Readme.md

Cifar 10 residual network models.

The script build_resnet.py can be used to generate cifar-10 models described in the residual networks paper.

usage: build_resnet.py [-h] -n N -m MAIN_BRANCH [-f FIRE_FILTER_MULT]
                       [-o OUTPUT_FOLDER]

This script generates cifar10 resnet train_val.prototxt files

optional arguments:
  -h, --help            show this help message and exit
  -n N, --N N           Number of block per stage (or N), as described in
                        paper. Total number of layers will be 3N + 2
  -m MAIN_BRANCH, --main_branch MAIN_BRANCH
                        normal, bottleneck
  -o OUTPUT_FOLDER, --output_folder OUTPUT_FOLDER
                        Train and Test prototxt will be generated as
                        train.prototxt and test.prototxt

As an example the resnet_110 folder contains the prototxt files generated for training the 110 layer network for cifar 10 dataset. The model files can be generated as follows:

python app/cifar10/build_resnet.py -m bottleneck -n 36 -o ./

Output:
....
....
Number of params:  0.220944  Million
Number of flops:  30.73296  Million

Some results

Note - We do not use augmentation, the numbers are 2-3% below the ones reported in original paper, this is just to demonstrate how to use pynetbuilder to reproduce residual networks.

• Training batch size 128
• LR - 0.1, gamma 0.1. Steps 32K, 48k.
• Iterations - 60K

Model	Accuracy
Resnet_20	0.8795
Resnet_32	0.8922
Resnet_44	0.892
Resnet_56	0.8896
Resnet_110	0.8921