search

ronghanghu / seg_every_thing

Code release for Hu et al., Learning to Segment Every Thing. in CVPR, 2018.
http://ronghanghu.com/seg_every_thing/
Apache License 2.0
424 stars 72 forks source link

train on own dataset #1

Open wangshuai09 opened 6 years ago

wangshuai09 commented 6 years ago

I convert my dataset to the coco format，the dataset has no segmentation，so i set the "segmentation" =[] in annotations，and set the cfg.TRAIN.MRCNN_LABELS_TO_KEEP = ( )，i also convert the pretrained coco MXRCNN model to my dataset followed the convert_coco_model_cityscapes.py

finally, i got the error: Exception encountered running PythonOp function: ValueError: min() arg is an empty sequence

Is the cfg.TRAIN.MRCNN_LABELS_TO_KEEP =( ) means that segmention would not be used? how should i train bbox branch on my dataset which without segmentation from the coco pretrained MXRCNN model?

jinxuan777 commented 6 years ago

@xiaozr @ronghanghu I met same problem, have you solved it?

wangshuai09 commented 6 years ago

i just set the "segmentation" = [[1,1,1,1,1,1]] ,and during the training ,the mask loss is 0.0

lssily commented 5 years ago

@xiaozr May I ask you that how do you install seg_every_thing? Looking forward to your reply!