Detecting single digits in 32x32 images from Format 2 The Street View House Numbers (SVHN).
Using virtualenv named SigTest. The virtualenv can be activated by running terminal command "source sigTest/bin/activate" when in the repositories folder. The virtualenv can be deactivated by running termnial command "deactivate" within same folder as activated.
Run terminal command "python trainCNN.py" to train model. If chosen model to train doesn't already exists a model will first be compiled and save in the chosen name and then trained and saved again. - Make sure paths for compile and train match what you want -
Run terminal command "python testCNN.py" to test model. Make sure function "test" in CNN.py has a matching path to the model you want tested.
Using Keras with Tensorflow backend to create a Convolutional Neural Network (CNN) which learns to detect the MNIST-like 32x32 image SVHN data set.
Testing is only done through evaluation which prints the loss and accuracy of the model on test data it has not seen before. Predictions are commented out, along with the code for users to choose through input if they want to evaluate the model or use it for prediction.
Predictions would be nice with a ROC curve together with its AUC to show performance. But to do this a new model has to be compiled using Keras Sequential model with the Keras-Classifier wrapper for sklearnto be able to use ROC and AUC from sklearn.
No changes needed in the code to use "bestSparse" model which uses sparse categorical crossentropy.
To use the "bestCheck" model, which uses categorical crossentropy as loss and metric, the function "to_categorical" has to be used to change the labels to a binary class matrix. To_categorical is commented out in both train and test file right after 10s are changed to 0s in labels arrays.