search

szq0214 / DSOD

DSOD: Learning Deeply Supervised Object Detectors from Scratch. In ICCV 2017.
Other
706 stars 210 forks source link

small batchsize has a lower mAP. #30

Open DL-ljw opened 6 years ago

DL-ljw commented 6 years ago

Hi, @szq0214: I only have two GTX 1080 GPUs. I want to reproduce you GRP-DSOD. When I change the batch_size and accum_batch_size to 6 and 30, the mAP is just 63%. What I should do to get the results as you paper? Thanks.

szq0214 commented 6 years ago

@DL-ljw, batch_size can be a small value (6 is fine), but accum_batch_size should be 128 or larger.

DL-ljw commented 6 years ago

When accum_batch_size is 128, it wil train a very long time. Maybe I should have a better machine,

DL-ljw commented 6 years ago

Hi, @szq0214: several days later. I set batch_size = 6 and accum_batch_size = 128, it run 250 hours, and mAP is 76.5%(lower than 79.0 % in paper on pascal voc 2007 test)? Is the batch_size too small? What should I do next? We are planing to buy four GTX1080, these can set a larger batchsize, can they work? Thanks.