Open gzhcv opened 5 years ago
Hi,
I also encountered this bug, it seems some details in "DCN" lib are not handled well enough to avoid this problem, so when it comes to the last step during training process, some tensors cannot be divided.
My solution is to set batch_size = 1
, and it will work but time-consumingly.
@chengdazhi Hope u can fix this bug, it shouldn't take u much time.
Thanks! Ian
Yes,when I set batch_size to 64,it will not wrong here
But sometimes my model likes to increase the loss but not decrease the loss when he reached at 22% accuracy , backward to the contrary side ? It's very funny
Hi, I came across the same problem, but it seems that it is not a bug. Because the "im2col_step" is a parameter. As is shown bellow:
We can always choose a proper "im2col_step" ensuring that "batch % im2colstep == 0".
To this end, I believe the problem is solved. However, I would like to know how "im2col_step" influences efficiency? Does a bigger value boost efficiency?
I agree with you, I also wonder how im2col_step influence the results.
I set batch_size = 64 ,and im2col_step=64,but still meet the error
Environment:
Error occurs as follows when the batch_size is 128 or 192 or 256. But when
batch_size = 64
is ok.I found that the
im2col_step = 1
by default, it's strange whybatch % im2col_step_ != 0
as the bug refers to?