SVHN-Classifier

Pretrained classifier (Convolutional Neural Network, CNN) to classify SVHN images, based on Keras with the Tensorflow backend.

Requirements:

• Keras 2.1.4
• Numpy 1.14.1

To predict images:

To predict existing images with the pre-trained model (95.45% accuracy on the SVHN test set)

• python svhn_classifier.py --predict --model weights.hdf5 --img_path path-to-images

Labels for the pretrained model are according to the digit, i.e. digit "0" has label "0", digit "1" has label "1", etc. This is different from the original data, in which digit "0" has label "10". When using the pretrained model to predict data make sure images depicting a "0" are stored in the folder labelled with "0", not "10".

Images should be stored in the following layout:

• path-to-images
  • class-0
    • img1.jpg
    • img2.jpg
    • ...
  • class-1
    • img1.jpg
    • img2.jpg
    • ...
  • ...

To train a new classifier

Download the SVHN data set:

• go to http://ufldl.stanford.edu/housenumbers/
• download the cropped digits (Format 2): train_32x32.mat and test_32x32.mat
• preprocess images: python preprocess_svhn.py --data path-to-the-downloaded-files --save_to where-to-save-normalized-data

To train a new classifier on the SVHN data:

• python svhn_classifier.py --train --data_set_path path-to-normalized-data

To view training statistics:

• tensorboard --logdir log_dir/

Check out command line arguments for further control over the hyperparameters used for training:

• python svhn_classifier.py --help