Comparison between state of the art computer vision models. Three popular models AlexNet, VGG and ResNet were put into the test. The task was image classification on CIFAR-10.
Rather than exploring many different hyper parameter settings for every model, we performed tests using three different optimization methods while employing six different activation functions as well as performing regularization with
Accuracies for each configuration can be seen in accuracies_with_weight_decay.csv (weight decay = 0.01, one run with learning rates set to 0.01 for RMSprop, 0.001 for Adam and 0.01 for SGD; and one run with learning rates set ti 0.1 for RMSprop, 0.0001 for Adam and 0.001 for SGD) and accuracies_without_weight_decay.csv (run with learning rates set to 0.01 for RMSprop, 0.001 for Adam and 0.01 for SGD).
Similarly, loss can be found in loss_with_weight_decay.csv and loss_without_weight_decay.csv. For each configuration, there are 2 rows: one for training loss and one for validation loss.