CIFAR-10 Image Classification#

Data Preprocessing#

Using tensorflow and keras, load the cifar-10 dataset into a (trainX,trainY) (testX,testY) split. Then normalize the pixels of trainX and testX to the range [0,1] to improve model stability. The possible labels for this dataset of images are [‘Airplane’,‘Automobile’,‘Bird’,‘Cat’,‘Deer’,‘Dog’,‘Frog’,‘Horse’,‘Ship’,‘Truck’]. Next, encode the labels (trainY, testY) into a 10 dimensional vector to do multiclass classification. Visuals of 16 labeled images are plotted using pyplot.

Model Initialization#

Since the dataset consists of images, a convoluted neural network is used to classify data. The general process of creating the model is as follows:

create 2 convoluted layers with relu activation to extract spatial features (32, (3,3))
add MaxPooling2D to reduce the feature map size
use Dropout to prevent overfitting by disabling neurons randomly when training (0.25 until flattening layers)
create 2 more convoluted layers (64,(3,3)) with MaxPooling2D and Dropout layers added after
create 2 more convoluted layers (128,(3,3)) with MaxPooling2D and Dropout layers added after
flatten the layers with model.add(layers.Flatten())
add a Dense layer with 512 and relu activation
add the final Dropout layer with 0.5
add the final Dense layer with 10 and softmax activation (used to get probabilty scores for classes)

Then, compile the model with the adam optimizer, accuracy metric, and categorical_crossentropy as the loss function.

Model Summary

Model Training#

The CNN model is trained with:

30 epochs (cycles of training)
batch_size of 64 (meaning 64 images per batch for the best gradient updates)
validation_split to help with overfitting & generalization The loss and accuracy results are stored in the history variable for plotting model accuracy and loss.

After 20 epochs, accuracy is 0.8182, loss is 0.5139, validation_accuracy is 0.7934 and validation_loss is 0.6363.

After all 30 epochs, accuracy is 0.8439, loss is 0.45, validation_accuracy is 0.08007 and validation_loss is 0.6123.

Accuracy & Loss

The more training cycles, the less training and validation loss. Similarly the more training cycles, accuracy is higher for both training and validation data.

Overall, model accuracy is 84% and model validation_accuracy is 80%, with loss averaging about 0.62.

Model Prediction#

Recall that the possible labels are [‘Airplane’,‘Automobile’,‘Bird’,‘Cat’,‘Deer’,‘Dog’,‘Frog’,‘Horse’,‘Ship’,‘Truck’]. The process for classifying an image is as follows: