analysis.md

🔬 Description/Usage of analysis folder

• Description of analysis Folder
  • Folder Structure
• Usage of analysis Folder
  • Find Filters For Specific Degradation
  • Compare against these methods
    • Mask Discovered Filters
  • Filter Distribution Visualization
  • Filter Modulation

Description of analysis Folder

All provided analysis scripts are contained in analysis folder.

Folder Structure

analysis
├── scripts
|   ├── cal_blur_gradient_loss.py
|   ├── cal_noise_gradient_loss.py
|   ├── filter_modulation.py
|   ├── generate_gradient_map.py
|   ├── vis_filters_distribution.py
├── tools
│   ├── srcnn_style
|   |   ├── faig.py
│   │   ├── ig.py
│   │   ├── absolute_filter_change.py
│   │   ├── random_neurons.py
│   │   ├── mask_neurons.py
│   │   ├── comparison.py
│   ├── srresnet
│   │   ├── faig.py
│   │   ├── ig.py
│   │   ├── absolute_filter_change.py
│   │   ├── random_neurons.py
│   │   ├── mask_neurons.py
│   │   ├── comparison.py

Usage of analysis Folder

Find Filters For Specific Degradation

For both SRCNN_Style network and SRResNet network, we provide four different methods to find filters for specific degradations, which are FAIG, IG, , and random. You just need to run the corresponding scripts and you can get the txt file that records the sorted filters of specific degradation in descending order. Note that you need to run all these four scripts. Take FAIG as an example,

python analysis/tools/srresnet/faig.py

Output results are saved in the results/Interpret/neuron-search folder.

---

Compare against these methods

To quantify the discovered filters' contribution to the network function, you first need to run the following script to use the target model to restore different degraded images. Take SRResNet as an example,

python inference/inference_srresnet.py

Then, for both SRCNN_Style network and SRResNet network, we provide a script analysis/Tools/SRResNet/comparison.py to compare the performance of FAIG, IG, and random. It automatically includes several steps:

• Replace the weights of discovered filters (1%, 3%, 5%, 10%) in target model with those filters (at the same location) in baseline model and obtain a substituted model.
• Use this substituted model to restore different degraded images (e.g, blurry or noisy images).
• Generate the gray gradient map of images (saved in npy format).
• Quantify the discovered filters' contribution to the network function by measuring output difference (MSE of gradient map) of target model and the substituted model.

python analysis/tools/srresnet/comparison.py

Output results are saved in results folder, the folder structure is

results
├── Interpret
|   ├── gradient  # contains the generated gray gradient map
|   ├── Masking   # contains the restored results of the substituted model
|   ├── metrics   # records the output difference

Mask Discovered Filters

Specifically, we introduce the naming rule of results/Interpret/Masking folder. When you delve into this folder, you will find some complicated directory structures (e.g, MaskdenoiseFilter/Blur2_LRbicx2), it means that the blurry input image is inferenced by the model that is masked deblurring filters (we hope this new model loses deblur function while contains denoise function). For the FAIG-discovered filters, you can observe the similar phenomenon as the picture shown below.

---

Filter Distribution Visualization

We provide a script analysis/scripts/vis_filters_distribution.py to help visualize the discovered 1% filters in SRResNet for different function.

python analysis/scripts/vis_filters_distribution.py

The deblurring filters are more located in the back part of the network while denoising filters locate more uniformly.

---

Filter Modulation

We provide a script analysis/scripts/filter_modulation.py to achieve controllable restoration based on the FAIG-discovered discriminative filters. After adjusting , you can get a similar phenomenon as shown in Figure 10 and Figure 11 of the supplementary file.

python analysis/scripts/filter_modulation.py

We can adjust coefficient for denoising filters to obtain continuous denoising effect.

We can adjust coefficient for deblurring filters to obtain continuous deblurring effect.