This project is an inquiry into DenseNet and CondenseNet, specifically, the effect of having dense connections. The project develops a network class with a configurable rate of dense connection, as well as an empirical analysis on the effect of dense connection on network performance and training time.
Interactive web application to show how dense connections impact model performance using a 6-layers CondenseNet architecture. Access the application here: cnn-dense-connection.boyu.io
DenseNet is a compact network architecture, it introduces dense connections between network layers. Dense connection constructs an architecture with more connectivity compare to a traditional sequential network with similar parameter size.
CondenseNet architecture is one step further on DenseNet. In addition to dense connection with in the same block, CondenseNet keeps doing dense connection and re-use feature over the entire network.
Training under default configuration
python main.py
Options
# model architecture setting
--cross_block_rate=0.5 # dense connection rate between dense blocks, use 0 for original DenseNet and 1 for CondenseNet
--end_block_reduction_rate=0.5 # feature channel reduction rate at the end of each dense block
# model configuration
--stages='10-10-10' # layers for each stage
--growth='8-16-32' # growth rate for each stage
--group_1x1=4 # number of 1-by-1 group convolution
--group_3x3=4 # number of 3-by-3 group convolution
--bottleneck=4 # bottleneck
--optimizer='sgd' # optimizer function
--lr=0.1 # learning rate
--scheduler='cos' # learning rate scheduler
--ep=120 # number of epoch to train
--bsize=512 # training batch size
--one_batch # train only the first batch of every epoch, for fast local code testing
# parallel gpu setting
--parallel # use parallel gpu training
--n_gpu=4 # number of gpu used for parallel
# save model
--save_folder='default' # folder to save model and training details
TBD
├── evaluation
│ ├── model_eval.py
│ └── model_select.py # plot performance figure
├── example
│ ├── result # json files, result for models in this report
│ └── example.ipynb # python notebook, demo training and cresult
├── model
│ ├── architectures.py
│ ├── densenet.py
│ └── helpers.py
├── playground
│ └── README.md # information on interactive playground web app
├── augment.py # data augmentation function
├── data.py # load data to pytorch dataloader
├── experiment.sh # bash script, train and test all models in this report
└── main.py # entry point from network training
Method | Depth | Params | C10 | C10+ |
---|---|---|---|---|
DenseNet | 30 | 0.24M | - | 15.45 |
DenseNet (cbr=0.2) | 30 | 0.26M | - | 14.18 |
DenseNet (cbr=0.4) | 30 | 0.29M | - | 14.80 |
DenseNet (cbr=0.6) | 30 | 0.31M | - | 15.65 |
CondenseNet | 30 | 0.32M | - | 13.95 |
CondenseNet (cbr=0.8) | 30 | 0.29M | - | 13.98 |
CondenseNet (cbr=0.6) | 30 | 0.27M | - | 15.06 |
CondenseNet (cbr=0.4) | 30 | 0.25M | - | 14.78 |
DenseNet | 78 | 1.09M | - | 8.08 |
DenseNet (cbr=0.2) | 72 | 1.08M | - | 7.99 |
CondenseNet (cbr=0.8) | 66 | 1.07M | - | 7.74 |
* C10+
represent cifar-10 dataset with data augmentation.