chainer v2 implementation of DenseNet-BC

This is an implementation of DenseNet-BC. The difference between DenseNet-BC and original DenseNet is using 1x1 convolutional layer before each 3x3 convolutional layer in DenseBlock.

3 kinds of DenseNet is defined in densenet.py: DenseNetCifar, DenseNetImagenet, & DenseNet. DenseNetCifar and DenseNetImagenet consists of 4 DenseBlocks, while you can change # of blocks and # of layers for each block by passing n_layers which is an argument of DenseNet.

Environment

• Python 3.5.2
• Chainer 2.0.0
• ChainerCV 0.5.1
• OpenCV w/ contrib 3.2.0.7

FYI: You can install OpenCV with or without contrib for python via pip: pip install opencv-python or pip install opencv-contrib-python.

• opencv-python
• opencv-contrib-python

Results of cifar100, no tuning

Whole training log is in result_cifar100/log.json

accuracy

loss

Training

In my environment (GTX 1080), one epoch including validation took around 210 seconds.

about model size

As to DenseNetImagenet or DenseNet in densenet.py, examples of # of layers are below.

# of layers	growth rate	# of layer of each block
121	32	6, 12, 24, 16
169	32	6, 12, 32, 32
201	32	6, 12, 48, 32
161	48	6, 12, 36, 32