ADAPT: Adaptive DCT Filters for Enhanced Image Recognition
For CIFAR10, ResNet20 model : For BCE loss:
| Corruption (sev=5) | Without Edge | With True Edge |
|---|---|---|
| Original | 92.45 | 92.45 |
| Fog | 59.76 | 75.97 |
| Gaussian Noise | 21.03 | 27.77 |
| Zoom Blur | 68.01 | 79.179 |
| Jpeg | 67.96 | 80.24 |
| Brightness | 86.63 | 84.28 |
| Contrast | 25.12 | 48.74 |
| Shot noise | 25.58 | 33.28 |
| Snow | 69.31 | 72.09 |
| Frost | 75.41 | 78.13 |
| Average without original | 52.9 | 63.5 |
- For brightness, it seems that better looking images does not always mean better performance. For all severities it doesnt seem to help..the color info goes awol ... maybe its only useful when everything is too occluded so adding some info then helps the model But for lower severities it seems to hurt performance. So maybe we need to combine it with TTA loss or metric to prevent overcorrecting ..?
For Cos loss: For CIFAR10, ResNet20 model :
| Corruption (sev=5) | Without Edge | With True Edge |
|---|---|---|
| Original | 92.45 | 92.45 |
| Fog | 60.2 | 87.57 |
| Gaussian Noise | 20.9 | 91.93 |
| Zoom Blur | 66.737 | 90.15 |
| Jpeg | 66.68 | 90.32 |
| Brightness | 85.05 | 83.18 |
| Contrast | 24.65 | 49.33 |
| Shot noise | 25.42 | 90.16 |
| Snow | 67.44 | 87.62 |
| Frost | 73.84 | 85.89 |
| Average without original | 52.9 | 82.5 |
Combined : For CIFAR10, ResNet20 model : Corruption (sev=5) | Without Edge | With Cos Edge | With True Edge --- | --- | --- Original | 92.45 | 92.45 | 92.45 Fog | 60.2 | 87.57 | 75.97 Gaussian Noise | 20.9 | 91.93 | 27.77 Zoom Blur | 66.737 | 90.15 | 79.179 Jpeg | 66.68 | 90.32 | 80.24 Brightness | 85.05 | 83.18 | 84.28 Contrast | 24.65 | 49.33 | 48.74 Shot noise | 25.42 | 90.16 | 33.28 Snow | 67.44 | 87.62 | 72.09 Frost | 73.84 | 85.89 | 78.13 Average without original | 52.9 | 82.5 | 63.5
Sequence of things to do :
- Run for clean and entropy on clean
- pure and entropy on cc
- run anal on both test and train
- run clean plot script
- run ensemble plot script
Credits : thanks to the following repos for the code and inspiration :
- https://github.com/alexrame/mixmo-pytorch.git for the pretrained wideresnet model of CIFAR100
Color map hex codes: ['#e3d9c1', '#d4aa9b', '#be7c89', '#965680', '#5f396a', '#27213f']
conda env :/home/machiraj/miniconda3