Using all convoluntional network to train and test on CIFAR10 and CIFAR100
This is an re-implement of the paper "STRIVING FOR SIMPLICITY: THE ALL CONVOLUTIONAL NET".
I have implemented two of basic network architectures in this ACN paper with Keras: ConvPool-CNN-C and All-CNN-C.
I have made all operations described in this paper except ZCA whitening, which is found not to help the metwork converge. This code file is conPool-CNN.py.
On CIFAR10 dataset, this model get 90.03% acc after 2000 epochs. Here is the curve for training processing.
On CIFAR100 dataset, this model get 61.99% acc after 1500 epochs.
I have made all operations described in this paper except ZCA whitening. The code file is allconv_cifar10.py and allconv_cifar100.py.
On CIFAR10 dataset, this model get 91.91% after 3000 epochs (test best in 2993).
On CIFAR100 dataset, this model get 66.13% after 3500 epochs (test best in 3229).