I have a question,
when query pass through the linear layer of sampling_offest, I found out that its value is between +-4 at the first epoch,
and then normalized by offset_normalizer, which stands for feature map size like 32, 16, etc.
After the above steps, the sampling_locations will not keep in the range [0,1], and then the sampling_grids = 2 * sampling_locations - 1 for F.grid_sample (I am using the pytorch version) will not follow the rule of range (-1, 1).
Thank you for the great work!
I have a question, when query pass through the linear layer of sampling_offest, I found out that its value is between +-4 at the first epoch, and then normalized by offset_normalizer, which stands for feature map size like 32, 16, etc.
After the above steps, the sampling_locations will not keep in the range [0,1], and then the sampling_grids = 2 * sampling_locations - 1 for F.grid_sample (I am using the pytorch version) will not follow the rule of range (-1, 1).
Could you help me figure out this problem?