search

DetectionTeamUCAS / RRPN_Faster-RCNN_Tensorflow

A tensorflow re-implementation of RRPN: Arbitrary-Oriented Scene Text Detection via Rotation Proposals.
254 stars 89 forks source link

training slow #12

Closed loscheris closed 5 years ago

loscheris commented 5 years ago

Hi, I found the training is extremely slow, it took about 2 mins to finish one step (since batch size is 1, I assumed for every step only one image is trained). All images have size about 800*1000 after reshaping. I run the code on K80 GPU and the trianing process only occupies 72MB GPU memory. I'm wondering if there's any problem in my environment. Thanks

yangxue0827 commented 5 years ago

Mainly slow in iou's calculations and nms processing, you can replace it by implementing a more efficient calculation method yourself. Or you can use this code. @loscheris

loscheris commented 5 years ago

Thanks! @yangxue0827 , I will have a try R2CNN later!

Btw, Is this a normal training speed in RRPN? What's your training speed in RRPN?

yangxue0827 commented 5 years ago

Not normal, you may set wrong gpu index. Please check the default gpu in cfgs.py, and set right index according to your environment. @loscheris

DianaZhang commented 5 years ago

Hi, I have the same problem. When I use train. py, a step takes about 5 minutes, and I have two gpUs, but only one GPU is called. I modified cfgs.py, line 15 with GPU_GROUP = 0,1, but it was of no use. I don't know what to do next.