99.542 % accuarcy on test set
I used a pre-trained VGG16 model and replaced the fully connected layers with untrained layers while freezing the fully convolutional layers.
I made 5 % validation set for my hyper-parameters tuning data was augmented to be x10 MNIST, I trained 5 classifiers for 5 epochs and did majority voting.
The results from 3 classifiers is exactly the same as 5 classifiers on the testset :3