We manually converted the original torch models into caffe format from https://github.com/liuzhuang13/DenseNet.
For details of these networks, please read the original paper:
The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN)
| Network | Top-1 | Top-5 | Download | Architecture |
|---|---|---|---|---|
| DenseNet 121 (k=32) | 74.91 | 92.19 | caffemodel (30.8 MB) | netscope |
| DenseNet 169 (k=32) | 76.09 | 93.14 | caffemodel (54.6 MB) | netscope |
| DenseNet 201 (k=32) | 77.31 | 93.64 | caffemodel (77.3 MB) | netscope |
| DenseNet 161 (k=48) | 77.64 | 93.79 | caffemodel (110 MB) | netscope |
Update (July 27, 2017): for your convenience, we also provide a link to these models on Baidu Disk.
Due to compatibility reasons, several modifications have been made:
- BGR mean values [103.94,116.78,123.68] are subtracted
- scale: 0.017 is used, instead of the original std values for image preprocessing
- ceil_mode: false is used in the first pooling layers ('pool1')