search

DrSleep / light-weight-refinenet

Light-Weight RefineNet for Real-Time Semantic Segmentation
Other
734 stars 163 forks source link

Broken pipe #77

Closed hotfinda closed 2 years ago

hotfinda commented 2 years ago

Hi, Thanks for this contribution. I am training the network with GTX 1080Ti, but it shows broken pipe. Except the batch size, which parameter can I change to reduce the cost of GPU? I only find the batch-size as (6,6,6)in the config.py, and I have change it to (1,1,1), but it still shows the problem

DrSleep commented 2 years ago

but it shows broken pipe

what do you mean by this?

hotfinda commented 2 years ago

Hi I have solved this problem. And another question is that why you have 3 stages during training?

DrSleep commented 2 years ago

The first stage is COCO+BSD+VOC, the second stage is BSD+VOC, the final stage is VOC. The idea is that the VOC data contains the best quality annotations out of all three, so you can think of the first two stages as a pre-training. That is how it was done in the original RefineNet work, and I just followed it.

hotfinda commented 2 years ago

so if i cite this, I should cite the original paper and your code ,right?

DrSleep commented 2 years ago

yes, I think so