Open WreckItBoss opened 10 months ago
Thanks for your interest! And I'm deeply sorry for the late response.
I think the number of GPUs might matter. We only used one GPU on P-Stance. Actually we also noticed that results would differ when varying the number of GPUs. You might also wanna just use one GPU to see if your results would match. Thank you!
Hello, I am currently doing stance detection research that is based off of this research paper and I'm having a bit of a problem. When running the model with 4 GPUS (a400x4), the results for BERTweet+ increases by 0.2 for Trump, Biden and 0.5 for Sanders (which is normal), but for WS-BERT-DUAl, none of the results gets even close to the results written on the paper. In the paper the results are Trump: 85.8, Biden: 83.5, Sanders 79.0, Average: 82.8. However, when I run it on the server, the results are Trump: 84.2, Biden: 82.7, Sanders: 75.6 averaging at 80.83. Were there changes made to the code? What was the environment you ran on and I'm wondering if there are potential ways to fix this problem.