The results can be quickly inspected / exported with the results.py script. Supported output modes are shown in table below. Data can be exported by appending --df {output dir} .
| Output mode | Description |
|---|---|
scatter-psnr / scatter-ssim |
show accuracy / image quality trade-off (NIP optimization) |
progress |
show training progress |
conf / conf-tex |
print confusion matrices as plain text or LaTeX tables |
ssim / psnr/accuracy |
boxplots with image quality / accuracy |
df |
print results' summary as a table |
auto |
automatically parse the results' structure and plot accuracy for various configurations |
For example, the following command shows the scatter plot with the trade-off between classification accuracy and image fidelity for the UNet model trained on Nikon D90 :
> python3 results.py --nip UNet --cam D90 scatter-psnr
To visualize variations of classification accuracy and image quality as the training progresses:
> python3 results.py --nip UNet --cam D90 progress
To show confusion matrices for all regularization strengths:
> python3 results.py --nip UNet --cam D90 confusion
Show Differences in NIP models
This command shows differences between a UNet model trained normally (A) and with manipulation detection objectives (B).
> python3 diff_nip.py --nip UNet --cam D90 --b ./data/m/D90/UNet/ln-0.1000/000/models/ --image 16