Xinyu Fu
Sorry, my current focus is not image processing and I pretty much have forgotten all the algorithm details. You might need to figure out the bug yourself.
You may have a look at this [paper](https://arxiv.org/abs/1708.08988).
In my original experiments, I tried homogeneous GNNs with (1) the original graph data without node/edge type information and (2) the metapath-based homogeneous graph constructed by the best-performing metapath. On...
> hi [@cynricfu](https://github.com/cynricfu) ,thanks for the report. This is likely due to the Transformer display logic using `total_batch_size` without accounting for `dp_world_size != world_size` You can ignore it for now...
> If I remember correctly, tp_plan is for vllm integration, where tp_plan is used for inference time sharding. > > AutoTP is originally for deepspeed's inference engine, just like vllm....
Can other models achieve normal performance on your extracted subgraph? You may also need to check if you properly run the MAGNN preprocessing code on your subgraph.
确实，数据预处理代码比较粗糙，没有考虑并行处理等加速/优化方法，对于较大的图来说需要的时间可能很久。您可以考虑修改数据预处理的代码，使用比如Python的multiprocessing包来进行多进程计算。