$ python main.py --block_type bottleneck --depth 100 --growth_rate 12 --compression_rate 0.5 --drop_rate 0 --outdir results
| Model | Test Error (median of 3 runs) | Test Error (in paper) | Training Time |
|---|---|---|---|
| DenseNet-100 (k=12) | 4.10 (1 run) | ||
| DenseNet-100 (k=24) | 3.74 (1 run) | ||
| DenseNet-BC-100 (k=12) | 4.69 | 4.51 (1 run) | 15h20m |
| DenseNet-BC-250 (k=24) | 3.62 (1 run) | ||
| DenseNet-BC-190 (k=40) | 3.46 (1 run) |
NOTE This model was trained with batch size 32 and initial learning rate 0.05. In the paper, they use batch size 64 and initial learning rate 0.1.
- Huang, Gao, et al. "Densely connected convolutional networks." The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4700-4708. arXiv:1608.06993, Torch implementation