CondenseNet-PyTorch

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

├── 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.

图片.png

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.