This is an unoffcial implementation of the ResNet in Keras (concept and architectue: https://arxiv.org/pdf/1512.03385.pdf).
This implementation alows in a fexible way to generate the ResNet 18,34,50,101,152 model variants.
Other ResNet configurations (like number of resnet blocks, or filter number) can be easily changed. Identity shortcuts on dimension transition use 1x1 convolution to match the desired output, see the paper for details.
- clone the project
- select one of the architectures (
RESNET_MODELS = {"RESNET_18":0,"RESNET_34":1,"RESNET_50":2,"RESNET_101":3,"RESNET_152":4}
) with i.e.SELECTED_ARCHITECTURE = RESNET_MODELS['RESNET_18']
- or customize the block / filter lists from the notebook. Default configuration is the ResNet18 configuration
block_list = [2,2,2,2]
andfilter_list = [64,128,256,512]
- Train the model on your data and use it in your projects
/Enjoy.