Open abhishekk06 opened 3 years ago
Hi- @puneeshkhanna do you have the script to generate the .csv files ? @abhishekk06 i will upload a file "ddadjcols100.c" which can be used to create the new data set. You can look at this file to see how the random graphs are created. BTW, you would require a package called very_nauty ( http://keithbriggs.info/very_nauty.html ) to be installed and the library linked to get this file to compile.
DD,
We used the ready made files as dataset. Didn’t try to create dataset yet.
Regards, Puneesh
Sent from my iPhone
On 23-Apr-2021, at 17:40, dasdibye @.***> wrote:
Hi- @puneeshkhanna do you have the script to generate the .csv files ? @abhishekk06 i will upload a file "ddadjcols100.c" which can be used to create the new data set. You can look at this file to see how the random graphs are created. BTW, you would require a package called very_nauty ( http://keithbriggs.info/very_nauty.html ) to be installed and the library linked to get this file to compile.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
Hi,
Let me try to run and understand "ddadjcols100.c" and if possible I can also extend it to dump csv for us to directly feed our model itself
Hi @dasdibye,
Can you please share the command line to compile "ddadjcols100.c"
clang -O2 -DPAD ddadjcols100.c -I../.. -o adjcols100 -lvn_graph -L../.. -lm.
You should have built the very nauty library which should be available as vn_graph.a i think. also need to include the path of the header file #include "vn_graph.h"
Hi @dasdibye / @puneeshkhanna,
I tried creating a synthetic dataset(~10K sample) using networkx/grinpy and trained my GNN model on it.
With my synthetic dataset, I was able to get a val. accuracy of around 90% but in order to do a better comparison I thought of retraining my model on the dataset available in the repo (adjcols100.csv) and for some reason, I see a ~15% drop in the accuracy. I suspect my GNN model needs more data to run and it will be really great if you can share the script to create new data points.
BTW I already tried playing with the regularization and pooling layer but no major impact.