You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, thanks for your job!
There are two points in the code I cannot understand:
1.line 133 in tasn-mxnet/example/tasn/model.py: mx.nd.batch_dot(b.reshape((n,c,1)), f.reshape((n,c,w*w)), True, False).reshape((n,1,w,w))
b is sorted but f not, i don't think their attention channels can match one by one.
2.the channels in both Structure_Att and Detail_Att are filtered to make sure that each channel is unique,but why you think the channels with the same sum value can be represented by one of these channels?
looking forward to your reply.
The text was updated successfully, but these errors were encountered:
@yogsin Sorry for my late reply, and thanks for your valuable comments.
I think you are right, and I will fix this bug.
I thought it is a rare case for two different attention maps share the same sum value, and of course, you can try a more strict unique selecting strategy.
Hi, thanks for your job!
There are two points in the code I cannot understand:
1.line 133 in tasn-mxnet/example/tasn/model.py: mx.nd.batch_dot(b.reshape((n,c,1)), f.reshape((n,c,w*w)), True, False).reshape((n,1,w,w))
b is sorted but f not, i don't think their attention channels can match one by one.
2.the channels in both Structure_Att and Detail_Att are filtered to make sure that each channel is unique,but why you think the channels with the same sum value can be represented by one of these channels?
looking forward to your reply.
The text was updated successfully, but these errors were encountered: