search

dnouri / cuda-convnet

My fork of Alex Krizhevsky's cuda-convnet from 2013 where I added dropout, among other features.
http://code.google.com/p/cuda-convnet/
254 stars 147 forks source link

[noccn] Make batch File wont work on 128 and 256 images #2

Closed teknosains closed 10 years ago

teknosains commented 10 years ago

Haloo, did it because of the size of the images too large ?

and do you have any suggestions about layers n params that best match for the 64 image ?

I trained using your dropout, 15 classes (14million images, 7 batches), cifar-cropped , it was 36% error..... how to get lower error ?

thanks

dnouri commented 10 years ago

Do you have a traceback for us to look at regarding your image size problem? What's the error you're getting? And please report your issue with noccn over in the noccn bug tracker. (I will close the issue here.)

For 15 classes and 14 million images you want to make your network very large. You might try something like the net that Krizhevsky used on ImageNet; five convolutional layers, some of them with max pooling, and two fully connected layers. There's some configuration files floating around on the internet, which you can use for inspiration. The syntax will slightly differ though.