Keras implementation of DenseNets
Original paper: https://arxiv.org/abs/1608.06993
Original implementation: https://github.com/liuzhuang13/DenseNet
@inproceedings{
huang2017densely,
title={Densely connected convolutional networks},
author={Huang, Gao and Liu, Zhuang and van der Maaten, Laurens and Weinberger, Kilian Q },
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
year={2017}
}
Introduction to each folder and file:
"Data": put the CIFAR-100 data set here
"log": training log for model with no augmentation, relates to main.py
"Pictures": the place store generated visualized samples, relates to get_img_samples_and_conv_results.py
"Preprocess": the folder stores pre-processing files
load_data.py: load CIFAR-100 data set
normalize.py: functions used to normalize data by mean and variance way
utils: some functions used to visualize samples from data set
"weights": stores trained model weights with no augmentation
main.py: run this Python script if you want to test the data with no augmentation
main_aug.py: run this Python script if you want to evaluate data with augmentation
model.py: model building using Keras
predict.py: Run this Python script to see predicted results and generate confusion matrix, change file names or file
paths to get corresponding data for successful running
Notice: this is the re-implementation of DenseNets, but not detailed ones as same as the original paper.
Differences like pre-processing, normalize method and hyper-parameters settings.