Enhance Caffe CIFAR-10 Quick model prediction accuracy
229 views
Skip to first unread message
Harsh Wardhan
Nov 2, 2015, 6:28:53 AM11/2/15
to Caffe Users
Hi,
I'm trying to enhance the accuracy of the Caffe CIFAR-10 Quick model. First I applied translation on all the images of the training set (50,000 images) and test set (10,000 images) and created the LMDB database for the complete 60,000x2 images. When I run the training I get an accuracy of 72%. The original dataset gives an accuracy of 75%. Thereafter, I replaced the test dataset with the original 10,000 images (without any translation) from the CIFAR-10 dataset. The train dataset remains the new enhanced dataset with translation (Total of 50,000x2 images). But then the accuracy drops to about 50%. Am I doing something wrong with the mean computation or dataset creation?
I'm trying to enhance the accuracy of the Caffe CIFAR-10 Quick model. First I applied translation on all the images of the training set (50,000 images) and test set (10,000 images) and created the LMDB database for the complete 60,000x2 images. When I run the training I get an accuracy of 72%. The original dataset gives an accuracy of 75%. Thereafter, I replaced the test dataset with the original 10,000 images (without any translation) from the CIFAR-10 dataset. The train dataset remains the new enhanced dataset with translation (Total of 50,000x2 images). But then the accuracy drops to about 50%. Am I doing something wrong with the mean computation or dataset creation?
Reply all
Reply to author
Forward
0 new messages