Open MarcCote opened 4 years ago
Hi, I think there's an error in AvgReadout with a mask. The mask summation should be performed along the first dimension only. https://github.com/PetarV-/DGI/blob/0afce4e36b5edbe1e735536d15b748d0381e4083/layers/readout.py#L15
AvgReadout
It is return torch.sum(seq * msk, 1) / torch.sum(msk) but should be return torch.sum(seq * msk, 1) / torch.sum(msk, 1) # Note the dimension here.
return torch.sum(seq * msk, 1) / torch.sum(msk)
return torch.sum(seq * msk, 1) / torch.sum(msk, 1) # Note the dimension here.
Hi, I think there's an error in
AvgReadoutwith a mask. The mask summation should be performed along the first dimension only. https://github.com/PetarV-/DGI/blob/0afce4e36b5edbe1e735536d15b748d0381e4083/layers/readout.py#L15It is
return torch.sum(seq * msk, 1) / torch.sum(msk)but should bereturn torch.sum(seq * msk, 1) / torch.sum(msk, 1) # Note the dimension here.