Pretrained model's performance is bad on vqa tasks

#63
by hcy511 - opened

Does anyone experience the same issue as I do? When I use this pretrained version to evaluate vqav2 the accuracy is super low (0.xx%) while the ft version is pretty good (76%). I checked the generated answers and they seemed random when I used the pretrained model.

Microsoft org

@hcy511 pre-training does not involve any VQA data.

haipingwu changed discussion status to closed

Sign up or log in to comment