Closed y-kl8 closed 4 years ago
You can reference this: https://github.com/juntang-zhuang/ShelfNet/blob/citys_lw/ShelfNet18_realtime/test_speed.py. Your test script contain the time of copying output tensor from gpu to cpu, which may take much more time than model inference because the size of output tensor is large.
my test code like this,. The speed of deeplabv3-pluse-mobilenetv2(use ASPP and Decoder) is faster than the deeplabv3-pluse-mobilenetv2(doesn't use ASPP and Decoder). Do somebody know why?