The performance of this code

[DavidLee-4032]

Hello, Louis.
Thanks for providing the code of pytorch version.
I've run the code and try to estimate the performance. I used 40000 graphs(all graphs are finished) and run for 5 epochs. (to speed up the procedure, I changed the method of adjacent matrix), and using 1000 graphs for test.
However, this code only have a performance of about 1.5 (in regards to the worst case) and an average performance of about 1.15 , lower than the ones in the origin code (worst case ~1.1 and avg. case ~1.001). I tried many approaches to improve but failed. Would you like to analyze the reasons? Thanks.