lbin/DCNv2

FP16 training error

qing3 opened this issue · 2 comments

qing3 commented

Hi, I want to use fp16 to train DCN, but error "RuntimeError: expected scalar type Float but found Half (data_ptr at /home/amax/.conda/envs/pytorch/lib/python3.6/site-packages/torch/include/ATen/core/TensorMethods.h:5747)" occurs. Do you know the reason? Hope for your reply!

qing3 commented

Training with FP32 is no problem.

lbin commented

pls follow 1.7 branch @qing3