NOthing about vqa_regat

[linjieli222]

Thank you for pointing it out. You are definitely right.

However, I don't think the current version would cause any error as we initialize q_dim to be the same as num_hid

VQA_ReGAT/model/regat.py

Line 115 in 9f6fe5b

joint_embedding = BUTD(args.relation_dim, args.num_hid, args.num_hid)

[alice-cool]

Dear scholar,
I feel the third picture could not corespond with the first formation.

I think the sum signal is some confused. Because in this code the visual data wasn't injected. I think maybe when injecting the visual data , the label_biases_att could be introduced( the label num is converted to the bias term)
The below picture is from https://github.com/linjieli222/VQA_ReGAT/blob/master/model/graph_att_layer.py