Densely Connected Convolutional Network (DenseNet)
This repository contains the caffe version code for the paper Densely Connected Convolutional Networks.
For a brief introduction of DenseNet, see our original Torch implementation.
ImageNet Pretrained Models
See https://github.com/shicai/DenseNet-Caffe for caffe prototxt and pre-trained models.
See https://github.com/liuzhuang13/DenseNet for Torch pre-trained models.
See http://pytorch.org/docs/torchvision/models.html?highlight=densenet for directly using the pretrained models in PyTorch.
Note
- The models in this repo are for CIFAR datasets only (input 32x32). If you feed images with larger resolution (e.g., ImageNet images), you need to use a different downsampling strategy to keep the memory usage reasonable. See our paper or Torch code for details on ImageNet models.
- The code in this repo doesn't support BC-structres. However, it should be easy to modify.
- This code is not the code we use to obtain the results in the original paper, the details (such as input preprocessing, data augmentation, training epochs) may be different. To reproduce the results reported in our paper, see our original Torch implementation.
Results
The default setting (L=40, k=12, dropout=0.2) in the code yields a 7.09% error rate on CIFAR10 dataset (without any data augmentation).
Usage
- Get the CIFAR data prepared following the Caffe's official CIFAR tutorial.
- make_densenet.py contains the code to generate the network and solver prototxt file. First change the data path in function make_net() and preprocessing mean file in function densenet() to your own path of corresponding data file.
- By default make_densenet.py generates a DenseNet with Depth L=40, Growth rate k=12 and Dropout=0.2. To experiment with different settings, change the code accordingly (see the comments in the code). Example prototxt files are already included. Use
python densenet_make.py
to generate new prototxt files. - Change the caffe path in train.sh. Then use
sh train.sh
to train a DenseNet.
Contact
liuzhuangthu at gmail.com
gh349 at cornell.edu
Any discussions, suggestions and questions are welcome!