![]()
I used this kernel in the Kannada MNIST Competition, getting a final Private Score of 0.99040 and final Public Score of 0.98920, approximately 60th out of 1213 ( Top 5% ) on LB. Here is how this kernel implemented. The CNN architecture is based on kernel of FWiktor. Thanks a lot to him. CNN Architecture First of all, here is the architecture of FWiktor’s network: Based on this summary, I implemented a network and reached an accuracy of 85% on val_set( Dig-MNIST.csv ). Well, the result is fairly good, but I want a even higher accuracy, something like 95% or even 99%. In order to achieve that, I adjusted some layers of the neural network structure and added some layers as well( Will mention it later ). Here is the summary of my network: Differences Between FWiktor’s And Mine Two More Conv2d Layers When I trained my model on the FWikor’s network, I found the accuracy of val_set drastically improved from 10% to 80% in a very short time, like in 5 epochs, then it continued to improve a little bit to 85% in the approximately next 40 epochs and finally remained the same no matter how much longer you t…
