question about the bug of Location Projection in trainer.py about meta-sr-pytorch HOT 4 CLOSED

the problem seems to be caused by the floor operation? Maybe int(4.0)==5. Since I can not reproduce this problem, i do not know the true the problem.
So, it is very helpful if you provide more information about the python version.

from meta-sr-pytorch.

Thank you for your answer.
my torch_version is 1.0.1, python version is 3.5 and in my python int(4.0)==4
I printed the shape of variable:
h_project_coord_shape: torch.Size([512])
h_offset_shape: torch.Size([128, 4, 1])
offset_h_coord_shape: torch.Size([512])

from meta-sr-pytorch.

check the int_project_coord matrix. maybe the torch.floor(h_project_coord)
or int_h_project_coord = int_h_project_coord.int() cause the floor problem.

some pixels apear 5 times(which will cause the index out problem).

from meta-sr-pytorch.

thanks, I know the problem.
It is the fault of tensor.mul(). I find that in my torch version 'torch.arange(0,outH, 1).mul(1.0/scale)=0'
change
h_project_coord = torch.arange(0,outH, 1).mul(1.0/scale).float()
to
h_project_coord = torch.arange(0,outH, 1).float().mul(1.0/scale)
It is ok

from meta-sr-pytorch.