Closed Anikily closed 5 years ago
@Anikily @liutinglt I found that when I was training, the memory became more and more until the GPU was full, and there was no error message such as out of memory. I used torch. cuda. empty_cache() and del variables, which had no effect. Have you ever encountered this kind of situation?
@Anikily This is the updated code, you can train with input size 473 as well, and the result will be higher.
@nemonameless please try smaller batch size
@liutinglt so,what's the right size to reproduce the performance in Readme using pretrained model?
@Anikily 384
I notice that the input size is 473 in paper, and the evaluate code set the size to 384