Can you tell me what accuracy have you achieved?

[greathope]

with this command on vqa v1.0: python train.py --vqa_trainsplit trainval --path_opt options/vqa/mutan_att_trainval.yaml

[Cadene]

@greathope As far as I remember, we got those results on the testset (with our old Torch7 code, but you should get the same results with this code in pytorch):

Overall Accuracy is: 64.77
Per Answer Type Accuracy is the following:
other : 54.47
number : 38.43
yes/no : 83.81

[greathope]

@Cadene with trainval datasets? Actually I got 65.42...

[Cadene]

That's good news, isn't it?

It might be related to this little change in architecture: https://github.com/Cadene/vqa.pytorch/issues/18

What do you think? @greathope

[eustcPL]

@Cadene how about: vqav1.0 python train.py --vqa_trainsplit trainval --path_opt options/vqa/mlb_att_trainval.yaml, what accuracy did you acheived?