how rarefied graph work?

[Shi-Mingyang]

Hello, and thank you for your contributions!

While testing ariadne-planner, I noticed that the node_manager.py file differs significantly from the file with the same name in ariadne. Specifically, it seems that the input to the observation comes from the rarefied graph, which appears to be different from the form of the observation input used during training. Do you think this discrepancy between training and testing is reasonable? Also, what is the difference between the rarefied graph and the graph output by node_manager during training?

Additionally, I noticed that NODE_PADDING_SIZE is set to 360. Is this because the maximum number of nodes in the images from the dataset is 360? Where does the value 360 come from?

I would greatly appreciate any clarification you can provide!

作者您好，感谢您的贡献！
我在ariadne-planner的测试中发现node_manager.py这个文件与ariadne中的同名文件差别很大，具体体现在observation的输入似乎来源于rarefied graph，这看起来跟训练过程中的observation的输入形式是不同的，您认为这种训练和测试之间的差异是合理的吗？rarefied graph与训练时的node_manager输出的graph有什么不同呢？
另外您的NODE_PADDING_SIZE为什么是360？是因为数据集中的图片的节点数量最多为360吗？360来源于哪里？

[caoyuhong001]

The rarefied graph is sparsified from the original graph, removing some zero-utility nodes. I believe that reasoning about the original dense graph is harder than the rarefied graph, therefore I choose to train the model on the dense graphs (but on a smaller scale). Of course, the rarefaction will cause some different observation distribution. But I did a test on a model trained with the rarefied graph and the performance was almost the same as the model trained on the dense original graph. Based on such a result, I think the difference is acceptable and validates the generalizability of the trained model,

And yes, 360 is the maximum number of nodes of a graph during training. To do the batch training, node observations at different timesteps need to be padded to this size. It is an empirical number based on the tests on the training dataset.