search

qqlu / Amodal-Instance-Segmentation-through-KINS-Dataset

Amodal-Instance-Segmentation-through-KINS-Dataset
133 stars 19 forks source link

About training Mask R-CNN #12

Closed YoruCathy closed 4 years ago

YoruCathy commented 4 years ago

Hello, you listed result on Mask R-CNN, with det result around 31. I tried to train Mask RCNN on KINS dataset but I just can't get that score. Pleazzzzzzzz tell me how did you train Mask R-CNN? How many GPUs did you use, and your batch size, lr, lrdecay and some other parameters? Thanks a lottt

qqlu commented 4 years ago

Hi, 8 GPU was used with batch size 16 and min size 1666. I think your image size may be very small.

YoruCathy commented 4 years ago

Ah, my images are resized to 1333*800 keeping ratio, maybe I have to make them larger By the way how many epochs or iterations did you train?

qqlu commented 4 years ago

The hyper-parameters are set in my reference code (configs/kitti/). The max iterations are 24000 with steps [0, 20000, 22000].

YoruCathy commented 4 years ago

Oh I see, thank you (●'◡'●)