Based on Deep Residual Learning for Image Recognition by Kaiming He et al, 2015 with pre-activation improvements from Identity Mappings in Deep Residual Networks.
Differences in implementation from original paper:
- model is trained on flowers dataset (download)
- due to smaller input image size, first layer has filter size [3x3] instead of [7x7] and there is no maxpool layer following it
- there is only bottleneck residual block implemented
- layers that downsample use bottleneck layers to adjust skip-connection size
- only implemented resnet-18 (
conf/resnet_18.json) but other networks can be easily implemented
Other ResNet architectures can be easily added by editing json files in
confdirectory
Sample training - takes ~11mins on MacBook Pro 2015
