Skip to content
No description, website, or topics provided.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
imagenet_refactored init Jan 29, 2019 init Jan 29, 2019 init Jan 29, 2019

Layerwise Learned CNN

This is code associated with the paper This is a peliminary research code and some more refinements are needed.


Imagenet experiments for 1-hidden layer use the standalone

Imagenet experiments for k=2+ can be run with

Note k in the paper corresponds to nlin in the code

To obtain the results for Imagenet


python IMAGENER_DIR -j THREADS --ncnn 8 --nlin 2 


python IMAGENER_DIR -j THREADS --ncnn 8 --nlin 1 

k=1 model

python IMAGENER_DIR -j THREADS --ncnn 8


The VGG-11 model was trained with a new refactored and more modular codebase different from the codebase used for the above models and is thus run from the standalone directory refactored_imagenet/

To train the VGG-11 with k=3

python IMAGENER_DIR -j THREADS --arch vgg11_bn --half --dynamic-loss-scale

to train the baseline:

python IMAGENER_DIR -j THREADS --arch vgg11_bn --half --dynamic-loss-scale

CIFAR experiments

CIFAR experiments can be reproduced using

The CIFAR-10 models can be trained:

k=3 (~91.7)

python --ncnn 4 --nlin 2 --feature_size 128 --down [1] --bn 1

k=2 (~90.4)

python --ncnn 4 --nlin 1 --feature_size 128 --down [1] --bn 1

k=1 (~88.3)

python --ncnn 5 --nlin 0 --feature_size 256 

Refactored (similar to imagenet_refactored) to train CIFAR-10 coming soon with some improvements in accuracy.

Contact: please send questions/comments/issues to

You can’t perform that action at this time.