search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.7k stars 11.71k forks source link

TRAINING SLOW #719

Open prasoonjha opened 6 years ago

prasoonjha commented 6 years ago

training on CPU takes about half an hour per epoch with only 10 steps per epoch.How do i speed up training process.please be elaborate. Thank you MY SYSTEM CONFIGURATION: i5 7th gen,8 gigs ddr4 ram, 2gb nvidia gtx 1050 also,the pretrained weight file im using is inception_resnetv2

bahmed11 commented 6 years ago

Mask RCNN is slow because it doesn't contain any batchnormalization. There is a new solution that has been just released in the paper below which is using groupnormalization. Groupnormalization has been implemented by few people and all are available on github. If you are using resnet 101 layers, you might need to lower the number of groupnormalization layers as it will run out of memory unless if you used higher resources.

This is the link to the paper: https://arxiv.org/pdf/1803.08494.pdf

engineer1109 commented 6 years ago

How can you run by nvidia gtx 1050 ??????

prasoonjha commented 6 years ago

why?is there something wrong with a gtx1050 video card?