VQA-2.0

[erjpc]

Hello Author:
I have recently reproduced your paper, and according to the data set you gave, it is 'number': 50.91 in vqa-2.0.
'other': 59.45,
'overall': 69.13,
'yes/no': 85.29}}
The result is a little different from yours. Could you tell me what went wrong

[rentainhe]

Hello, would you like to provide the config and training dataset you use for this results

[erjpc]

Hello，
config
torch 2.3.1
spacy 3.7.5
en-core-web-lg 3.7.1
numpy 2.0.0
The training dataset use is provided by you to download from Baidu Cloud disk

[erjpc]

Hello， config torch 2.3.1 spacy 3.7.5 en-core-web-lg 3.7.1 numpy 2.0.0 The training dataset use is provided by you to download from Baidu Cloud disk

gpu 3090

[rentainhe]

We've listed our training hyper-param in model zoo: https://github.com/rentainhe/TRAR-VQA/blob/main/MODEL.md

Would you like to tell us which hyper-param do you use in your experiments

[erjpc]

We've listed our training hyper-param in model zoo: https://github.com/rentainhe/TRAR-VQA/blob/main/MODEL.md

Would you like to tell us which hyper-param do you use in your experiments

谢谢作者，已顺利解决但是目前train+val+vg跑出来只有71.42 没有达到72

[rentainhe]

We've listed our training hyper-param in model zoo: https://github.com/rentainhe/TRAR-VQA/blob/main/MODEL.md
Would you like to tell us which hyper-param do you use in your experiments

谢谢作者，已顺利解决但是目前train+val+vg跑出来只有71.42 没有达到72

You can resume from the 10-epoch checkpoint trained on train + val + vg and continue training it with train + val for the last 2 or 3 epochs, which may boost the final performance