search

IBM / EvolveGCN

Code for EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs
Apache License 2.0
508 stars 164 forks source link

About the datasplit settings #5

Closed gfizhu closed 4 years ago

gfizhu commented 4 years ago

Hi, I'm trying to produce the results in the article, but I found the settings are different. In the splitter.py there is a special handling , so the train sets are args.num_hist_steps less than the numbers in the article. Is this code produce your results or the comment out?

only the training one requires special handling on start, the others are fine with the split IDX.

        start = tasker.data.min_time + args.num_hist_steps #-1 + args.adj_mat_time_window

I tried to use the comment out codes, but index errors come out, please help.

gfizhu commented 4 years ago

Sorry, I found in the link_pred_tasker there is a minus operation, so the datasplit is the same as in the article. def get_sample(self,idx,test, **kwargs): hist_adj_list = [] hist_ndFeats_list = [] hist_mask_list = [] existing_nodes = [] for i in range(idx - self.args.num_hist_steps, idx+1):