CondenseNet-PyTorch

A PyTorch implementation of CondenseNet: An Efficient DenseNet using Learned Group Convolutions

Table of Contents:

• CondenseNet-PyTorch
  • Project structure
  • Data
  • Model
  • Experiment configs
  • Usage
  • Results
  • Requirements
  • Future Work
  • References
  • License

Project structure:

├── agents
|  └── condensenet.py # the main training agent
├── graphs
|  └── models
|  |  └── condensenet.py
|  |  └── denseblock.py
|  |  └── layers.py
|  └── losses
|  |  └── loss.py # contains cross entropy loss definition
├── datasets  # contains all dataloaders for the project
|  └── cifar10.py # dataloader for cifar10 dataset
├── data
|  └── cifar10  # contains raw dataset
├── utils # utilities folder containing metrics , config parsing, etc
|  └── assets
├── main.py
└── run.sh

Data:

Dataloader is responsible for downloading (first time only) and preparing cifar10 data.

Model:

To be able to reproduce the results from the official implementation, we use the default model of cifar10 and its configs as given here.

CondenseNet	Feature map
3x3 Conv (stride =1)	32x32
	32x32
2×2 average pool, stride 2	16x16
	16x16
2×2 average pool, stride 2	8x8
	8x8
8x8 global average pool	1x1
10-dim fully-connected

Experiment configs:

- Input size: 32x32x3
- Batch size: 64
- Learning rate: 0.1 following a consine type
- Optimizer: SGD
- Number of epochs: 300
- Condensation Stages: [14, 14, 14]
- Growth Rate: [8, 16, 32]

Usage:

• To run the project, you need to add your configurations into the folder configs/ as found here
• sh run.sh
• To run on a GPU, you need to enable cuda in the config file.

Results:

Metric	Reproduced	Official
Top1 error	4.78%
Top5 error	0.15%

Requirements:

• Pytorch: 0.4.0
• torchvision: 0.2.1
• tensorboardX: 1.2

Check requirements.txt.

Future Work:

• Add network profiling for counting the number of FLOPS
• Implement the condensation layers to optimize the model for inference

References:

• CondenseNet official implementation in PyTorch: https://github.com/ShichenLiu/CondenseNet

License:

This project is licensed under MIT License - see the LICENSE file for details.