salesforce
Finetune vqa by own data
Hi, I'm here again.
This time I meet a new question.
I want to finetune the vqa checkpoints with my own data. There are two questions confused me.
The 1st one is that which init checkpoints should I choose. The pre-trained checkpoints or the finetuned checkpoints on VQA? I noticed both checkpoints you provide download :)
The 2nd question is how to create the vqa dataset in the correct format. I have some photos now, and for all these photos, I just want to ask an yes/no question on it. I haved checked the coco vqa annotations, I found for every picture, It has five keys per image and it seems have 3 important keys for me : 'question_id', 'question', 'answer'.
If I want to create my own question, which id I should use for the question? Maybe just choose an unique number?
And there are many repetitions words in the 'answer', I can't understand what does this mean. If I set a yes or no question, how to set the 'answer'?
Looking forward to your reply and thanks for your help.
