Relation_Networks-pytorch icon indicating copy to clipboard operation
Relation_Networks-pytorch copied to clipboard
verified publisher icon heefe92

wrong in PositionalEmbedding？

Open fregulationn opened this issue 6 years ago • 3 comments

I compare your code with mxnet(official) version, find that in PositionalEmbedding function，maybe the dim_mat should not be divided by 1？

dim_mat = mx.sym.broadcast_power(lhs=mx.sym.full((1,), wave_length), rhs=(8. / feat_dim) * feat_range)

does mx.sym.full((1,wave_length) mean divided? i am not really sure

fregulationn avatar Mar 24 '19 02:03 fregulationn

@heefe92 @fregulationn I meet a problem. When I run the train.py file after I modified the config.py file,there is a AttributerError:module 'model' has no attribute 'resnet101'. Does this model come from torchvision.models..resnet101 or the author's own definition? in model.py? Can you helpe me?Thank you

shajie17 avatar Jul 10 '19 01:07 shajie17

mx.sym.full returns a new array of given shape and type, filled with the given value val so I think mx.sym.full((1,), represents [1000]

wsx66848 avatar Apr 13 '20 13:04 wsx66848

the reason why dim_mat is divied by 1 is that in the official version, div_mat = mx.sym.broadcast_div(lhs=position_mat, rhs=dim_mat),while in the autor's @heefe92 version, it is mul_mat = position_mat * dim_mat

wsx66848 avatar Apr 13 '20 13:04 wsx66848