For more information, please visit https://vision-mamba-mender.github.io/.

Requirements

• causal_conv1d: pip install causal-conv1d
• mamba: pip install mamba-ssm
• pip install -U openmim
• mim install mmcv
• pip install mmsegmentation
• pip install mmengine
• pip install ftfy

Algorithms

🐍 Preparation

1. Train a pre-trained model:

bash scripts/train.sh

2. Select high-confidence samples and low-confidence samples：

bash scripts/sample_selection.sh

3. Extract intermediate features from the model, including activations and gradients of samples:

bash scripts/feature_selection.sh

🔍 Interpreting

4. Visualize external interactions of states or internal interactions of states

bash state_external_visualize.sh
bash state_internal_visualize.sh

5. Binarize external interactions of states or internal interactions of states

bash state_external_mask.sh
bash state_internal_mask.sh # Steps necessary for calibration!

6. Compute interpretability scores for external interactions of states or internal interactions of states

bash state_external_score.sh
bash state_internal_score.sh

🔧 Calibrating

7. Calibrate external interactions of states or internal interactions of states

bash scripts/train_repair.sh

Future Plans

This repository contains the initial version of the code for the Vision Mamba Mender paper. Please stay tuned for further refinements and detailed explanations in future updates.

Others

• The directory structure of the files outputted by the algorithm:

output_path # Overall output directory as defined by you
    ├── exp_name # Experiment name defined by you
            ├── models
            ⎪       └── xxx.pth       
            ├── samples
            ⎪       ├── htrain # Selected high-confidence samples
            ⎪       └── ltrain # Selected low-confidence samples
            ├── features
            ⎪       ├── hdata # Selected intermediate features of the high-confidence samples
            ⎪       ⎪       └── xxx.pkl 
            ⎪       └── ldata # Selected intermediate features of the low-confidence samples
            ├── visualize
            ⎪       ├── hdata
            ⎪       ⎪       ├── external # Visualization results of external interactions of states
            ⎪       ⎪       ⎪       └── xxx.JPEG/PNG
            ⎪       ⎪       └── internal # Visualization results of internal interactions of states
            ⎪       └── ldata # Same as above
            ├── masks
            ⎪       ├── hdata
            ⎪       ⎪       ├── external # Binarization results of external interactions of states
            ⎪       ⎪       ⎪       └── xxx.pt
            ⎪       ⎪       └── internal # Binarization results of internal interactions of states
            ⎪       └── ldata # Same as above
            └── scores
                    ├── hdata
                    ⎪       ├── external # Interpretability scores of external interactions of states
                    ⎪       ⎪       └── xxx.PNG
                    ⎪       └── internal # Interpretability scores of internal interactions of states
                    └── ldata # Same as above