Reimplementation of DenseNet on Image Recognition
Densely Connected Convolutional Networks(DenseNet), won the Best Paper Award on CVPR 2017.
This is an (re-)implementation of DenseNet in TensorFlow for image recognition tasks. The (re-)implementation is based on official Torch DenseNet with Tensorflow Slim.
In the paper, DenseNet-264 seems to be a typo, since there is no way the number of layers to be an even number. See DenseNet-121, 169, 201, if you add up, which is 1 + 6 x 2 + 1 + 12 x 2 + 1 + 64 x 2 + 1 + 48 x 2 + 1 + 1 = 265.
- DenseNet-B, DenseNet-C, DenseNet-BC
- DenseNet-121, DenseNet-169, DenseNet-201, DenseNet-265
- Training on CIFAR
- Training on SVHN
- Training on ImageNet
python 3.5
tensorflow 1.4
CUDA 8.0
cuDNN 6.0
python 3.5
tensorflow 1.2
CUDA 8.0
cuDNN 5.1
pip3 install -r requirements.txt
| L=100, k=12 | Accuracy | Top 1 Error |
|---|---|---|
| paper | 94.08% | 5.92% |
| repo | 95.75% | 4.25% |
| L=100, k=12 | Accuracy | Top 1 Error |
|---|---|---|
| paper | 75.85% | 24.15% |
| repo | 77.25% | 22.75% |
