An image classifier CNN to recognise handwritten digits.
Data was loaded from Keras library.
Keras requires an extra dimension in the end which corresponds to channels. Images from
MNIST dataset are gray-scaled, so only one channel is used.
The labels (numbers 0 to 9) were encoded to one hot vectors (e.g : 1 : [1,0,0,0,0,0,0,0,0,0]).
Neural network models often cannot be trained on raw pixel values, such as pixel values in the range of 0 to 255. The reason is that the network uses a weighted sum of inputs, and for the network to both be stable and train effectively, weights should be kept small. Instead, the pixel values must be scaled prior to training. Normalization is often the default approach as we can assume pixel values are always in the range 0-255, making the procedure very simple and efficient to implement. Conse- quently, pixel values are scaled to the range 0-1.
• 1 2D-Convolution Layer: (3,3) is the dimensionality space of output, ReLU is the activation function. HE initializer performs better than normal that's why is selected.
• 1 Max-Pooling Layer: Downsampling of data
• 1 Flatten Layer: Data is attened so it can be passed to dense layer (keeping 1 dimension)
• 2 Dense Layers: Dense layers are used when association can exist among any feature to any other feature in data point. Since between two layers of size n1 and n2, there can n1n2 connections and these are referred to as Dense. The first one contains a ReLU activation function while the second is the softmax layer.
• Optimizer: Gradient descent is a good generalized optimizer (Adam can be used as well) • Loss function: Since a softmax output layer was used, the Cross-Entropy loss was the right loss function
In order to avoid over-fitting problem, we need to expand artificially our handwritten digit dataset. Data augmentation is a strategy that enables to significantly increase the diversity of data available for our training model. For this part, ImageDataGenerator from Keras Image Data Preprocessing was used to generate batches of tensor image data.
The following arguments were configured to aug-
ment the original data:
• Randomly rotation by 10 degrees
• Randomly Zoom by 10%
• Randomly shift images horizontally by 10% of the width
• Randomly shift images vertically by 10% of the height
• Divide inputs by std of the dataset
• Divide each input by its std
Last but not least, a manual gamma correction function was created and passed along the ImageDataGenerator to apply gamma processing to each image.
Model was finally fitted to the original dataset, merged with the augmented.
Validation was performed on the validation data.
The steps per epoch was calculated as train-length / batch-size, since this uses all of the
data points, one batch size worth at a time.
With the ReduceLROnPlateau function from Keras.callbacks, we reduced the Learning
Rate by half if the accuracy is not improved after 3 epochs.