I am looking for a simple CNN script using keras that performs close to the state fo the art on CIFAR10. I've found some scripts that can barely touch 90% accuracy, but they all suffer the same defaults. Namely, they are not reliable (several runs needed to reach 90%), consume a huge amount of memory and run for ages.
Does anybody out there have a reasonable script that reliably achieve over 90% without taking days to train?
Do you know Caffe?
Caffe is a library developed by Berleley University that uses CUDA and is very fast. Caffe have many examples, among them have a Cifar10-example.
I hope that this information can help you.
http://caffe.berkeleyvision.org/
I know about caffe. I'm not looking for a deep net library, since I'm already using one (Keras). I'm just looking for a model that can be easily trained and performs over 90% (or even better over 92%) accuracy on CIFAR10.
Thank you nonetheless.
Your question is very specific.
Your best chance should be the forum at Kaggle. Just ask them.
https://www.kaggle.com/c/cifar-10/forums
Or ask the 20+ submitters with a score above 0.9
https://www.kaggle.com/c/cifar-10/leaderboard/public
Answering my own question: The following code from the official examples do indeed achieve over 90% reliably.
https://github.com/keras-team/keras/blob/master/examples/cifar10_resnet.py
- Badges
- Science topic