https://github.com/gaohuang/DenseNet_lite

A more memory efficient Torch implementation of "Densely Connected Convolutional Networks".
https://github.com/gaohuang/DenseNet_lite

Last synced: 3 months ago
JSON representation

A more memory efficient Torch implementation of "Densely Connected Convolutional Networks".

• Host: GitHub
• URL: https://github.com/gaohuang/DenseNet_lite
• Owner: gaohuang
• License: mit
• Created: 2016-09-21T19:48:26.000Z (almost 8 years ago)
• Default Branch: master
• Last Pushed: 2017-05-11T20:06:43.000Z (about 7 years ago)
• Last Synced: 2024-01-22T10:08:36.798Z (5 months ago)
• Language: Lua
• Size: 4.88 KB
• Stars: 30
• Watchers: 4
• Forks: 5
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Lists

• awesome-very-deep-learning - Torch Implementation

README

        
# DenseNet_lite

This implements the DenseNet architecture introduced in [Densely Connected Convolutional Network](http://arxiv.org/abs/1608.06993).The original Torch implementation can be found at https://github.com/liuzhuang13/DenseNet, and please find more details about DenseNet there. The only difference here is that we write a customed container "DenseLayer.lua" to implement the dense connections in a more memory efficient way. This leads to **~25% reduction** in memory consumption during training, while keeps the accuracy and training time the same. 

## Usage 

0. Install Torch ResNet (https://github.com/facebook/fb.resnet.torch) following the instructions there. To reduce memory consumption, we recommend to install the [optnet](https://github.com/fmassa/optimize-net) package. 

1. Add the files ```densenet_lite.lua``` and ```DenseLayer.lua``` to the folder models/;

2. Insert ```require 'models/DenseLayer``` at Line.89 of ```models/init.lua```, if you need to use multiple GPUs;

3. Change the learning rate schedule at function learningRate() in ```train.lua``` (line 171/173),

from 

```decay = epoch >= 122 and 2 or epoch >= 81 and 1 or 0```

to 

 ```decay = epoch >= 225 and 2 or epoch >= 150 and 1 or 0 ```

 

4. Train a DenseNet (L=40, k=12) on CIFAR-10+ using

```

th main.lua -netType densenet_lite -depth 40 -dataset cifar10 -batchSize 64 -nEpochs 300 -optnet true

```