This official implement of CIDER is based on the [PyTorch Geometric]
conda create -n cider requirements.yaml
conda activate ciderpython main.py --dataset=[nci1, mutag]To do this, you should write some code to prepare the dataset and a model, and then use the example code to do this task. The example code is as follows:
from cider import CIDER
from util import train_one_epoch
from torch_geometric.data import DataLoader
from model import ExampleNet
from dataset import ExampleDataset
# prepare the dataset and model
dataset = ExampleDataset()
model = ExampleNet()
loader = DataLoader(dataset, batch_size=32, shuffle=True)
# train the example model
for data in loader:
## train the model
## ...
# use CIDER to explain the model
## train the cider model
cider = CIDER(...)
for epoch in range(100):
## please see main.py for more details
train_one_epoch(...)
## use the trained cider model to explain the model for a specific data(e.g. dataset[0])
explainations = cider.get_explainations(dataset[0].x, dataset[0].edge_index)