cifar100-cnn-01

Image Classifier using CNNs on the CIFAR100 Image Dataset

Data from Learning Multiple Layers of Features from Tiny Images, Alex Krizhevsky, 2009.

Implementation:

• This method uses a rather simple CNN with ELU(Exponential Linear Unit) rather than more conventional RELUs.
• Has about 1.6 Million Parameters for the "Model_B" which is the more accurate Model.
• After Training for 100 epochs we get ~ 44 % Accuracy on the validation(20% of the training data) and Test Set.
• Weights are saved and so is the model, so as to skip training from scratch.
• Has another model(Model_A) which is a much more deeper network, but that yields only ~30% accuracy after 45 epochs of training.

Confusion Matrix with Model_B after 100 epochs: