https://github.com/markdtw/condensenet-tensorflow

tensorflow implementation of CondenseNet: An Efficient DenseNet using Learned Group Convolutions
https://github.com/markdtw/condensenet-tensorflow

Last synced: 8 days ago
JSON representation

tensorflow implementation of CondenseNet: An Efficient DenseNet using Learned Group Convolutions

• Host: GitHub
• URL: https://github.com/markdtw/condensenet-tensorflow
• Owner: markdtw
• Created: 2018-01-26T08:34:42.000Z (almost 7 years ago)
• Default Branch: master
• Last Pushed: 2018-02-01T14:07:45.000Z (almost 7 years ago)
• Last Synced: 2024-08-01T22:50:09.226Z (3 months ago)
• Language: Python
• Size: 20.5 KB
• Stars: 29
• Watchers: 4
• Forks: 15
• Open Issues: 3
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-image-classification - unofficial-tensorflow : https://github.com/markdtw/condensenet-tensorflow
• awesome-image-classification - unofficial-tensorflow : https://github.com/markdtw/condensenet-tensorflow

README

        
# CondenseNet tensorflow

Tensorflow implementation of [CondenseNet: An Efficient DenseNet using Learned Group Convolutions](https://arxiv.org/abs/1711.09224). The code is tested with cifar10, *inference phase not implemented yet*.

![Model architecture](https://i.imgur.com/f98IK2e.png)

Official PyTorch implementation by @ShichenLiu [here](https://github.com/ShichenLiu/CondenseNet).

## Prerequisites

- Python 2.7+ (3.5+ is recommended)

- [NumPy](http://www.numpy.org/)

- [TensorFlow 1.0+](https://www.tensorflow.org/)

## Data

- [CIFAR-10](https://www.cs.toronto.edu/~kriz/cifar.html)

## Preparation

- Go to `data/` folder and run `python2 generate_cifar10_tfrecords.py --data-dir=./cifar-10-data`. This code is directly borrowed from tensorflow official repo and have to be run with python 2.7+.

## Train

Use default parameters:

```bash

python main.py

```

Check out tunable hyper-parameters:

```bash

python main.py --help

```

Other parameters including `stages, groups, condense factor, and growth rate` are in `experiment.py`.

## Notes

- Training for 300 epochs with the default settings reach testing accuracy 93.389% (paper report is 94.94%). There might be some details I didn't notice, feel free to point them out.

- All the default parameters settings follow the paper/official pytorch implementation.

- Current implmentations of standard group convolution and learned group convolution are very inefficient (a bunch of reshape, transpose and concat), looking for help to build much more efficient graph.

- Evaluation phase (index select) has not been implemented yet, looking for potential help as well :D.

- Issues are welcome!

## Resources

- [The paper](https://arxiv.org/abs/1711.09224).

- [Official PyTorch Implementation](https://github.com/ShichenLiu/CondenseNet).