A simple project that is created to support the blog post @ http://hackevolve.com/recognize-handwritten-digits-1 and http://hackevolve.com/recognize-handwritten-digits-2
It is simply an Hand written digit recognizer, It takes an image and recognizes the digits in them. This can be extended to recognizing characters as well.
utils/dataset.py- A simple utility that loads the dataset from disk, encodes, plots the datasetrecognize.py- It is the driver script which takes an image and recognizes the digits in it. It takes two command line arguments namely --model and --image.train_keras.py- A simple script which trains a model on MNIST dataset and save to disk using keras.train_tf.py- It's just an tensorflow implementation of the above.train_keras_cnn.py- It trains the model on Convolution Neural Network and save the model to disk.train_tf_cnn.py- It's just an tensorflow implementation of the above.
The below text represents the CNN model implemented in this project
Input---> shape(-1,28,28,1)
Convolution Layer 1 ---> filter_shape:(3,3), no.of.filters: 32, strides: 1, padding: valid
Activation ---> Relu
MaxPooling ---> pool_size: (2,2) i.e., halves the image, padding: valid
Convolution Layer 2 ---> filter_shape:(3,3), no.of.filters: 32, strides: 1, padding: valid
Activation ---> Relu
MaxPooling ---> pool_size: (2,2) i.e., halves the image, padding: valid
Dropout ---> 0.75 (keep probability)
Fully Connected Layer 1 ---> num_units: 128,input_dim: 5*5*32
Activation ---> Relu
Dropout ---> 0.5 (keep probability)
Fully Connected Layer 2 ---> num_units: 10, input_dim: 128
Optimizer ---> Adadelta, learning_rate: 0.001
fit ---> batch_size: 128, nb_epochs: 12
The accuracy obtained is more than 99% using the Convolution Neural Network model 96% while using simple Multi Layer Perceptron.
