XAI

• eXplainable AI
• AI college Recording Repo

can see the demo at http://app.soopace.com (not always on) or run it by docker at local

Requirements

python == 3.7
pytorch == 1.7.1
torchvision == 0.8.2
jupyter >= 1.0.0

Run in Docker

better use nvidia-docker to run it, if you don't have nvidia-docker, please remove --runtime=nvidia option

$ docker pull simonjisu/xaivision:v1.0
$ docker run -it -d --name xai -p 7013:7013 --runtime=nvidia simonjisu/xaivision:v1.0

Run in local

1. install requirements packages

• [pytorch] install with your environments
• [jupyter] better to install with anaconda
• [streamlit] install via pip

2. quick tutorial

run following scripts to download all model weights(1.4GB), you can also download from google drive

$ sh download-weight.sh

run streamlit application

$ streamlit run app.py

after this go to browser http://localhost:8501

3. training from scratch

you can also train from scratch if you want. You can choose 3rd argument in "experiments" by following:

• 1: plain
• 2: rcd
• 3: rcd-fgm
• 4: rcd-noabs

options means:

• plain: basic setting
• rcd: gray scale for all attribution methods(means that reducing the color dimension to 1)
• fgm: fill the masks with global mean of all datas instead of zeros.
• noabs: not to absolute attribution scores in some methods

# 1: data-type: one of 'mnist', 'cifar10'
# 2: eval-type: one of 'roar', 'kar'
# 3: experiments: one of 1~4
# 4: if you train the model first time(for each data-type), 
#    ensure this variable to `true`
$ sh fast-run $1 $2 $3 $4