Does anybody used cifar-100 in Caffe

[Ihsan Ullah]

Hi,
I am trying to use Cifar-100.
I know i have to change the last layer instead of 10 to 100 outputs. And there are no more difference in the network.
But to prepare the data, I tried to change the files used for cifar-10 to cifar-100.
But i got errors. 
Can anybody share the get_cifar100.sh , create_cifar100.sh , convert_cifar_data.sh files.
Regards
khan

[Guillaume Chevalier]

I have made an HDF5 version of the cifar-100 dataset while playing witht the cifar-10 dataset, but I only practiced on the cifar-10. 

I would very appreciate to see how to do classification with multiple labels at once, maybe add two loss layers and two accuracy layers for the coarse and fine labels. In my HDF5 data, instead of having a field named "label", I have both a "label_coarse" and a "label_fine" field.

The link: 

https://github.com/guillaume-chevalier/caffe-cifar-10-and-cifar-100-datasets-preprocessed-to-HDF5

[Guillaume Chevalier]

I finally did it in Python:

https://github.com/guillaume-chevalier/python-caffe-custom-cifar-100-conv-net

The CNN architecture I built is inspired from NIN and some others. 

On Friday, 5 June 2015 19:04:54 UTC-4, Ihsan Ullah wrote:

[junyu...@gmail.com]

在 2016年8月25日星期四 UTC+8下午8:58:51，junyu...@gmail.com写道：

Hi,

Recently, I try to use Cifar100 on caffe and write some script to convert cifar100 data to lmdb. Now, I successfully train with cifar100 on caffe

You can get the scripts from my github:https://github.com/fish145/uncommon-datasets-caffe/tree/master/scripts

在 2015年6月6日星期六 UTC+8上午7:04:54，Ihsan Ullah写道：

[Hossein Hasanpour]

you might want to check this branch : https://github.com/Coderx7/caffe 

the data/example section also has cifar100, it can be downloaded/created and run like cifar10.