This is quick evaluation of #filters impact on performance on ImageNet-2012.
The architecture is similar to CaffeNet, but has differences:
- Images are resized to small side = 128 for speed reasons.
- fc6 and fc7 layers have 2048 neurons instead of 4096.
- Networks are initialized with LSUV-init
- No LRN layers.
Default augmentation: random crop 128x128 from 144xN image, 50% random horizontal flip.
| Name | Accuracy | LogLoss | Comments |
|---|---|---|---|
| 4sqrt(2)x wider | 0.565 | 1.96 | Start overfitting |
| 4x wider | 0.563 | 1.92 | Still no overfitting %) |
| 2sqrt(2)x wider | 0.552 | 1.94 | |
| 2x wider | 0.533 | 2.04 | |
| sqrt(2)x wider | 0.506 | 2.17 | |
| Default | 0.471 | 2.36 | |
| sqrt(2)x narrower | 0.460 | 2.41 | |
| 2x narrower | 0.416 | 2.68 | |
| 2sqrt(2)x narrower | 0.340 | 3.11 | no group conv |
| 2sqrt(2)x narrower | 0.318 | 3.25 | |
| 4x narrower | 0.256 | 3.33 |
