search

libuyu / GHM_Detection

The implementation of “Gradient Harmonized Single-stage Detector” published on AAAI 2019.
MIT License
617 stars 110 forks source link

add ghm loss to other network #18

Open congxin513 opened 5 years ago

congxin513 commented 5 years ago

Thank you for your good paper！I encountered some problems when I added ghm loss to other one stage networks. When adding ghmc_loss alone, it does not decrease during training, the network can't learn anything from it. And when adding ghmr_loss alone, it is nan. What parameters should I adjust, or which step I ignored?

libuyu commented 5 years ago

This issue may help you: https://github.com/libuyu/GHM_Detection/issues/10

congxin513 commented 5 years ago

@libuyu Thank you for your reply. I tried momentum 0.5 and mu 0.02, but it still doesn't work. How to adjust the momentum and mu when the dataset is small?

xiaowei-ui commented 5 years ago

这个问题可能对您有所帮助：＃10

Hello, I just get in touch with the field, not very clear. Now I want to apply GHM to SSD or DSSD. What should I do specifically? Need to replace the loss function of SSD for the loss function of GHM ?or other methods？

wanglaotou commented 4 years ago

@libuyu Thank you for your reply. I tried momentum 0.5 and mu 0.02, but it still doesn't work. How to adjust the momentum and mu when the dataset is small?

hi, may i know the details about how to add the ghm loss to the ssd? thanks.