search

midas-research / sthan-sr-aaai

98 stars 32 forks source link

Each epoch results are the same #9

Open From-zero-tohero opened 1 year ago

From-zero-tohero commented 1 year ago

The training results of each round are the same. It's strange. The training results after the first round will not change

TompaBay commented 1 year ago

May I ask how did you train without the hypergraph npy file? Seems like their preprocessing codes ouptut a 3d matrix instead of a 2d matrix, and it does not work for the sparse.coo_matrix function. Thx!

M1stF0rest commented 8 months ago

True

May I ask how did you train without the hypergraph npy file? Seems like their preprocessing codes ouptut a 3d matrix instead of a 2d matrix, and it does not work for the sparse.coo_matrix function. Thx!

Hi, in the README file, the authors referred another repository which contained the relations of companies which are used to build the hyper-graph. You can downloaded them and composed the .npy file by yourself. The author did not mention too much about the details of it, but it seems like an instance matrix with rows being nodes and columns being edges. (If mth row is connected by nth hyper-graph, then the [m,n]th entry of this matrix will be 1, otherwise 0). I constructed one and the model runs successfully, but the results I got was very different from the authors'. In the paper, the authors used 862 hyper-edges, but I only got 140 such edges based on the data. I guess maybe this is where the problem lies. Hope you find this info useful!