It seems that the function defined here needs to be modified as follows:

    def forward(self, q, k, v, mask=None):
        bs = k.shape[0]
        k = k.view(bs, -1, self.head, self.d_k)
        q = F.relu(self.q_linear(q).view(bs, self.head, self.d_k))
        ...

Please help check this.
BTW, thanks for sharing this implementation.

Hi, thanks for this helpful code! Have you replicated performance mentioned in the original paper?

The paper mentions that the output from the trunk is sliced to get the central frame's feature map and is then passed through the RPN.

"We slice out the temporally-central frame from this feature map and pass it through a region proposal network
(RPN)..."

How should I slice the output from the trunk? What is the input dimension expected by Tail?

Do I need to edit the BNClassifier class for regression problems? Training with mse loss and sgd set to 1e-3 doesn't converge but I can't intuitively work out why

Hi,
Thanks for your awesome project. I found a dependence issue when I try to run transformer_v3.py. Could you please tell me which dependence do I need? Thanks in advance!

from tools import BNClassifier, BottleSoftmax
ImportError: cannot import name 'BNClassifier' from 'tools' (unknown location)

Hi!

In the paper, it mentions that the query is an ROI, and key and values are from the clip around an ROI. But in this implementation, it seems that all query, key and value are from the features in the same spatial locations. In that case, it will not be able to aggregate another person that the target person is talking to.

Please clarify this I am mistaken. Thank you very much!