代码问题请教 about udp-pose HOT 4 CLOSED

@cyndixxxxx
if you use the combined target, the output channels of network should be 51, i.e. shape=(64,51,64,48), 17 channels for heatmap and 34 channels for offsetmap
otherwise if you use gaussian heatmap for target, the output channels of network is 17, and then, you should use API 'flip_back' instead of 'flip_back_offset'
you can figure out how the algorithm discriminate these two situations by referring line 157-162 in "UDP-Pose/deep-high-resolution-net.pytorch/lib/core/function.py"

from udp-pose.

How to generate the output to be shape=(64,51,64,48)? .yml file is filled with ‘offset’

from udp-pose.

@cyndixxxxx just use the default TARGET_TYPE('offset') as in default .yml ( UDP-Pose/deep-high-resolution-net.pytorch/experiments/coco/hrnet/w32_256x192_adam_lr1e-3_offset_ofm.yaml )

from udp-pose.

Thank you for your reply

from udp-pose.