About the reproduction of VCR experiment results about villa

[zhegan27]

Thanks for your interest in the code. In the leaderboard (https://visualcommonsense.com/leaderboard/), if you see Entry #18, there seems to be a third party running this code to reproduce the results.

Since you only have 1 gpu, please change the hyper-parameter "gradient_accumulation_steps" from 5 to 20, and have a try again. Let me know how this goes, and hopefully the performance can catch up.

[Tclz]

Thanks for your interest in the code. In the leaderboard (https://visualcommonsense.com/leaderboard/), if you see Entry #18, there seems to be a third party running this code to reproduce the results.

Since you only have 1 gpu, please change the hyper-parameter "gradient_accumulation_steps" from 5 to 20, and have a try again. Let me know how this goes, and hopefully the performance can catch up.

Thanks for your reply! I am trying to modify the configuration file and re-experiment, after which I will feed back the results.

[Tclz]

Hi, this strategy worked. After I adjusted the hyper-parameter, when the training level reached 90%, I got the following results.
88%|########7 | 7000/8000 [15:35:25<2:10:08, 7.81s/it][1,0]:09/14/2021 18:04:58 - INFO - main - ============Step 7000=============
[1,0]:09/14/2021 18:04:58 - INFO - main - 4481024 examples trained at 79 ex/s
[1,0]:09/14/2021 18:04:58 - INFO - main - ===========================================
[1,0]: [1,0]:09/14/2021 18:04:58 - INFO - main - start running validation...
[1,0]: [[[[1,0]:09/14/2021 18:10:05 - INFO - main - validation finished in 307 seconds, score_qa: 75.16 score_qar: 78.46 score: 59.18 1,0]:
89%|########8 | 7100/8000 [15:53:25<1:57:14, 7.82s/it][1,0]:09/14/2021 18:22:58 - INFO - main - ============Step 7100=============
Thanks for your solution!