lpzhang / pre-resnet-gen-caffe

Generate Caffe prototxt of pre-activation ResNet on ImageNet

Pre-ResNet Generator

A simple python script to generate Caffe prototxt of pre-activation ResNet on the ImageNet (ILSVRC) dataset. See Identity Mappings in Deep Residual Networks and fb's torch implementation.

Usage

1. Set the CAFFE_ROOT (line 8) properly.
2. python2 net_def.py --depth 200

Depth can be 50, 101, 152, or 200. Will output resnet{depth}_trainval.txt respectively by default.

Note

1. The script only serves as a network scaffold. More data augmentation schemes should be added to the data layers to achieve good performance.

2. Some pre-activation BatchNorm layers could be further made inplace to save memory. Will double check the correctness of doing so.