Interpreting and Disentangling Feature Components of Various Complexity from DNNs

PyTorch implementation of Interpreting and Disentangling Feature Components of Various Complexity from DNNs (ICML2021). [pdf] [arxiv]

We will re-arrange the code for better readability and reusability soon.

Environment

• python 3.6
• pytorch==1.2.0

Dataset

• CIFAR10
• CUB-200-2011
• DOG120

The images in CUB-200-2011 and DOG120 are cropped using the provided bounding boxes. For example, to crop the CUB-200-2011 dataset:

python3 ./dataset/generate_dataset.py --dataset=dogs

Also, in this paper, datasets of different training samples are needed. For example, to generate a subset of CUB-200-2011 with 2000 training samples:

python3 ./dataset/generate_dataset.py --dataset=cub --data-size=2000

Note: In our experiments, we used CIFAR10 with 200/500/1000/2000/5000 training samples, CUB-200-2011 with 2000/3000/4000/5000 training samples, and DOG120 with 1200/2400/3600/4800 training samples.

Train and disentangle

Given the feature of a certain layer in a pretrained DNN, our method

• disentangles feature components of different complexity orders from the feature
• disentangles reliable feature in the target feature

For example, to train ResNet20 on CIFAR10-2000, and disentangle feature components of different complexity orders and reliable feature:

python3 train_net.py --model-name=resnet20 --data-size=2000 --dataset=cifar10
python3 disentangle.py --model-name=resnet20 --data-size=2000 --dataset=cifar10
python3 disentangle_reliability.py --model-name=resnet20 --data-size=2000 --dataset=cifar10

Evaluation

We further design a set of metrics to diagnose the disentangled feature components, including

• the reliability of feature components
• the effectiveness of feature components (codes will be updated later)
• the over-fitting level of feature components being over-fitted

Below is an example.

python3 eval_disentangle.py --model-name=resnet20 --data-size=2000 --dataset=cifar10
python3 eval_reliability.py --model-name=resnet20 --data-size=2000 --dataset=cifar10

Demos

Several notebooks are provided for visualization demos.

• Visualization of disentangled feature components, feature of different complexity orders, and feature with higher complexity orders is given in ./visualization_demo/vgg16_conv43_vis.ipynb

  

• Significance of disentangled feature components, reliability of feature components, effectiveness of feature components, and confidence of feature components being over-fitted are given in ./vis_demo.ipynb

  

  The last two will be updated later.

• Visualization of relationship between the distribution of feature components and the accuracy is given in ./vis_demo.ipynb

  

• Significance of disentangled feature components from the DNNs for different tasks is given in ./task_complexity/task_complexity_vis.ipynb