Default ResNet-18 with one quarter of the usual width and its performance can be seen in this notebook (with outputs).
The best accuracy achieved (on K49) is 98.29%. Please excuse K49 being called KMNIST inside of the notebook.
As minor dataset augmentation and adaptive learning rate scheduling are used, which are both not part of the original ResNet paper, this might have to be marked differently though.
Below "Dense" means that the InDeDeNet was used. Exploration is done on K49, implying that results can be shared as well.
Using the notebook above, here are a few more configurations
| Dense |
Width Factor |
Depth |
Parameters |
Accuracy |
| False |
1/4 |
10 |
535.2k |
96.82 |
| False |
1/3 |
10 |
584.4k |
97.22 |
| True |
1/4 |
10 |
598.3K |
97.52 |
| False |
1/3 |
18 |
1.2M |
98.14 |
| False |
1/4 |
101 |
2.6M |
98.20 |
| False |
1/2 |
18 |
2.8M |
98.28 |
| True |
1/2 |
10 |
2.4M |
98.28 |
Default ResNet-18 with one quarter of the usual width and its performance can be seen in this notebook (with outputs).
The best accuracy achieved (on K49) is 98.29%. Please excuse K49 being called KMNIST inside of the notebook.
As minor dataset augmentation and adaptive learning rate scheduling are used, which are both not part of the original ResNet paper, this might have to be marked differently though.
Below "Dense" means that the InDeDeNet was used. Exploration is done on K49, implying that results can be shared as well.
Using the notebook above, here are a few more configurations