Prabhu
Feb 19, 2015, 6:01:19 PM2/19/15
to caffe...@googlegroups.com
i am trying to run regression on a lenet style deep neural network on grayscale images with 5 conv layers, 2 pooling, RELU, 2 IP layers and 1 dropout layer. The network is giving a poor performance. Look at the loss numnbers below
Simple Perceptron model - 33% error in kaggle
DNN with caffe - training 10% error, 28% test error, 44% (error when uploadded to kaggle)
It seems a simple NN perceptron model outperforms caffe.
Why is the 44% so large difference. Is the network need to be optimized? If yes what kind of optimization. i already tried normalizing inputs 0 and 1, but i cant get past the 44%
Should convolutional deep neural network perform much better on images.
Simple Perceptron model - 33% error in kaggle
DNN with caffe - training 10% error, 28% test error, 44% (error when uploadded to kaggle)
It seems a simple NN perceptron model outperforms caffe.
Why is the 44% so large difference. Is the network need to be optimized? If yes what kind of optimization. i already tried normalizing inputs 0 and 1, but i cant get past the 44%
Should convolutional deep neural network perform much better on images.
Reply all
Reply to author
Forward
0 new messages