Implement awesome CNNs with Keras

Keras2.1.0+, Tensorflow1.4+

1.resnet_cifar.py

nb_classes: the number of your dataset classes, for cifar-10, nb_classes should be 10

img_dim: the input shape of the model input

nb_blocks: the number of blocks in each stage,the depth of the model = 2 + 2 x (nb_blocks[0] + nb_blocks[1] + nb_blocks[2])

k: the widen fatcor, k=1 indicates that the model is original ResNet, when k>1 the model is a wide ResNet

weight_decay: weight decay for L2 regularization

droprate: the dropout between two convolutons of each block is added and the default drop rate is set to 0.0

return: ResNet model or WRN model

from resnet_cifar import create_ResNet

resnet = create_ResNet(

                    nb_classes = 10,
                    
                    img_dim = (32, 32, 3),
                    
                    nb_blocks = [18, 18, 18],
                    
                    k = 1,
                    
                    droprate = 0.0
                    )