
python 3.6.8
numpy 1.16.3
pandas 0.24.2
scikit-learn 0.21.1
scipy 1.2.1
tensorflow 1.13.1 (to CPU process)
tensorflow-gpu 1.13.1 (to GPU process)
It's a 3 hidden layers MLP.
You just need to specify a .csv file with the data to training, test or validation in dataset folder.
You can set up the training/test file in the variable train_path and test_path neural_network.py
It's set to iterate up to 300 epochs with an early stopping logic to avoid divergence.
LAYER1 = Neurons number in the first hidden layer
LAYER2 = Neurons number in the second hidden layer
LAYER3 = Neurons number in the third hidden layer
LR = learning rate
BETA = beta parameter to regularization. 0 to ignore
The Hyperparameter optimization search is performed by a Random Search algorithm with 3 trails of hyperparameter
You can change the number of trails in the main function
The method ExtraTreesClassifier is automatically applied to reduce the data dimension.
In the MNIST example, the data dimension is reduced of 784 features to 274.
The best result founded in the Random Search is saved in the folder TF_model.
The test/validation classification is inside this folder (classification.csv)
The Tensorflow model (session) is saved inside the sess folder.
At the end of the execution, some metrics like error rate and accuracy will be presented.
A learning plot will also be displayed with train and test error rate.
